perf(mcp): checkNewComments — параллелизм с капом (#490)

checkNewComments делал O(N) последовательных REST-вызовов listComments по страницам
working set — большой space линеен по round-trip'ам. Теперь per-page фетчи идут с
ограниченным параллелизмом (cap 6, середина полосы 5–8): независимые чтения не ждут
друг друга, но и не заваливают сервер/сокеты.

mapWithConcurrency — крошечный пул без зависимости от p-limit: N воркеров тянут
следующий индекс с общего курсора. Порядок результатов сохраняется (по входному
порядку страниц), поэтому вывод детерминирован независимо от того, какой фетч
завершился первым. Серверный batch-эндпоинт «comments updated since T по space» —
опционально, отдельным заходом.

Тест (mock-HTTP): 13 страниц, задержанный /api/comments — maxInFlight > 1 и <= 6
(последовательная реализация дала бы 1), порядок результатов = порядок обхода.

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

.env.example

@@ -225,11 +225,26 @@ MCP_DOCMOST_PASSWORD=
 
 # Silence timeout (ms) for EXTERNAL-MCP transport ONLY (not the chat provider).
 # Tighter than AI_STREAM_TIMEOUT_MS so a byte-silent/hung MCP server is broken in
-# ~1 min instead of 15. Note it also cuts a legitimately long but byte-silent
-# single tool call (a slow crawl that emits nothing until done) and an SSE
-# transport idling >1 min BETWEEN tool calls. Default 60000 (1 min).
+# ~1 min instead of 15. It cuts a legitimately long but byte-silent single tool
+# call (a slow crawl that emits nothing until done) on the HTTP (streamable)
+# transport, which opens a fresh request per call. The SSE transport — one
+# long-lived body across many calls — is NO LONGER governed by this timeout
+# (as of #489): its idle-BETWEEN-calls window has its own, raised bodyTimeout,
+# AI_MCP_SSE_BODY_TIMEOUT_MS below. Default 60000 (1 min).
 # AI_MCP_STREAM_TIMEOUT_MS=60000
 
+# bodyTimeout (ms) for the EXTERNAL-MCP SSE transport ONLY (#489). The SSE
+# transport holds ONE response body open across many tool calls, so undici's
+# bodyTimeout (time between body bytes) counts the LEGITIMATE silence BETWEEN the
+# model's tool calls, not just a hung single call. At the tight 1-min silence
+# timeout above, a normal >1-min gap between calls would break the SSE socket and
+# the cache would serve a dead client until TTL — so the SSE transport gets its
+# OWN, RAISED bodyTimeout. A single stuck call is still bounded by the per-call
+# cap (AI_MCP_CALL_TIMEOUT_MS), and a socket that does break is healed by the
+# in-run transport-error retry. The HTTP (streamable) transport keeps the tight
+# timeout. Default 600000 (10 min).
+# AI_MCP_SSE_BODY_TIMEOUT_MS=600000
+
 # Total wall-clock cap (ms) for ONE external MCP tool call (app-level, not
 # transport). Aborts a tool that keeps the socket warm (SSE heartbeats / trickle)
 # but never returns a result — which the silence timeout above never breaks.
@@ -288,6 +303,29 @@ MCP_DOCMOST_PASSWORD=
 # registry is process-local).
 # AI_CHAT_RESUMABLE_STREAM=false
 
+# --- Run lifecycle tunables (#487) ---
+# These govern the universal run machinery (every turn is now a first-class run,
+# both modes) and rarely need changing.
+#
+# How long a server-side SUPERSEDE ("interrupt and send now") waits for the target
+# run to settle after issuing Stop before it degrades to a 409 SUPERSEDE_TIMEOUT
+# (nothing sent, the composer keeps the user's text). 10s is generous under a
+# healthy DB; do NOT raise it to paper over a slow DB — a SUPERSEDE_TIMEOUT is the
+# honest signal. Default 10000 (10s).
+# AI_CHAT_SUPERSEDE_TIMEOUT_MS=10000
+#
+# How often the periodic bidirectional reconcile job runs (heals runs/messages
+# left dangling by a crash or a lost terminal write). Default 120000 (2 min).
+# AI_CHAT_RECONCILE_INTERVAL_MS=120000
+#
+# Wall-clock cap for a SINGLE in-app tool call (a long paginated read, or a content
+# write whose collab commit hangs) — the per-call half of the composite abort
+# signal every in-app tool is wrapped with (the other half is the turn's Stop).
+# The reconcile staleness floor is derived as max(2 x this cap, 15min), so a very
+# high value delays stale-run recovery (the server boot-warns above 30min). Default
+# 120000 (2 min).
+# AI_CHAT_INAPP_TOOL_CALL_CAP_MS=120000
+
 # --- Anonymous public-share AI assistant ---
 # Opt-in per workspace (AI settings -> "public share assistant"; off by default).
 # When enabled, anonymous visitors of a published share can ask an AI about that
@@ -335,6 +373,20 @@ 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

.gitignore

@@ -29,6 +29,10 @@ packages/mcp/build/
 # is a build artifact like build/ — never committed, always fresh.
 packages/mcp/src/registry-stamp.generated.ts
 
+# token-estimate compiled output (#490; built in CI/Docker via `pnpm build` /
+# the server `pretest`, never committed, so src/ and prod can never diverge).
+packages/token-estimate/dist/
+
 # Logs
 logs
 *.log

AGENTS.md

@@ -455,7 +455,7 @@ The API server is a Fastify app with a global `/api` prefix (`main.ts` excludes
    - `core/ai-chat/tools/` — the agent's ~40 read+write tools. Every tool runs under the **calling user's** CASL permissions via a per-user loopback access token (`docmost-client.loader.ts`), so the agent can never exceed what the user could do. Only **reversible** operations are exposed (page history + trash; no permanent delete). Agent edits get an "AI agent" provenance badge in page history (`20260616T130000-agent-provenance` migration).
    - `core/ai-chat/embedding/` — RAG indexer + a BullMQ consumer on `AI_QUEUE` that embeds pages into `page_embeddings` (vector search), complementing Postgres full-text search. Pages are (re)indexed on edit; `AI_EMBEDDING_TIMEOUT_MS` bounds a hung embeddings endpoint.
    - `core/ai-chat/external-mcp/` — admins can attach external MCP servers (e.g. Tavily) to give the agent web access. **`ssrf-guard.ts` validates outbound MCP URLs against SSRF** — keep that guard in the path when touching external-MCP connection logic.
-   - `core/ai-chat/ai-chat-run.service.ts` + `ai_chat_runs` — **detached/autonomous agent runs** (`#184`), behind the per-workspace `settings.ai.autonomousRuns` flag (off by default). When on, a turn becomes a server-side RUN that survives a browser disconnect; only an explicit `POST /ai-chat/stop` ends it, and a client reconnects/live-follows via `POST /ai-chat/run`. **DEPLOY CONSTRAINT — single-instance only in phase 1:** Stop and the AbortController that backs it are process-local, so a Stop only aborts a run executing on the **same** replica that owns it (cross-instance pub/sub stop is phase 2). Do **not** enable `autonomousRuns` on a horizontally-scaled deployment (multiple replicas behind a load balancer, or Docmost cloud `CLOUD=true`) — run a single instance instead. The server logs a startup WARNING when it detects a multi-instance deployment (`CLOUD=true`) so the constraint is visible. The startup sweep settles any run left dangling by a restart.
+   - `core/ai-chat/ai-chat-run.service.ts` + `ai_chat_runs` — **every agent turn is now a first-class server-side RUN** (`#184`, universalized in `#487`): its lifecycle is tracked in `ai_chat_runs` in **both** modes, and the single-active-run-per-chat concurrency gate is enforced universally (a legacy second tab now gets a clean `409 A_RUN_ALREADY_ACTIVE` instead of a second parallel stream that interleaved history). The per-workspace `settings.ai.autonomousRuns` flag (off by default) **no longer gates whether a turn is a run** — it now controls **only the browser-disconnect semantics**: when ON the run is *detached* (a disconnect leaves it executing server-side; only an explicit `POST /ai-chat/stop` ends it, and a client reconnects/live-follows via `POST /ai-chat/run`); when OFF (legacy) a disconnect ends the turn by stopping its run via the run's stop lever. `#487` also adds a server-side **supersede** CAS ("interrupt and send now") to `POST /ai-chat/stream` (`supersede: { runId }`): it atomically stops the chat's currently-active run and waits for it to settle before the new turn claims the slot, returning `SUPERSEDE_INVALID` / `SUPERSEDE_TARGET_MISMATCH` / `SUPERSEDE_TIMEOUT` on the non-proceed branches. **DEPLOY CONSTRAINT — single-instance only in phase 1:** Stop and the AbortController that backs it are process-local, so a Stop only aborts a run executing on the **same** replica that owns it (cross-instance pub/sub stop is phase 2). Do **not** enable `autonomousRuns` on a horizontally-scaled deployment (multiple replicas behind a load balancer, or Docmost cloud `CLOUD=true`) — run a single instance instead. The server logs a startup WARNING when it detects a multi-instance deployment (`CLOUD=true`) so the constraint is visible. The startup sweep settles any run left dangling by a restart.
 
 ### Client structure
 Vite SPA. Code is organized by feature under `apps/client/src/features/*` (mirrors the server domains: `page`, `space`, `comment`, `ai-chat`, `editor`, …). Conventions:
@@ -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
@@ -190,6 +202,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   dangling by a restart. Phase 1 is single-instance-only (cross-instance Stop is
   not yet reliable); the server warns at startup on a horizontally-scaled
   deployment. (#184)
+- **Server-side "interrupt and send now" (supersede) for AI chat.** `POST
+  /ai-chat/stream` now accepts a `supersede: { runId }` field: when the user sends
+  a new message while a run is active, the server atomically stops that run and
+  waits for it to settle before the new turn claims the chat's single run slot,
+  instead of the send being rejected as concurrent. The compare-and-set surfaces
+  three codes on its non-proceed branches — `SUPERSEDE_INVALID` (the targeted run
+  is malformed / belongs to another chat), `SUPERSEDE_TARGET_MISMATCH` (a
+  different run is now active; carries the current `activeRunId`), and
+  `SUPERSEDE_TIMEOUT` (the previous run did not stop within the settle window, so
+  nothing was sent and the composer keeps the text). Tunable via
+  `AI_CHAT_SUPERSEDE_TIMEOUT_MS` (default 10s). (#487)
 - **Out-of-band page transfer via an in-RAM blob sandbox (`stash_page`).** A
   new MCP tool serializes a whole page (its full ProseMirror JSON, with every
   internal image/file mirrored) into an ephemeral in-RAM blob and returns only
@@ -270,6 +293,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Changed
 
+- **Every AI-chat turn is now a first-class server-side run, and one run per chat
+  is enforced in both modes.** The run machinery from `#184` was universalized: a
+  turn is tracked in `ai_chat_runs` and gated by the single-active-run-per-chat
+  index regardless of the `settings.ai.autonomousRuns` flag. **Behavior change:**
+  a second tab (or a double-submit) that starts a turn while one is already active
+  on the chat is now rejected up front with `409 A_RUN_ALREADY_ACTIVE` (carrying
+  the `activeRunId`); previously, on the legacy path, it opened a second parallel
+  stream on the same chat that interleaved history. The `autonomousRuns` flag no
+  longer controls whether a turn is a run — it now governs **only** the
+  browser-disconnect semantics (ON = detached/survives a disconnect; OFF = a
+  disconnect stops the run). (#487)
 - **Client markdown paste/copy and AI-chat rendering now go through the canonical
   converter.** Pasting markdown into the editor, "Copy as markdown", the AI title
   generator, and the AI-chat markdown renderer all now use
@@ -302,6 +336,23 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- **A chat with one malformed message part no longer 500s on every turn, and a
+  failed send no longer duplicates the user's message.** Incoming client parts
+  are now whitelisted to `text` (a forged tool-result part can no longer reach
+  the persisted history or the model context), and the turn is converted BEFORE
+  the user row is inserted, so a mid-flight failure cannot leave a duplicate
+  user row that a retry then compounds. A single part that still fails to convert
+  degrades to a `[tool context omitted]` marker on that one row instead of
+  bricking the whole chat. (#489)
+- **A transport drop to an external MCP server now heals within the same turn.**
+  On an undici transport error, a read-only MCP tool reconnects its server and
+  retries once within the run; a write is never auto-retried (it may already have
+  applied). One flapping server no longer nulls the shared client cache, so other
+  servers' cached clients are untouched. The SSE transport also gets a raised
+  body-timeout so a legitimate >1-min idle between the model's tool calls no
+  longer breaks a long-lived SSE socket (new `AI_MCP_SSE_BODY_TIMEOUT_MS`, default
+  10 min; see `.env.example`). (#489)
+
 - **The server no longer runs out of heap during long autonomous agent runs.** A
   new pnpm patch on `ai@6.0.134` stops the SDK from building a cumulative
   snapshot of the ENTIRE turn text on every streamed text-delta when no output
@@ -310,6 +361,39 @@ 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 +470,24 @@ 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/package.json

@@ -22,6 +22,7 @@
     "@casl/react": "5.0.1",
     "@docmost/editor-ext": "workspace:*",
     "@docmost/prosemirror-markdown": "workspace:*",
+    "@docmost/token-estimate": "workspace:*",
     "@excalidraw/excalidraw": "0.18.0-3a5ef40",
     "@mantine/core": "8.3.18",
     "@mantine/dates": "8.3.18",

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/count-stream-tokens.test.ts

@@ -6,10 +6,13 @@ describe("estimateTokens", () => {
     expect(estimateTokens("")).toBe(0);
   });
 
-  it("ceils chars/4 so any non-empty text is at least 1 token", () => {
+  // #490: migrated onto the shared @docmost/token-estimate module (chars/2.5, up
+  // from the old client-only chars/4) so the client counter and the server replay
+  // budgeter can never diverge.
+  it("ceils chars/2.5 so any non-empty text is at least 1 token", () => {
     expect(estimateTokens("a")).toBe(1);
-    expect(estimateTokens("abcd")).toBe(1);
-    expect(estimateTokens("abcde")).toBe(2);
-    expect(estimateTokens("12345678")).toBe(2);
+    expect(estimateTokens("ab")).toBe(1);
+    expect(estimateTokens("abcde")).toBe(2); // 5 / 2.5 = 2
+    expect(estimateTokens("x".repeat(10))).toBe(4); // 10 / 2.5 = 4
   });
 });

apps/client/src/features/ai-chat/utils/count-stream-tokens.ts

@@ -2,18 +2,10 @@
  * Rough client-side token estimation for AI-chat UI affordances.
  *
  * No provider streams exact per-token usage mid-stream, so any in-flight figure
- * is a CLIENT ESTIMATE (chars/≈4 heuristic). Pure + unit-testable: it never runs
- * a real BPE tokenizer (that would be O(n²) on the hot path, bloat the bundle,
- * and be wrong for Gemini/Ollama anyway). Used by the in-body reasoning counter
- * ("Thinking · N tokens").
+ * is a CLIENT ESTIMATE. This re-exports the SHARED estimator from
+ * `@docmost/token-estimate` (chars/2.5) so the in-body counter and the server's
+ * replay budgeter use the SAME heuristic — two divergent estimators would mean
+ * "the badge shows 60%" while "the budgeter already trimmed" (#490). Used by the
+ * in-body reasoning counter ("Thinking · N tokens").
  */
-
-/**
- * Rough token estimate for a piece of text using the standard chars/≈4 heuristic.
- * Returns 0 for empty/whitespace-free-of-content input, and ceils so any
- * non-empty text counts as at least one token.
- */
-export function estimateTokens(text: string): number {
-  if (!text) return 0;
-  return Math.ceil(text.length / 4);
-}
+export { estimateTokens } from "@docmost/token-estimate";

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/package.json

@@ -23,7 +23,7 @@
     "migration:reset": "tsx src/database/migrate.ts down-to NO_MIGRATIONS",
     "migration:codegen": "kysely-codegen --dialect=postgres --camel-case --env-file=../../.env --out-file=./src/database/types/db.d.ts",
     "lint": "eslint \"{src,apps,libs,test}/**/*.ts\" --fix",
-    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build",
+    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build && pnpm --filter @docmost/token-estimate build",
     "test": "jest",
     "test:int": "jest --config test/jest-integration.json",
     "test:watch": "jest --watch",
@@ -44,6 +44,7 @@
     "@docmost/mcp": "workspace:*",
     "@docmost/pdf-inspector": "1.9.6",
     "@docmost/prosemirror-markdown": "workspace:*",
+    "@docmost/token-estimate": "workspace:*",
     "@fastify/compress": "^9.0.0",
     "@fastify/cookie": "^11.0.2",
     "@fastify/multipart": "^10.0.0",
@@ -206,6 +207,7 @@
       "^@docmost/db/(.*)$": "<rootDir>/database/$1",
       "^@docmost/transactional/(.*)$": "<rootDir>/integrations/transactional/$1",
       "^@docmost/ee/(.*)$": "<rootDir>/ee/$1",
+      "^@docmost/token-estimate$": "<rootDir>/../../../packages/token-estimate/src/index.ts",
       "^src/(.*)$": "<rootDir>/$1",
       "^@tiptap/react$": "<rootDir>/../test/stubs/tiptap-react.js"
     }

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

@@ -43,6 +43,9 @@ function makeRepo(overrides: Record<string, jest.Mock> = {}) {
       workspaceId: v.workspaceId,
     })),
     update: jest.fn(async () => ({ id: 'run-1' })),
+    // #487: terminal finalize now goes through the CONDITIONAL write. Default
+    // returns a truthy row (the run WAS active -> this call wrote it).
+    finalizeIfActive: jest.fn(async () => ({ id: 'run-1', status: 'succeeded' })),
     markStopRequested: jest.fn(async () => ({ id: 'run-1' })),
     findActiveByChat: jest.fn(async () => undefined),
     findLatestByChat: jest.fn(async () => undefined),
@@ -336,14 +339,12 @@ describe('AiChatRunService run lifecycle', () => {
     await svc.finalizeRun('run-1', 'ws-1', 'error', 'provider blew up');
 
     expect(svc.isLocallyActive('run-1')).toBe(false);
-    expect(repo.update).toHaveBeenCalledWith(
+    // #487: the terminal write is CONDITIONAL (finalizeIfActive); finishedAt is
+    // stamped inside the repo method, so the service passes just status + error.
+    expect(repo.finalizeIfActive).toHaveBeenCalledWith(
       'run-1',
       'ws-1',
-      expect.objectContaining({
-        status: 'failed',
-        error: 'provider blew up',
-        finishedAt: expect.any(Date),
-      }),
+      expect.objectContaining({ status: 'failed', error: 'provider blew up' }),
     );
   });
 
@@ -366,8 +367,8 @@ describe('AiChatRunService run lifecycle', () => {
     // A second settle (e.g. a streamText callback firing after the catch) no-ops.
     await svc.finalizeRun('run-1', 'ws-1', 'completed', undefined);
 
-    expect(repo.update).toHaveBeenCalledTimes(1);
-    expect(repo.update).toHaveBeenCalledWith(
+    expect(repo.finalizeIfActive).toHaveBeenCalledTimes(1);
+    expect(repo.finalizeIfActive).toHaveBeenCalledWith(
       'run-1',
       'ws-1',
       expect.objectContaining({ status: 'failed', error: 'first' }),
@@ -389,8 +390,8 @@ describe('AiChatRunService run lifecycle', () => {
     const updateGate = new Promise((res) => {
       resolveUpdate = res;
     });
-    const update = jest.fn(() => updateGate);
-    const repo = makeRepo({ update });
+    const finalizeIfActive = jest.fn(() => updateGate);
+    const repo = makeRepo({ finalizeIfActive });
     const svc = new AiChatRunService(repo as never, makeEnv() as never);
     await svc.beginRun({
       chatId: 'chat-1',
@@ -399,23 +400,23 @@ describe('AiChatRunService run lifecycle', () => {
     });
 
     // Fire both before the (pending) update resolves. The first synchronously
-    // claims the entry (active.delete) and awaits update; the second, started in
-    // the same macrotask, finds the entry already gone and returns at the claim
-    // WITHOUT ever calling update.
+    // claims the entry (active.delete) and awaits the write; the second, started
+    // in the same macrotask, finds the entry already gone and returns at the claim
+    // WITHOUT ever writing.
     const p1 = svc.finalizeRun('run-1', 'ws-1', 'completed');
     const p2 = svc.finalizeRun('run-1', 'ws-1', 'error', 'safety-net');
 
     // The decisive assertion: exactly one caller reached the terminal UPDATE.
-    expect(update).toHaveBeenCalledTimes(1);
+    expect(finalizeIfActive).toHaveBeenCalledTimes(1);
 
     // Let the single in-flight update land; both calls resolve cleanly.
-    resolveUpdate({ id: 'run-1' });
+    resolveUpdate({ id: 'run-1', status: 'succeeded' });
     await Promise.all([p1, p2]);
 
-    expect(update).toHaveBeenCalledTimes(1);
+    expect(finalizeIfActive).toHaveBeenCalledTimes(1);
     // The winner is the FIRST caller ('completed' -> 'succeeded'); the late
     // 'error' settle never wrote, so it could not clobber the real status.
-    expect(update).toHaveBeenCalledWith(
+    expect(finalizeIfActive).toHaveBeenCalledWith(
       'run-1',
       'ws-1',
       expect.objectContaining({ status: 'succeeded' }),
@@ -431,10 +432,10 @@ describe('AiChatRunService run lifecycle', () => {
     // 409s until a restart. The fix updates FIRST and retries.
     let calls = 0;
     const repo = makeRepo({
-      update: jest.fn(async () => {
+      finalizeIfActive: jest.fn(async () => {
         calls += 1;
         if (calls === 1) throw new Error('deadlock detected');
-        return { id: 'run-1' };
+        return { id: 'run-1', status: 'succeeded' };
       }),
     });
     jest.spyOn(Logger.prototype, 'warn').mockImplementation(() => undefined);
@@ -447,26 +448,29 @@ describe('AiChatRunService run lifecycle', () => {
 
     await svc.finalizeRun('run-1', 'ws-1', 'completed');
 
-    // The retry landed the terminal write: the entry is dropped (slot freed) and
-    // the row carries the real terminal status — NOT stranded at 'running'.
+    // The retry landed the terminal write: the entry is dropped (slot freed), no
+    // zombie left, and the row carries the real terminal status.
     expect(svc.isLocallyActive('run-1')).toBe(false);
-    expect(repo.update).toHaveBeenCalledTimes(2);
-    expect(repo.update).toHaveBeenLastCalledWith(
+    expect(svc.hasZombie('run-1')).toBe(false);
+    expect(repo.finalizeIfActive).toHaveBeenCalledTimes(2);
+    expect(repo.finalizeIfActive).toHaveBeenLastCalledWith(
       'run-1',
       'ws-1',
       expect.objectContaining({ status: 'succeeded' }),
     );
   });
 
-  it('F6: if the terminal write keeps failing, the entry is RETAINED and a LATER settle completes it (chat not permanently 409d)', async () => {
+  it('#487 give-up: if the terminal write keeps failing, finalizeRun leaves a ZOMBIE (does NOT restore the entry) and settleZombie re-drives it', async () => {
     // Worst case: the DB is down for the whole first finalize (all attempts fail).
-    // The run must NOT be silently lost — the entry stays so a subsequent settle
-    // (a streamText callback, requestStop -> onAbort, or a future sweep) can retry.
+    // #487 changes the give-up behaviour: the entry is NOT restored (a restored
+    // entry is indistinguishable from a live run). Instead a ZOMBIE record holds
+    // the intended terminal status, and a re-drive (settleZombie — called by the
+    // reconcile / supersede / opportunistic paths) applies it later.
     let healthy = false;
     const repo = makeRepo({
-      update: jest.fn(async () => {
+      finalizeIfActive: jest.fn(async () => {
         if (!healthy) throw new Error('pool exhausted');
-        return { id: 'run-1' };
+        return { id: 'run-1', status: 'succeeded' };
       }),
     });
     jest.spyOn(Logger.prototype, 'warn').mockImplementation(() => undefined);
@@ -480,35 +484,83 @@ describe('AiChatRunService run lifecycle', () => {
       userId: 'user-1',
     });
 
-    // First settle: every bounded attempt fails -> entry retained, NOT settled.
+    // First settle: every bounded attempt fails -> ZOMBIE, entry NOT restored.
     await svc.finalizeRun('run-1', 'ws-1', 'completed');
-    expect(svc.isLocallyActive('run-1')).toBe(true);
-    // F12: the give-up emits ONE explicit, greppable ERROR (run + chat context)
-    // so an operator can tell "gave up, run held in memory" from a per-attempt
-    // blip — distinct from the per-attempt warns.
+    expect(svc.isLocallyActive('run-1')).toBe(false); // NOT a live entry
+    expect(svc.hasZombie('run-1')).toBe(true);
+    expect(svc.zombieRunIds()).toContain('run-1');
+    // The give-up emits ONE explicit, greppable ERROR mentioning the zombie.
     const gaveUp = errorSpy.mock.calls.some(
       (c) =>
         /NON-TERMINAL/.test(String(c[0])) &&
+        /ZOMBIE/.test(String(c[0])) &&
         /run-1/.test(String(c[0])) &&
         /chat-1/.test(String(c[0])),
     );
     expect(gaveUp).toBe(true);
+    // The settle notifier resolved as terminalWriteFailed (a subscriber learns the
+    // slot still needs the intended status applied).
+    const outcome = await svc.peekSettled('run-1');
+    expect(outcome).toEqual({
+      status: 'succeeded',
+      error: null,
+      terminalWriteFailed: true,
+    });
 
-    // The DB recovers; a later settle now succeeds and frees the slot.
+    // The DB recovers; a re-drive settles the zombie via the conditional UPDATE.
     healthy = true;
-    await svc.finalizeRun('run-1', 'ws-1', 'completed');
-    expect(svc.isLocallyActive('run-1')).toBe(false);
-    expect(repo.update).toHaveBeenLastCalledWith(
+    const redriven = await svc.settleZombie('run-1');
+    expect(redriven).toBe(true);
+    expect(svc.hasZombie('run-1')).toBe(false);
+    expect(repo.finalizeIfActive).toHaveBeenLastCalledWith(
       'run-1',
       'ws-1',
       expect.objectContaining({ status: 'succeeded' }),
     );
 
-    // And it is now idempotent: a further settle no-ops (terminal row already
-    // written), so a double-settle can never clobber the real status.
-    const callsBefore = repo.update.mock.calls.length;
+    // A later finalizeRun is idempotent (row already terminal): it no-ops at the
+    // once-gate, never re-writing.
+    const callsBefore = repo.finalizeIfActive.mock.calls.length;
     await svc.finalizeRun('run-1', 'ws-1', 'error', 'late');
-    expect(repo.update).toHaveBeenCalledTimes(callsBefore);
+    expect(repo.finalizeIfActive).toHaveBeenCalledTimes(callsBefore);
+  });
+
+  it('#487 double-settle collapses to a benign no-op (conditional write; notifier resolves once)', async () => {
+    // A second concurrent settle is stopped at the synchronous active.delete
+    // claim, so the terminal write runs exactly once and the notifier resolves
+    // exactly once with the FIRST settler's outcome.
+    const repo = makeRepo();
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    await svc.beginRun({ chatId: 'chat-1', workspaceId: 'ws-1', userId: 'u1' });
+
+    await svc.finalizeRun('run-1', 'ws-1', 'aborted');
+    await svc.finalizeRun('run-1', 'ws-1', 'error', 'late'); // no-op
+
+    expect(repo.finalizeIfActive).toHaveBeenCalledTimes(1);
+    const outcome = await svc.peekSettled('run-1');
+    // peekSettled after resolve+delete falls through (notifier dropped, no zombie)
+    // -> undefined; the FIRST settler already resolved any earlier subscriber.
+    expect(outcome).toBeUndefined();
+  });
+
+  it('#487 late settledPromise subscriber gets the resolved outcome', async () => {
+    const repo = makeRepo();
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    await svc.beginRun({ chatId: 'chat-1', workspaceId: 'ws-1', userId: 'u1' });
+
+    // Subscribe BEFORE settle: hold the promise reference (as supersede does).
+    const early = svc.peekSettled('run-1');
+    expect(early).toBeDefined();
+
+    await svc.finalizeRun('run-1', 'ws-1', 'completed');
+
+    // The reference grabbed before settle resolves with the written outcome, even
+    // though the notifier was dropped from the map on resolve (bounded).
+    await expect(early).resolves.toEqual({
+      status: 'succeeded',
+      error: null,
+      terminalWriteFailed: false,
+    });
   });
 
   it('recordStep / linkAssistantMessage are best-effort: a repo failure is swallowed', async () => {
@@ -525,3 +577,197 @@ describe('AiChatRunService run lifecycle', () => {
     ).resolves.toBeUndefined();
   });
 });
+
+describe('#487 AiChatRunService.supersede (CAS)', () => {
+  const chat = 'chat-1';
+  const ws = 'ws-1';
+
+  it('degrade: no active run on the chat -> caller sends a normal turn', async () => {
+    const repo = makeRepo({
+      findById: jest.fn(async () => undefined),
+      findActiveByChat: jest.fn(async () => undefined),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    expect(await svc.supersede(chat, 'run-x', ws)).toEqual({ kind: 'degrade' });
+  });
+
+  it('invalid: the target run belongs to a DIFFERENT chat -> 400', async () => {
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({
+        id: 'run-x',
+        chatId: 'other-chat',
+        workspaceId: ws,
+      })),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    expect(await svc.supersede(chat, 'run-x', ws)).toEqual({ kind: 'invalid' });
+  });
+
+  it('mismatch: a DIFFERENT run is active than the one targeted -> current runId', async () => {
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({ id: 'run-x', chatId: chat, workspaceId: ws })),
+      findActiveByChat: jest.fn(async () => ({
+        id: 'run-live',
+        chatId: chat,
+        workspaceId: ws,
+        status: 'running',
+      })),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    expect(await svc.supersede(chat, 'run-x', ws)).toEqual({
+      kind: 'mismatch',
+      activeRunId: 'run-live',
+    });
+  });
+
+  it('ready: the target IS active -> stop it, await its (fast) settle, free the slot', async () => {
+    // Simulate a live long TOOL (NOT a slow UPDATE): the run stays active until an
+    // explicit Stop unwinds it; commit-1's race makes that settle land quickly.
+    // The abort listener stands in for streamText's onAbort -> finalizeRun.
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({
+        id: 'run-1',
+        chatId: chat,
+        workspaceId: ws,
+        status: 'aborted',
+        error: null,
+      })),
+      findActiveByChat: jest.fn(async () => ({
+        id: 'run-1',
+        chatId: chat,
+        workspaceId: ws,
+        status: 'running',
+      })),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    const handle = await svc.beginRun({ chatId: chat, workspaceId: ws, userId: 'u1' });
+    handle.signal.addEventListener('abort', () => {
+      void svc.finalizeRun('run-1', ws, 'aborted');
+    });
+
+    // supersede: getRun -> getActiveByChat(==target) -> requestStop -> the abort
+    // listener settles the run -> awaitSettled resolves -> ready.
+    expect(await svc.supersede(chat, 'run-1', ws, 10_000)).toEqual({
+      kind: 'ready',
+    });
+    expect(handle.signal.aborted).toBe(true); // Stop reached the run
+  });
+
+  it('timeout: the target never settles within W -> 409 SUPERSEDE_TIMEOUT (nothing persisted)', async () => {
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({ id: 'run-1', chatId: chat, workspaceId: ws })),
+      findActiveByChat: jest.fn(async () => ({
+        id: 'run-1',
+        chatId: chat,
+        workspaceId: ws,
+        status: 'running',
+      })),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    await svc.beginRun({ chatId: chat, workspaceId: ws, userId: 'u1' });
+    // Do NOT settle the run: a tiny W elapses -> timeout.
+    const result = await svc.supersede(chat, 'run-1', ws, 30);
+    expect(result).toEqual({ kind: 'timeout' });
+  });
+
+  it('ready then a DUPLICATE supersede POST degrades (the run is already gone)', async () => {
+    let active: unknown = {
+      id: 'run-1',
+      chatId: chat,
+      workspaceId: ws,
+      status: 'running',
+    };
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({
+        id: 'run-1',
+        chatId: chat,
+        workspaceId: ws,
+        status: 'aborted',
+        error: null,
+      })),
+      findActiveByChat: jest.fn(async () => active),
+      finalizeIfActive: jest.fn(async () => {
+        active = undefined; // settling frees the active slot
+        return { id: 'run-1', status: 'aborted' };
+      }),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    const handle = await svc.beginRun({ chatId: chat, workspaceId: ws, userId: 'u1' });
+    handle.signal.addEventListener('abort', () => {
+      void svc.finalizeRun('run-1', ws, 'aborted');
+    });
+
+    expect(await svc.supersede(chat, 'run-1', ws, 10_000)).toEqual({
+      kind: 'ready',
+    });
+    // The duplicate POST for the same target now finds no active run -> degrade.
+    expect(await svc.supersede(chat, 'run-1', ws)).toEqual({ kind: 'degrade' });
+  });
+
+  it('reconcileStaleRuns: aborts a stale run with NO entry/zombie; NEVER touches a live entry', async () => {
+    const finalizeIfActive = jest.fn(async () => ({ id: 'x', status: 'aborted' }));
+    const repo = makeRepo({
+      insert: jest.fn(async (v: any) => ({
+        id: 'live-1',
+        status: 'running',
+        chatId: v.chatId,
+        workspaceId: v.workspaceId,
+      })),
+      finalizeIfActive,
+      findStaleActive: jest.fn(async () => [
+        { id: 'orphan-1', workspaceId: ws, chatId: 'c-orphan' },
+        { id: 'live-1', workspaceId: ws, chatId: 'c-live' },
+      ]),
+    });
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    // A LIVE run this replica owns (in the `active` map).
+    await svc.beginRun({ chatId: 'c-live', workspaceId: ws, userId: 'u1' });
+    expect(svc.isLocallyActive('live-1')).toBe(true);
+
+    const aborted = await svc.reconcileStaleRuns(15 * 60 * 1000);
+    expect(aborted).toBe(1);
+    // The orphan (no entry) was aborted; the live entry was NEVER passed to the DB.
+    expect(finalizeIfActive).toHaveBeenCalledTimes(1);
+    expect(finalizeIfActive).toHaveBeenCalledWith(
+      'orphan-1',
+      ws,
+      expect.objectContaining({ status: 'aborted' }),
+    );
+    expect(svc.isLocallyActive('live-1')).toBe(true);
+  });
+
+  it('gave-up zombie: supersede applies the intended status (settleZombie) then is ready', async () => {
+    let healthy = false;
+    let active: unknown = {
+      id: 'run-1',
+      chatId: chat,
+      workspaceId: ws,
+      status: 'running',
+    };
+    const repo = makeRepo({
+      findById: jest.fn(async () => ({ id: 'run-1', chatId: chat, workspaceId: ws })),
+      findActiveByChat: jest.fn(async () => active),
+      finalizeIfActive: jest.fn(async () => {
+        if (!healthy) throw new Error('db down');
+        active = undefined;
+        return { id: 'run-1', status: 'aborted' };
+      }),
+    });
+    jest.spyOn(Logger.prototype, 'warn').mockImplementation(() => undefined);
+    jest.spyOn(Logger.prototype, 'error').mockImplementation(() => undefined);
+    const svc = new AiChatRunService(repo as never, makeEnv() as never);
+    await svc.beginRun({ chatId: chat, workspaceId: ws, userId: 'u1' });
+
+    // The run's terminal write gives up -> zombie (row still 'running').
+    await svc.finalizeRun('run-1', ws, 'aborted');
+    expect(svc.hasZombie('run-1')).toBe(true);
+
+    // The DB recovers; supersede awaits the (already-resolved, terminalWriteFailed)
+    // settle, then settleZombie applies the intended status -> ready.
+    healthy = true;
+    expect(await svc.supersede(chat, 'run-1', ws, 10_000)).toEqual({
+      kind: 'ready',
+    });
+    expect(svc.hasZombie('run-1')).toBe(false);
+  });
+});

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

@@ -34,6 +34,88 @@ export class RunAlreadyActiveError extends Error {
 export type TurnTerminalStatus = 'completed' | 'error' | 'aborted';
 export type RunTerminalStatus = 'succeeded' | 'failed' | 'aborted';
 
+/** The terminal run statuses — the row is done once it reads one of these. */
+export const RUN_TERMINAL_STATUSES: readonly RunTerminalStatus[] = [
+  'succeeded',
+  'failed',
+  'aborted',
+];
+
+/** Whether a persisted run status is terminal (settled). */
+export function isRunTerminal(status: string | null | undefined): boolean {
+  return (
+    status === 'succeeded' || status === 'failed' || status === 'aborted'
+  );
+}
+
+/**
+ * #487: the outcome a run's {@link AiChatRunService.finalizeRun} settled with.
+ * `terminalWriteFailed` = the terminal write GAVE UP after the bounded retry, so
+ * the row is still non-terminal ('running') and a ZOMBIE record holds the
+ * `intended` status for a later re-drive (reconcile / supersede / boot sweep). A
+ * subscriber (supersede, #487 commit 3) uses this to decide whether the slot is
+ * genuinely free or must first have the intended status applied.
+ */
+export interface RunSettleOutcome {
+  status: RunTerminalStatus;
+  error: string | null;
+  terminalWriteFailed: boolean;
+}
+
+/**
+ * #487: how long a supersede waits for the target run to settle after Stop before
+ * it degrades to `SUPERSEDE_TIMEOUT`. W=10s is generous under a HEALTHY DB: commit
+ * 1's race-on-abort makes an in-app tool abort->settle in ms/hundreds of ms, so a
+ * live run releases its slot well within the window. Under a DB brownout the
+ * timeout is normal (the write cannot land); W must NOT be raised to paper
+ * over a slow DB — a SUPERSEDE_TIMEOUT is the honest signal (nothing persisted,
+ * the composer keeps the user's text). Env-tunable for ops, default 10s.
+ */
+export const SUPERSEDE_SETTLE_TIMEOUT_MS = (() => {
+  const raw = Number(process.env.AI_CHAT_SUPERSEDE_TIMEOUT_MS);
+  return Number.isFinite(raw) && raw > 0 ? raw : 10_000;
+})();
+
+/**
+ * #487: the result of the supersede CAS ({@link AiChatRunService.supersede}).
+ *  - `degrade`  : no active run on the chat (it ended between click and POST) —
+ *                 the caller sends a NORMAL turn (NOT a mismatch);
+ *  - `invalid`  : the target runId belongs to a DIFFERENT chat (malformed CAS 400);
+ *  - `mismatch` : a DIFFERENT run is active than the one the client targeted —
+ *                 409 SUPERSEDE_TARGET_MISMATCH carrying the current `activeRunId`
+ *                 (the client does NOT auto-retry);
+ *  - `timeout`  : the target did not settle within W — 409 SUPERSEDE_TIMEOUT,
+ *                 nothing persisted;
+ *  - `ready`    : the target was stopped AND settled (or its zombie's intended was
+ *                 applied) — the slot is free; the caller may beginRun the new run.
+ */
+export type SupersedeResult =
+  | { kind: 'degrade' }
+  | { kind: 'invalid' }
+  | { kind: 'mismatch'; activeRunId: string }
+  | { kind: 'timeout' }
+  | { kind: 'ready' };
+
+/** A one-shot settle notifier (#487): `resolve` is called EXACTLY ONCE. */
+interface Deferred<T> {
+  promise: Promise<T>;
+  resolve: (value: T) => void;
+}
+
+/**
+ * #487: a run whose terminal write GAVE UP (every bounded attempt failed). The
+ * row is stranded non-terminal ('running'); this record is the ONLY thing that
+ * distinguishes it from a live run, and carries the `intended` terminal status so
+ * a re-drive can apply it via the conditional UPDATE. Process-local (phase-1
+ * single-process assumption): a restart drops it, and the boot sweep then writes
+ * 'aborted' over the intended — a documented loss (see finalizeRun).
+ */
+interface ZombieRun {
+  workspaceId: string;
+  chatId: string;
+  intended: { status: RunTerminalStatus; error: string | null };
+}
+
 export function mapTurnStatusToRun(
   status: TurnTerminalStatus,
 ): RunTerminalStatus {
@@ -101,6 +183,22 @@ export class AiChatRunService implements OnModuleInit {
   // uptime — negligible in phase 1's single process.
   private readonly settled = new Set<string>();
 
+  // #487 runId -> one-shot settle notifier. Kept in a SEPARATE map from `active`
+  // ON PURPOSE: it must OUTLIVE the `active.delete` claim inside finalizeRun (the
+  // claim frees the slot the instant finalize starts), so a subscriber can still
+  // await the outcome after the entry is gone. Created in beginRun, resolved
+  // EXACTLY ONCE in finalizeRun, then removed (bounded). Absence => this replica
+  // has no live notifier: a subscriber falls back to the zombie map, then to the
+  // row (see peekSettled). Process-local (phase-1 single-process assumption).
+  private readonly settledPromises = new Map<string, Deferred<RunSettleOutcome>>();
+
+  // #487 runId -> ZOMBIE record: a run whose terminal write gave up (row stranded
+  // non-terminal). BOUNDED — an entry is added only on give-up and removed on a
+  // successful re-drive (settleZombie) or when the row is found already terminal;
+  // a process restart clears it (and the boot sweep settles the stranded row).
+  // Process-local (phase-1 single-process assumption).
+  private readonly zombies = new Map<string, ZombieRun>();
+
   // Bounded retry for the terminal write (F6): a single PK UPDATE can fail
   // transiently under many fire-and-forget writes (pool exhaustion, deadlock, a
   // brief connection blip). Riding out that blip in-place matters because the
@@ -224,6 +322,10 @@ export class AiChatRunService implements OnModuleInit {
       chatId: args.chatId,
       workspaceId: args.workspaceId,
     });
+    // #487: arm the one-shot settle notifier BEFORE returning, so a subscriber
+    // that races in immediately after begin always finds a promise to await. It
+    // is resolved exactly once when the run settles (or gives up).
+    this.settledPromises.set(run.id, this.makeDeferred<RunSettleOutcome>());
     return { runId: run.id, signal: controller.signal };
   }
 
@@ -263,47 +365,43 @@ export class AiChatRunService implements OnModuleInit {
   }
 
   /**
-   * Finalize a run to its terminal status (succeeded / failed / aborted),
-   * stamping finishedAt + any error. Best-effort, but ROBUST against a transient
-   * terminal-write failure (F6) AND atomically safe against a concurrent settle.
+   * Finalize a run to its terminal status (succeeded / failed / aborted) via a
+   * CONDITIONAL UPDATE, stamping finishedAt + any error. Atomically safe against a
+   * concurrent settle AND robust against a transient terminal-write failure.
    *
    * ATOMIC ONCE-CLAIM (the gate must close in ONE synchronous tick): two
    * finalizeRun calls for the SAME run can race — the documented real path is
    * AiChatService.stream's safety-net catch settling the turn to 'error' while a
    * streamText terminal callback (onFinish/onAbort/onError) ALSO settles it. The
-   * `settled.has` check alone is NOT a gate: it is read BEFORE the awaited UPDATE,
-   * so two callers can both see `false` and both write the row (last-write-wins
-   * clobbers the real terminal status, and the bounded retry only widens that
-   * window). The claim therefore happens via `active.delete`, a SYNCHRONOUS
-   * check-and-clear with NO await between the gate and the entry removal: the
-   * second concurrent caller finds the entry already gone and returns in the same
-   * tick, before any UPDATE. The transition "nobody is finalizing" -> "I am
-   * finalizing" is thus a single atomic step.
+   * claim happens via `active.delete`, a SYNCHRONOUS check-and-clear with NO await
+   * between the gate and the entry removal: the second concurrent caller finds the
+   * entry already gone and returns in the same tick, before any UPDATE.
    *
-   * ORDER MATTERS (F6): once we own the claim, the terminal UPDATE happens FIRST;
-   * only once it SUCCEEDS do we record the run as settled. If the UPDATE fails on
-   * every bounded attempt we RESTORE the in-memory entry, leave the run UNsettled,
-   * and emit an ERROR signal that the row is left non-terminal 'running' (which
-   * would 409 every future turn in the chat until recovery). An in-process retry
-   * by a LATER settle is only POSSIBLE, never guaranteed: it needs (a) the entry
-   * to have been restored at the give-up path AND (b) a fresh settler to arrive
-   * AFTER that restore. A concurrent settler that arrives DURING the retry window
-   * — while the entry is deleted for backoff and not yet restored — is consumed at
-   * the synchronous `active.delete` claim (it finds nothing to delete and returns
-   * a no-op), so it does NOT become an in-process retrier. The NO-streamText path
-   * (the turn threw before streamText was wired, so ONLY the safety-net ever
-   * settles) likewise has no second in-process settler at all. The UNCONDITIONAL
-   * backstop in every case is the boot sweep on the next restart (phase 1 has no
-   * periodic in-process sweep); the retained entry is bounded (cleared on restart)
-   * and harmless meanwhile.
+   * ALL TERMINAL WRITES ARE CONDITIONAL (#487): `finalizeIfActive` only flips a
+   * row still in pending|running (mirror of the assistant message's
+   * `onlyIfStreaming`). So even a settle that DID reach the UPDATE (e.g. a
+   * reconcile stamp racing an owner finalize) can never clobber a terminal status
+   * — the loser matches nothing and is a benign no-op. `active.delete` is the
+   * fast, in-process gate; the conditional WHERE is the authoritative one.
    *
-   * IDEMPOTENT on SUCCESS (#184 review): the terminal write happens AT MOST ONCE
-   * per run. After a successful write the once-gate keys off {@link settled} (the
-   * terminal row already written) so a settle arriving AFTER the entry was already
-   * dropped-and-settled returns early; a settle racing the in-flight write is
-   * stopped earlier still, by the `active.delete` claim. Either way a genuine
-   * double-settle collapses to a single write and a late settle can never clobber
-   * the real terminal status or double-write the row.
+   * ZOMBIE ON GIVE-UP (#487): if every bounded attempt THROWS (the DB is down for
+   * the whole finalize), we do NOT restore the entry. The row is stranded
+   * non-terminal ('running'); we record a ZOMBIE `{ terminalWriteFailed, intended
+   * }` (the ONLY thing distinguishing this dead run from a live one) and resolve
+   * the settle notifier with `terminalWriteFailed: true`. A restore would make the
+   * zombie indistinguishable from a live run to every reader; instead a re-drive
+   * (settleZombie, called by the periodic reconcile / supersede / opportunistic
+   * paths) applies the intended status later via the same conditional UPDATE.
+   *
+   * DOCUMENTED LOSS (#487, single-process phase 1): if the process RESTARTS before
+   * a zombie is re-driven, the in-memory zombie map is gone and the boot sweep
+   * (unconditional) writes 'aborted' over the ACTUAL intended status. This is
+   * unavoidable while the run lifecycle is single-process — there is no durable
+   * record of `intended`; a cross-process durable intent is deferred to phase 2.
+   *
+   * IDEMPOTENT: the settle notifier resolves EXACTLY ONCE; a second settle is
+   * stopped at `settled.has` or the `active.delete` claim, so a double-settle
+   * collapses to a single write and can never double-resolve or clobber the row.
    */
   async finalizeRun(
     runId: string,
@@ -314,13 +412,17 @@ export class AiChatRunService implements OnModuleInit {
     // ---- Atomic once-claim (synchronous; NO await before the gate closes) ----
     // Already terminally written -> idempotent no-op.
     if (this.settled.has(runId)) return;
-    // Capture the entry BEFORE the delete so a total-failure path can restore it.
+    // Capture the entry BEFORE the delete for the give-up log context.
     const entry = this.active.get(runId);
     // SYNCHRONOUS check-and-clear: the FIRST caller deletes (claims) the entry;
     // any concurrent SECOND caller finds nothing to delete and returns HERE, in
     // the same tick, before any await — so it can never reach the UPDATE.
     if (!this.active.delete(runId)) return;
 
+    const status = mapTurnStatusToRun(turnStatus);
+    const err = error ?? null;
+    const chatId = entry?.chatId ?? 'unknown';
+
     let lastError: unknown;
     for (
       let attempt = 1;
@@ -328,47 +430,294 @@ export class AiChatRunService implements OnModuleInit {
       attempt++
     ) {
       try {
-        await this.runRepo.update(runId, workspaceId, {
-          status: mapTurnStatusToRun(turnStatus),
-          finishedAt: new Date(),
-          error: error ?? null,
+        const row = await this.runRepo.finalizeIfActive(runId, workspaceId, {
+          status,
+          error: err,
         });
-        // Terminal write landed: arm the once-gate. The entry is already gone
-        // (claimed above); we do NOT restore it. The slot is now free.
+        // No throw => the row is now terminal (we wrote it, or it was ALREADY
+        // terminal — another writer won the conditional UPDATE, a benign no-op).
         this.settled.add(runId);
+        this.zombies.delete(runId);
+        // Resolve with the persisted outcome: our status when WE wrote it, else
+        // the row's real terminal status (re-read on the already-terminal path so
+        // a subscriber never sees a status we did not actually persist).
+        const outcome: RunSettleOutcome = row
+          ? { status, error: err, terminalWriteFailed: false }
+          : await this.readTerminalOutcome(runId, workspaceId, status, err);
+        this.resolveSettled(runId, outcome);
         return;
-      } catch (err) {
-        lastError = err;
+      } catch (err2) {
+        lastError = err2;
         this.logger.warn(
           `Failed to finalize run ${runId} (attempt ${attempt}/${
             AiChatRunService.FINALIZE_MAX_ATTEMPTS
-          }): ${err instanceof Error ? err.message : 'unknown error'}`,
+          }): ${err2 instanceof Error ? err2.message : 'unknown error'}`,
         );
         if (attempt < AiChatRunService.FINALIZE_MAX_ATTEMPTS) {
           await this.delay(AiChatRunService.FINALIZE_RETRY_BASE_MS * attempt);
         }
       }
     }
-    // Every attempt failed: this is a give-up, materially worse than a per-attempt
-    // blip — the row is left NON-TERMINAL ('running'), so emit ONE explicit,
-    // greppable ERROR so an operator can tell "survived a blip" from "gave up, run
-    // held in memory until recovery" (the last warn alone says only "attempt 3/3").
+    // Every attempt threw: GIVE UP. The row is stranded non-terminal ('running').
+    // Do NOT restore the entry (a restored entry is indistinguishable from a live
+    // run); leave a ZOMBIE record instead, and resolve the notifier as
+    // terminalWriteFailed so a subscriber knows the slot still needs the intended
+    // status applied. One explicit, greppable ERROR so an operator can tell a
+    // give-up from a per-attempt blip.
     this.logger.error(
-      `Run ${runId} (chat ${entry?.chatId ?? 'unknown'}) left NON-TERMINAL ` +
-        `('running'): terminal write failed after ${
-          AiChatRunService.FINALIZE_MAX_ATTEMPTS
-        } attempts; entry retained in memory, recovery deferred to next settle / ` +
-        `boot sweep`,
+      `Run ${runId} (chat ${chatId}) left NON-TERMINAL ('running'): terminal ` +
+        `write failed after ${AiChatRunService.FINALIZE_MAX_ATTEMPTS} attempts; ` +
+        `ZOMBIE recorded (intended '${status}'), recovery deferred to reconcile / ` +
+        `supersede / boot sweep`,
       lastError,
     );
-    // RESTORE the claimed entry (and leave the run UNsettled) so a LATER settle
-    // that arrives AFTER this restore MAY retry the terminal write — but that
-    // in-process retry is NOT guaranteed (a concurrent settler caught in the retry
-    // window above is consumed at the `active.delete` claim, and the no-streamText
-    // path has no second settler at all). The UNCONDITIONAL backstop in every case
-    // is the boot sweep on the next restart; the restored entry is bounded and
-    // cleared on restart.
-    if (entry) this.active.set(runId, entry);
+    this.zombies.set(runId, {
+      workspaceId,
+      chatId,
+      intended: { status, error: err },
+    });
+    this.resolveSettled(runId, { status, error: err, terminalWriteFailed: true });
+  }
+
+  /**
+   * #487: re-drive a zombie run's intended terminal write (the conditional
+   * UPDATE). Called by the periodic reconcile (commit 4), an opportunistic
+   * single-chat reconcile, and supersede (commit 3). On success — the row is now
+   * terminal (written OR found already terminal) — the zombie is cleared and the
+   * once-gate armed; on another failure the zombie is kept for a later retry.
+   * Returns true when the row is now terminal. Best-effort; never throws.
+   */
+  async settleZombie(runId: string): Promise<boolean> {
+    const z = this.zombies.get(runId);
+    if (!z) return false;
+    try {
+      await this.runRepo.finalizeIfActive(runId, z.workspaceId, {
+        status: z.intended.status,
+        error: z.intended.error,
+      });
+      this.zombies.delete(runId);
+      this.settled.add(runId);
+      return true;
+    } catch (err) {
+      this.logger.warn(
+        `Re-drive of zombie run ${runId} (chat ${z.chatId}) failed; will retry ` +
+          `later: ${err instanceof Error ? err.message : 'unknown error'}`,
+      );
+      return false;
+    }
+  }
+
+  /**
+   * #487 reconcile clause (c): abort runs the DB still shows active (pending|
+   * running) but that this replica does NOT own — NO live entry AND NO zombie —
+   * and that have been UNTOUCHED past `staleMs` (from last-progress `updated_at`,
+   * NOT startedAt, so a legit long marathon is never a candidate). "No entry" is
+   * the PRIMARY gate: a live entry (an actively-executing run on this replica) is
+   * NEVER aborted, whatever its age. Returns the number aborted. Best-effort —
+   * never throws (a periodic-job failure must not crash the process).
+   */
+  async reconcileStaleRuns(staleMs: number): Promise<number> {
+    let candidates: Array<{ id: string; workspaceId: string; chatId: string }>;
+    try {
+      candidates = await this.runRepo.findStaleActive(staleMs);
+    } catch (err) {
+      this.logger.warn(
+        `Reconcile (stale runs) query failed: ${
+          err instanceof Error ? err.message : 'unknown error'
+        }`,
+      );
+      return 0;
+    }
+    let aborted = 0;
+    for (const c of candidates) {
+      // PRIMARY gate: never touch a live entry, and never race a zombie we are
+      // already re-driving (settleZombie owns those).
+      if (this.active.has(c.id) || this.zombies.has(c.id)) continue;
+      try {
+        const row = await this.runRepo.finalizeIfActive(c.id, c.workspaceId, {
+          status: 'aborted',
+          error: 'Run aborted by reconcile: no live runner (stale).',
+        });
+        if (row) {
+          aborted += 1;
+          this.settled.add(c.id);
+        }
+      } catch (err) {
+        this.logger.warn(
+          `Reconcile abort of stale run ${c.id} failed: ${
+            err instanceof Error ? err.message : 'unknown error'
+          }`,
+        );
+      }
+    }
+    return aborted;
+  }
+
+  /**
+   * #487: the run's settle outcome as seen by THIS replica, or undefined when it
+   * has no record (the caller then reads the row — the DB is the source of truth).
+   * A LIVE deferred (still settling, or resolved-but-not-yet-consumed) wins; a
+   * ZOMBIE synthesizes the give-up outcome. A subscriber (supersede) races this
+   * against a timeout.
+   */
+  peekSettled(runId: string): Promise<RunSettleOutcome> | undefined {
+    const d = this.settledPromises.get(runId);
+    if (d) return d.promise;
+    const z = this.zombies.get(runId);
+    if (z) {
+      return Promise.resolve({
+        status: z.intended.status,
+        error: z.intended.error,
+        terminalWriteFailed: true,
+      });
+    }
+    return undefined;
+  }
+
+  /**
+   * #487: await a run's settle outcome, bounded by `timeoutMs`. Returns the
+   * outcome on settle, or undefined on TIMEOUT (or when this replica has no record
+   * of the run and its row is not terminal). Uses the LIVE settle notifier / the
+   * zombie synth when present; else reads the row (the DB is the source of truth
+   * once the in-memory record is gone). The subscriber (supersede) grabs this
+   * right after Stop; commit 1's race makes the settle land in ms on a healthy DB.
+   */
+  async awaitSettled(
+    runId: string,
+    workspaceId: string,
+    timeoutMs: number,
+  ): Promise<RunSettleOutcome | undefined> {
+    const pending = this.peekSettled(runId);
+    if (pending) {
+      let timer: ReturnType<typeof setTimeout> | undefined;
+      const timeout = new Promise<undefined>((resolve) => {
+        timer = setTimeout(() => resolve(undefined), timeoutMs);
+        timer.unref?.();
+      });
+      try {
+        return await Promise.race([pending, timeout]);
+      } finally {
+        if (timer) clearTimeout(timer);
+      }
+    }
+    // No live notifier and no zombie: read the row (already settled-and-written,
+    // or unknown here). A terminal row is an outcome; anything else -> undefined.
+    const row = await this.runRepo.findById(runId, workspaceId);
+    if (row && isRunTerminal(row.status)) {
+      return {
+        status: row.status as RunTerminalStatus,
+        error: row.error ?? null,
+        terminalWriteFailed: false,
+      };
+    }
+    return undefined;
+  }
+
+  /**
+   * #487: the SERVER supersede CAS for `POST /stream { supersede: { runId: X } }`.
+   * Atomically transitions "X is the chat's active run" -> "X is stopped, settled,
+   * slot free" so the caller can start a replacement run. See {@link
+   * SupersedeResult} for the branch semantics.
+   *
+   * On a `ready` result the caller MUST still go through the normal beginRun gate
+   * (the partial unique index) — between the slot freeing here and beginRun a
+   * neighbouring tab's ordinary POST can win the slot (documented SLOT-THEFT: the
+   * loser then gets a MISMATCH carrying the NEW runId). There is also NO side-
+   * effect quiescence: an in-flight write of the stopped run may still land AFTER
+   * the new run starts (commit 1 stops the NEXT call, not one already committing),
+   * so the caller adds a prompt note to the new run.
+   */
+  async supersede(
+    chatId: string,
+    targetRunId: string,
+    workspaceId: string,
+    timeoutMs: number = SUPERSEDE_SETTLE_TIMEOUT_MS,
+  ): Promise<SupersedeResult> {
+    // Validate the target belongs to THIS chat (a CAS targeting another chat's run
+    // is malformed -> 400). A missing row is NOT invalid: the run may have ended
+    // and been pruned; the active-run check below decides degrade vs mismatch.
+    const target = await this.getRun(targetRunId, workspaceId);
+    if (target && target.chatId !== chatId) return { kind: 'invalid' };
+
+    const active = await this.getActiveForChat(chatId, workspaceId);
+    // No active run: it ended between the client's click and this POST — this is a
+    // DEGRADE to a normal send, NOT a mismatch (the user's intent still holds).
+    if (!active) return { kind: 'degrade' };
+    // A DIFFERENT run is active than the one the client saw -> mismatch. The
+    // client does not auto-retry; it surfaces the new runId.
+    if (active.id !== targetRunId) {
+      return { kind: 'mismatch', activeRunId: active.id };
+    }
+
+    // The target IS active: stop it, then await its settle within W.
+    await this.requestStop(targetRunId, workspaceId);
+    const outcome = await this.awaitSettled(targetRunId, workspaceId, timeoutMs);
+    if (!outcome) return { kind: 'timeout' };
+    // Gave up (terminal write failed): apply the intended status via the
+    // conditional UPDATE so the slot actually frees. If that ALSO fails, the row
+    // is still stranded -> treat as a timeout (nothing persisted for the new run).
+    if (outcome.terminalWriteFailed) {
+      const settled = await this.settleZombie(targetRunId);
+      if (!settled) return { kind: 'timeout' };
+    }
+    return { kind: 'ready' };
+  }
+
+  /** #487 test/diagnostic seam: whether a give-up zombie is held for this run. */
+  hasZombie(runId: string): boolean {
+    return this.zombies.has(runId);
+  }
+
+  /** #487: every zombie runId held on this replica (reconcile clause a, commit 4). */
+  zombieRunIds(): string[] {
+    return [...this.zombies.keys()];
+  }
+
+  /** #487: create a one-shot deferred (resolve captured for a later single call). */
+  private makeDeferred<T>(): Deferred<T> {
+    let resolve!: (value: T) => void;
+    const promise = new Promise<T>((r) => {
+      resolve = r;
+    });
+    return { promise, resolve };
+  }
+
+  /** #487: resolve a run's settle notifier EXACTLY ONCE, then drop it (bounded).
+   *  A subscriber that already grabbed the promise still resolves; a later one
+   *  falls back to the zombie map / the row (see peekSettled). */
+  private resolveSettled(runId: string, outcome: RunSettleOutcome): void {
+    const d = this.settledPromises.get(runId);
+    if (!d) return;
+    this.settledPromises.delete(runId);
+    d.resolve(outcome);
+  }
+
+  /** #487: read the persisted terminal outcome when the conditional finalize was a
+   *  no-op (the row was already terminal). Falls back to the intended status when
+   *  the read fails or the row is unexpectedly missing/non-terminal. */
+  private async readTerminalOutcome(
+    runId: string,
+    workspaceId: string,
+    fallbackStatus: RunTerminalStatus,
+    fallbackError: string | null,
+  ): Promise<RunSettleOutcome> {
+    try {
+      const row = await this.runRepo.findById(runId, workspaceId);
+      if (row && isRunTerminal(row.status)) {
+        return {
+          status: row.status as RunTerminalStatus,
+          error: row.error ?? null,
+          terminalWriteFailed: false,
+        };
+      }
+    } catch {
+      // Fall through to the intended status — best-effort only.
+    }
+    return {
+      status: fallbackStatus,
+      error: fallbackError,
+      terminalWriteFailed: false,
+    };
   }
 
   /** Small async backoff between terminal-write retries (F6). Isolated so it is

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.controller.export.spec.ts

@@ -115,7 +115,7 @@ describe('finalizeAssistant dispatch (planFinalizeAssistant + applyFinalize)', (
 
   // Drive the SAME applyFinalize the service calls (no duplicated logic).
   async function dispatchFinalize(
-    repo: { insert: jest.Mock; update: jest.Mock },
+    repo: { insert: jest.Mock; finalizeOwner: jest.Mock },
     assistantId: string | undefined,
     flushed: AssistantFlush,
   ): Promise<void> {
@@ -135,21 +135,22 @@ describe('finalizeAssistant dispatch (planFinalizeAssistant + applyFinalize)', (
     expect(planFinalizeAssistant(undefined)).toEqual({ kind: 'insert' });
   });
 
-  it('(a) upfront insert succeeded -> finalize UPDATEs the row by id', async () => {
-    const repo = { insert: jest.fn(), update: jest.fn() };
+  it('(a) upfront insert succeeded -> finalize CONDITIONALLY updates the row by id (#487 owner-write)', async () => {
+    const repo = { insert: jest.fn(), finalizeOwner: jest.fn() };
     const flushed = flushAssistant([], 'final answer', 'completed', {
       finishReason: 'stop',
     });
     await dispatchFinalize(repo, 'a1', flushed);
-    expect(repo.update).toHaveBeenCalledWith('a1', workspaceId, flushed);
+    // #487: the owner write is the CONDITIONAL finalizeOwner, not a raw update.
+    expect(repo.finalizeOwner).toHaveBeenCalledWith('a1', workspaceId, flushed);
     expect(repo.insert).not.toHaveBeenCalled();
   });
 
   it('(b) upfront insert failed -> finalize INSERTs the terminal payload', async () => {
-    const repo = { insert: jest.fn(), update: jest.fn() };
+    const repo = { insert: jest.fn(), finalizeOwner: jest.fn() };
     const flushed = flushAssistant([], 'partial', 'error', { error: 'boom' });
     await dispatchFinalize(repo, undefined, flushed);
-    expect(repo.update).not.toHaveBeenCalled();
+    expect(repo.finalizeOwner).not.toHaveBeenCalled();
     expect(repo.insert).toHaveBeenCalledTimes(1);
     const arg = repo.insert.mock.calls[0][0];
     // The fallback insert carries the terminal content/status/metadata.

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

@@ -0,0 +1,279 @@
+import {
+  BadRequestException,
+  ConflictException,
+  ForbiddenException,
+  HttpException,
+} from '@nestjs/common';
+import { AiChatController } from './ai-chat.controller';
+import type { User, Workspace } from '@docmost/db/types/entity.types';
+
+/**
+ * #487 commit 3 — the single concurrency GATE (both modes) + the server supersede
+ * CAS, at the controller boundary. The gate + CAS run BEFORE res.hijack(), so a
+ * rejected concurrent start / a CAS branch returns clean JSON (an HttpException
+ * the controller's post-hijack catch re-serializes). These assert the OBSERVABLE
+ * HTTP contract against the real controller + a stubbed run service.
+ */
+describe('#487 AiChatController.stream — gate + supersede', () => {
+  const user = { id: 'u1' } as User;
+
+  function wsWith(autonomousRuns: boolean): Workspace {
+    return {
+      id: 'ws1',
+      settings: { ai: { chat: true, autonomousRuns } },
+    } as unknown as Workspace;
+  }
+
+  function makeReqRes(body: Record<string, unknown>) {
+    const req = {
+      raw: { sessionId: 'sess', once: jest.fn(), destroyed: false },
+      body,
+    };
+    const res = {
+      raw: {
+        writableEnded: false,
+        headersSent: false,
+        on: jest.fn(),
+        once: jest.fn(),
+        setHeader: jest.fn(),
+        end: jest.fn(),
+        statusCode: 200,
+        flushHeaders: jest.fn(),
+      },
+      hijack: jest.fn(),
+      status: jest.fn().mockReturnThis(),
+      send: jest.fn(),
+    };
+    return { req, res };
+  }
+
+  function makeController(
+    runServiceOverrides: Record<string, jest.Mock>,
+    // The chat assertOwnedChat resolves. Default: a chat OWNED by `user` (u1), so
+    // the ownership gate is transparent to the gate/CAS assertions below. Pass a
+    // foreign-owner (or undefined) chat to exercise the #487 owner rejection.
+    chat: { creatorId: string } | undefined = { creatorId: 'u1' },
+  ) {
+    const aiChatService = {
+      resolveRoleForRequest: jest.fn().mockResolvedValue(null),
+      getChatModel: jest.fn().mockResolvedValue({}),
+      stream: jest.fn().mockResolvedValue(undefined),
+    };
+    const aiChatRunService = {
+      getActiveForChat: jest.fn().mockResolvedValue(undefined),
+      supersede: jest.fn(),
+      beginRun: jest.fn().mockResolvedValue({
+        runId: 'run-new',
+        signal: new AbortController().signal,
+      }),
+      linkAssistantMessage: jest.fn(),
+      recordStep: jest.fn(),
+      finalizeRun: jest.fn(),
+      requestStop: jest.fn(),
+      ...runServiceOverrides,
+    };
+    const aiChatRepo = { findById: jest.fn().mockResolvedValue(chat) };
+    const controller = new AiChatController(
+      aiChatService as never,
+      aiChatRunService as never,
+      aiChatRepo as never, // aiChatRepo
+      {} as never, // aiChatMessageRepo
+      {} as never, // aiTranscription
+      {} as never, // pageRepo
+    );
+    return { controller, aiChatService, aiChatRunService, aiChatRepo };
+  }
+
+  const codeOf = (err: unknown) =>
+    (((err as HttpException).getResponse() as Record<string, unknown>) ?? {})
+      .code;
+
+  describe('single concurrency gate — BOTH modes reject the second tab with 409', () => {
+    for (const autonomousRuns of [true, false]) {
+      it(`rejects a concurrent start with 409 A_RUN_ALREADY_ACTIVE (autonomousRuns=${autonomousRuns})`, async () => {
+        const { controller, aiChatRunService } = makeController({
+          getActiveForChat: jest
+            .fn()
+            .mockResolvedValue({ id: 'run-live', chatId: 'c1' }),
+        });
+        const { req, res } = makeReqRes({ chatId: 'c1' });
+        let thrown: unknown;
+        try {
+          await controller.stream(
+            req as never,
+            res as never,
+            user,
+            wsWith(autonomousRuns),
+          );
+        } catch (e) {
+          thrown = e;
+        }
+        expect(thrown).toBeInstanceOf(ConflictException);
+        expect((thrown as HttpException).getStatus()).toBe(409);
+        expect(codeOf(thrown)).toBe('A_RUN_ALREADY_ACTIVE');
+        // Rejected BEFORE committing to the stream (no hijack, no service.stream).
+        expect(res.hijack).not.toHaveBeenCalled();
+        expect(aiChatRunService.getActiveForChat).toHaveBeenCalledWith(
+          'c1',
+          'ws1',
+        );
+      });
+    }
+  });
+
+  // #487 [security, F1]: stream() MUST owner-gate an existing chat exactly like its
+  // six sibling endpoints, BEFORE the supersede CAS. Otherwise a same-workspace
+  // non-owner could POST a supersede against another user's chat and (a) harvest
+  // that user's active runId from the 409 SUPERSEDE_TARGET_MISMATCH body, then (b)
+  // requestStop the foreign run. The gate must reject FIRST — no run lookup, no
+  // supersede, no stop, no runId leak.
+  describe('cross-user ownership gate (F1)', () => {
+    it('a non-owner streaming against someone else\'s chat is rejected (403) with NO runId leak and NO foreign requestStop', async () => {
+      // A live run exists on the victim's chat. Without the gate the supersede CAS
+      // would run and (faithful to the run service) return a MISMATCH carrying the
+      // victim's runId — the exact leak. With the gate it must never be reached.
+      const getActiveForChat = jest
+        .fn()
+        .mockResolvedValue({ id: 'run-victim', chatId: 'c-other' });
+      const supersede = jest
+        .fn()
+        .mockResolvedValue({ kind: 'mismatch', activeRunId: 'run-victim' });
+      const requestStop = jest.fn();
+      const { controller, aiChatService } = makeController(
+        { getActiveForChat, supersede, requestStop },
+        { creatorId: 'someone-else' }, // the chat is NOT owned by u1
+      );
+      const { req, res } = makeReqRes({
+        chatId: 'c-other',
+        supersede: { runId: 'guessed-uuid' },
+      });
+      let thrown: unknown;
+      try {
+        await controller.stream(req as never, res as never, user, wsWith(true));
+      } catch (e) {
+        thrown = e;
+      }
+      // Rejected by the ownership gate (403), the SAME shape the neighbors use.
+      expect(thrown).toBeInstanceOf(ForbiddenException);
+      expect((thrown as HttpException).getStatus()).toBe(403);
+      // Crucially NOT a 409 that would carry activeRunId — no runId is leaked.
+      const payload = JSON.stringify(
+        (thrown as HttpException).getResponse() ?? {},
+      );
+      expect(payload).not.toContain('run-victim');
+      expect(codeOf(thrown)).not.toBe('SUPERSEDE_TARGET_MISMATCH');
+      // The gate short-circuits BEFORE any run machinery runs.
+      expect(getActiveForChat).not.toHaveBeenCalled();
+      expect(supersede).not.toHaveBeenCalled();
+      expect(requestStop).not.toHaveBeenCalled();
+      expect(aiChatService.stream).not.toHaveBeenCalled();
+      expect(res.hijack).not.toHaveBeenCalled();
+    });
+  });
+
+  it('supersede MISMATCH -> 409 SUPERSEDE_TARGET_MISMATCH carrying the current runId', async () => {
+    const { controller } = makeController({
+      supersede: jest
+        .fn()
+        .mockResolvedValue({ kind: 'mismatch', activeRunId: 'run-other' }),
+    });
+    const { req, res } = makeReqRes({
+      chatId: 'c1',
+      supersede: { runId: 'run-x' },
+    });
+    let thrown: unknown;
+    try {
+      await controller.stream(req as never, res as never, user, wsWith(true));
+    } catch (e) {
+      thrown = e;
+    }
+    expect(thrown).toBeInstanceOf(ConflictException);
+    expect(codeOf(thrown)).toBe('SUPERSEDE_TARGET_MISMATCH');
+    expect(
+      ((thrown as HttpException).getResponse() as Record<string, unknown>)
+        .activeRunId,
+    ).toBe('run-other');
+    expect(res.hijack).not.toHaveBeenCalled();
+  });
+
+  it('supersede TIMEOUT -> 409 SUPERSEDE_TIMEOUT, nothing streamed', async () => {
+    const { controller } = makeController({
+      supersede: jest.fn().mockResolvedValue({ kind: 'timeout' }),
+    });
+    const { req, res } = makeReqRes({
+      chatId: 'c1',
+      supersede: { runId: 'run-x' },
+    });
+    let thrown: unknown;
+    try {
+      await controller.stream(req as never, res as never, user, wsWith(false));
+    } catch (e) {
+      thrown = e;
+    }
+    expect(thrown).toBeInstanceOf(ConflictException);
+    expect(codeOf(thrown)).toBe('SUPERSEDE_TIMEOUT');
+    expect(res.hijack).not.toHaveBeenCalled();
+  });
+
+  it('supersede INVALID (target on another chat) -> 400 SUPERSEDE_INVALID', async () => {
+    const { controller } = makeController({
+      supersede: jest.fn().mockResolvedValue({ kind: 'invalid' }),
+    });
+    const { req, res } = makeReqRes({
+      chatId: 'c1',
+      supersede: { runId: 'run-x' },
+    });
+    let thrown: unknown;
+    try {
+      await controller.stream(req as never, res as never, user, wsWith(true));
+    } catch (e) {
+      thrown = e;
+    }
+    expect(thrown).toBeInstanceOf(BadRequestException);
+    expect(codeOf(thrown)).toBe('SUPERSEDE_INVALID');
+  });
+
+  it('supersede without chatId -> 400 SUPERSEDE_INVALID', async () => {
+    const { controller, aiChatRunService } = makeController({});
+    const { req, res } = makeReqRes({ supersede: { runId: 'run-x' } });
+    let thrown: unknown;
+    try {
+      await controller.stream(req as never, res as never, user, wsWith(true));
+    } catch (e) {
+      thrown = e;
+    }
+    expect(thrown).toBeInstanceOf(BadRequestException);
+    expect(codeOf(thrown)).toBe('SUPERSEDE_INVALID');
+    expect(aiChatRunService.supersede).not.toHaveBeenCalled();
+  });
+
+  it('supersede READY -> proceeds to stream with superseded=true', async () => {
+    const { controller, aiChatService } = makeController({
+      supersede: jest.fn().mockResolvedValue({ kind: 'ready' }),
+      getActiveForChat: jest.fn().mockResolvedValue(undefined), // slot free after CAS
+    });
+    const { req, res } = makeReqRes({
+      chatId: 'c1',
+      supersede: { runId: 'run-x' },
+    });
+    await controller.stream(req as never, res as never, user, wsWith(true));
+    expect(res.hijack).toHaveBeenCalled();
+    expect(aiChatService.stream).toHaveBeenCalledTimes(1);
+    expect(aiChatService.stream.mock.calls[0][0].superseded).toBe(true);
+    // The run hooks are always present now (both modes).
+    expect(aiChatService.stream.mock.calls[0][0].runHooks).toBeDefined();
+  });
+
+  it('supersede DEGRADE -> proceeds to a normal send (superseded=false)', async () => {
+    const { controller, aiChatService } = makeController({
+      supersede: jest.fn().mockResolvedValue({ kind: 'degrade' }),
+    });
+    const { req, res } = makeReqRes({
+      chatId: 'c1',
+      supersede: { runId: 'run-x' },
+    });
+    await controller.stream(req as never, res as never, user, wsWith(false));
+    expect(aiChatService.stream).toHaveBeenCalledTimes(1);
+    expect(aiChatService.stream.mock.calls[0][0].superseded).toBe(false);
+  });
+});

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

@@ -418,6 +418,19 @@ export class AiChatController {
 
     const body = (req.body ?? {}) as AiChatStreamBody;
 
+    // #487 [security]: gate cross-user access to an EXISTING chat BEFORE anything
+    // reads its runs. Every sibling endpoint (getRun/stop/history/rename/delete/
+    // attachRunStream) owner-checks the chat via assertOwnedChat; stream() must too.
+    // Without this a same-workspace member who is NOT the chat owner could POST a
+    // supersede against another user's chat and (a) harvest that user's active runId
+    // out of the 409 SUPERSEDE_TARGET_MISMATCH body, then (b) requestStop the foreign
+    // run. Gate on the chatId the client sent, when present — a brand-new chat (no
+    // chatId) has no prior owner to check. Mirrors /stop's owner check (403 as the
+    // neighbors do), and runs pre-hijack so it returns clean JSON.
+    if (body.chatId) {
+      await this.assertOwnedChat(body.chatId, user, workspace);
+    }
+
     // Resolve the agent role for this turn BEFORE hijack: existing chats read it
     // from ai_chats.role_id (authoritative), a new chat from body.roleId. The
     // role drives both the persona and the optional model override below.
@@ -432,12 +445,66 @@ export class AiChatController {
     // HttpException) instead of breaking mid-stream.
     const model = await this.aiChatService.getChatModel(workspace.id, role);
 
-    // #184: one active run per chat. For an EXISTING chat reject a concurrent
-    // start with a clean 409 BEFORE hijack (the common double-submit / second-tab
-    // case), so the user gets JSON, not a mid-stream error. A brand-new chat
-    // (no chatId) cannot have a prior run, and the DB partial unique index is the
-    // backstop against any race that slips past this check.
-    if (autonomousRuns && body.chatId) {
+    // #487: server-side supersede CAS ("interrupt and send now"). When the client
+    // asks to replace a live run, atomically STOP it and wait for it to settle
+    // before this turn claims the slot. Runs BEFORE hijack so every branch returns
+    // clean JSON (the client keeps the composer text on a 409). See
+    // AiChatRunService.supersede for the branch semantics.
+    let superseded = false;
+    const supersedeRunId = body.supersede?.runId;
+    if (supersedeRunId) {
+      if (!body.chatId) {
+        throw new BadRequestException({
+          message: 'supersede requires chatId',
+          code: 'SUPERSEDE_INVALID',
+        });
+      }
+      const result = await this.aiChatRunService.supersede(
+        body.chatId,
+        supersedeRunId,
+        workspace.id,
+      );
+      switch (result.kind) {
+        case 'invalid':
+          throw new BadRequestException({
+            message: 'The run to supersede does not belong to this chat',
+            code: 'SUPERSEDE_INVALID',
+          });
+        case 'mismatch':
+          // A DIFFERENT run is active than the one the client targeted. Surface
+          // the CURRENT runId; the client does NOT auto-retry (a stale CAS).
+          throw new ConflictException({
+            message: 'A different agent run is now active on this chat',
+            code: 'SUPERSEDE_TARGET_MISMATCH',
+            activeRunId: result.activeRunId,
+          });
+        case 'timeout':
+          // The target did not settle within W — nothing was persisted, the
+          // composer keeps the text. NOT a rollback: the stop is already issued.
+          throw new ConflictException({
+            message:
+              'The previous run did not stop in time; nothing was sent — please try again',
+            code: 'SUPERSEDE_TIMEOUT',
+          });
+        case 'ready':
+          // The target stopped and settled: the slot is free. Prompt the new run
+          // that the old run's last operations may still be applying.
+          superseded = true;
+          break;
+        case 'degrade':
+          // The run already ended between click and POST — send normally.
+          break;
+      }
+    }
+
+    // #487: one active run per chat — ENFORCED IN BOTH MODES now (legacy mode used
+    // to have NO gate, so two tabs streamed two parallel turns on one chat, which
+    // interleaved history and crashed convertToModelMessages). Reject a concurrent
+    // start with a clean pre-hijack 409 (double-submit / second-tab). A brand-new
+    // chat (no chatId) cannot have a prior run, and the DB partial unique index in
+    // beginRun is the authoritative backstop for any race that slips past here
+    // (including a slot stolen between a supersede release and beginRun).
+    if (body.chatId) {
       const active = await this.aiChatRunService.getActiveForChat(
         body.chatId,
         workspace.id,
@@ -446,107 +513,94 @@ export class AiChatController {
         throw new ConflictException({
           message: 'An agent run is already in progress for this chat',
           code: 'A_RUN_ALREADY_ACTIVE',
+          activeRunId: active.id,
         });
       }
     }
 
-    // Run-lifecycle hooks (#184), only when the flag is on. They wrap the turn in
-    // a durable run whose abort is governed by the run (explicit stop), persist
-    // its progress, and settle its terminal status — see AiChatRunService.
-    const runHooks: AiChatRunHooks | undefined = autonomousRuns
-      ? {
-          begin: async (chatId) => {
-            const handle = await this.aiChatRunService.beginRun({
-              chatId,
-              workspaceId: workspace.id,
-              userId: user.id,
-              trigger: 'user',
-            });
-            // #184 phase 1.5: register the run-stream entry at BEGIN (before any
-            // frame) so a tab that attaches in the begin->seed window finds an
-            // entry to wait on. Gated on AI_CHAT_RESUMABLE_STREAM: with the flag
-            // off nothing is registered and attach always 204s.
-            if (
-              handle?.runId &&
-              this.environment?.isAiChatResumableStreamEnabled?.()
-            ) {
-              this.streamRegistry?.open(chatId, handle.runId);
-            }
-            return handle;
-          },
-          onAssistantSeeded: (runId, messageId) =>
-            this.aiChatRunService.linkAssistantMessage(
-              runId,
-              workspace.id,
-              messageId,
-            ),
-          onStep: (runId, stepCount) =>
-            void this.aiChatRunService.recordStep(
-              runId,
-              workspace.id,
-              stepCount,
-            ),
-          onSettled: (runId, status, error) =>
-            this.aiChatRunService.finalizeRun(
-              runId,
-              workspace.id,
-              status,
-              error,
-            ),
+    // #487: the turn is ALWAYS a first-class RUN now (both modes). The mode
+    // difference is only the abort semantics on a browser disconnect (onClose
+    // below). currentRunId is captured at begin so a legacy disconnect can stop
+    // the run through its stop lever.
+    let currentRunId: string | undefined;
+    const runHooks: AiChatRunHooks = {
+      begin: async (chatId) => {
+        const handle = await this.aiChatRunService.beginRun({
+          chatId,
+          workspaceId: workspace.id,
+          userId: user.id,
+          trigger: 'user',
+        });
+        currentRunId = handle?.runId;
+        // #184 phase 1.5: register the run-stream entry at BEGIN (before any
+        // frame) so a tab that attaches in the begin->seed window finds an entry
+        // to wait on. Gated on AI_CHAT_RESUMABLE_STREAM.
+        if (
+          handle?.runId &&
+          this.environment?.isAiChatResumableStreamEnabled?.()
+        ) {
+          this.streamRegistry?.open(chatId, handle.runId);
         }
-      : undefined;
+        return handle;
+      },
+      onAssistantSeeded: (runId, messageId) =>
+        this.aiChatRunService.linkAssistantMessage(
+          runId,
+          workspace.id,
+          messageId,
+        ),
+      onStep: (runId, stepCount) =>
+        void this.aiChatRunService.recordStep(runId, workspace.id, stepCount),
+      onSettled: (runId, status, error) =>
+        this.aiChatRunService.finalizeRun(runId, workspace.id, status, error),
+    };
 
-    // Abort the agent loop when the client disconnects. `close` also fires on
-    // normal completion, so only abort when the response has not finished
-    // writing (a genuine disconnect). `once` fires at most once and self-removes;
-    // we also drop it on response `finish` so it never lingers after the stream
-    // completes normally (the AI SDK pipes the response fire-and-forget, so we
-    // cannot simply remove it once `stream()` returns).
+    // Handle a client disconnect. `close` also fires on normal completion, so only
+    // act when the response has not finished writing (a genuine disconnect). `once`
+    // fires at most once and self-removes; we also drop it on response `finish`.
     // DIAGNOSTIC (Safari stream-drop investigation) — temporary: wall-clock at
     // which a Safari disconnect is observed, measured from request receipt.
     const reqStartedAt = Date.now();
     const controller = new AbortController();
     const onClose = (): void => {
-      // A genuine disconnect leaves the response unfinished (unlike a normal
-      // completion, which also fires `close`). Such a drop — e.g. a reverse
-      // proxy cutting the SSE mid-answer — is otherwise invisible server-side,
-      // so log it here.
       if (!res.raw.writableEnded) {
         if (autonomousRuns) {
-          // #184: the turn is a DETACHED run. A disconnect must NOT abort it —
-          // the run keeps executing and persisting server-side; the client
-          // reconnects via /ai-chat/run (or re-stops via /ai-chat/stop). Log only.
+          // #184: a DETACHED run — a disconnect must NOT stop it. The run keeps
+          // executing and persisting server-side; the client reconnects via
+          // /ai-chat/run (or re-stops via /ai-chat/stop). Log only.
           this.logger.log(
             `AI chat stream: client disconnected; run continues server-side ` +
               `(elapsed=${Date.now() - reqStartedAt}ms since request received)`,
           );
         } else {
+          // #487: legacy — a disconnect ENDS the turn, but the turn is now a RUN,
+          // so stop it through the run's stop lever (requestStop). streamText no
+          // longer consumes the socket signal (effectiveSignal is the run signal),
+          // so aborting `controller` would do nothing; requestStop aborts the run.
           this.logger.warn(
-            `AI chat stream: client disconnected before completion; aborting turn ` +
-              `(elapsed=${Date.now() - reqStartedAt}ms since request received)`,
+            `AI chat stream: client disconnected before completion; stopping the ` +
+              `run (elapsed=${Date.now() - reqStartedAt}ms since request received)`,
           );
-          controller.abort();
+          if (currentRunId) {
+            void this.aiChatRunService.requestStop(currentRunId, workspace.id);
+          }
         }
       }
     };
     req.raw.once('close', onClose);
     res.raw.once('finish', () => req.raw.off('close', onClose));
 
-    // #184: in detached mode the turn is NOT aborted on disconnect, so the SDK's
-    // pipe keeps writing to a socket the client may have dropped — for the rest of
-    // the (continuing) run. A write to the dead socket can emit an 'error' on the
-    // raw response; without a listener that surfaces as an unhandled error event.
-    // Swallow it (the run continues server-side regardless). Legacy mode aborts on
-    // disconnect, so it does not need this and keeps its exact prior behavior.
-    if (autonomousRuns) {
-      res.raw.on('error', (err) => {
-        this.logger.debug(
-          `AI chat detached stream: post-disconnect socket error swallowed: ${
-            err instanceof Error ? err.message : String(err)
-          }`,
-        );
-      });
-    }
+    // #184/#487: the run/pipe can outlive the socket in BOTH modes now (autonomous
+    // keeps going; legacy keeps going until requestStop's abort unwinds the turn).
+    // The SDK's pipe may then write to a dropped socket and emit an 'error' on the
+    // raw response — swallow it so it never surfaces as an unhandled error event.
+    res.raw.on('error', (err) => {
+      this.logger.debug(
+        `AI chat stream: post-disconnect socket error swallowed: ${
+          err instanceof Error ? err.message : String(err)
+        }`,
+      );
+    });
 
     // Commit to streaming: hijack so Fastify stops managing the response and
     // the AI SDK can write the UI-message stream directly to the Node socket.
@@ -562,8 +616,10 @@ export class AiChatController {
         signal: controller.signal,
         model,
         role,
-        // #184: present only when the flag is on; wraps the turn in a durable run.
+        // #487: the turn is always run-wrapped now (both modes).
         runHooks,
+        // #487: warn the new run that a superseded run's last ops may still apply.
+        superseded,
       });
     } catch (err) {
       // Any failure AFTER hijack can no longer go through Nest's exception

apps/server/src/core/ai-chat/ai-chat.parts-resilience.spec.ts

@@ -0,0 +1,142 @@
+// #489 — client-parts validation + resilient history conversion.
+//
+// These unit tests exercise the two exported helpers against the REAL
+// `convertToModelMessages` from `ai` (NOT a mock): a genuinely malformed part
+// (a `null` element inside a parts array) makes the real converter throw
+// ("Cannot read properties of null"), which is the actual production
+// "bricked chat" mechanism this fix defends against. Asserting against the real
+// converter (rather than a mock-shaped error) is the whole point — a mock would
+// hide a version change in the converter's throw behaviour.
+import { convertToModelMessages, type UIMessage } from 'ai';
+import {
+  sanitizeUserParts,
+  convertHistoryResilient,
+  TOOL_CONTEXT_OMITTED_MARKER,
+} from './ai-chat.service';
+
+type Row = Omit<UIMessage, 'id'> & { id: string };
+
+describe('sanitizeUserParts (#489, branch: validation on receipt)', () => {
+  it('keeps whitelisted text parts unchanged', () => {
+    const drops: string[] = [];
+    const out = sanitizeUserParts(
+      [
+        { type: 'text', text: 'a' },
+        { type: 'text', text: 'b' },
+      ] as UIMessage['parts'],
+      (t) => drops.push(t),
+    );
+    expect(out).toEqual([
+      { type: 'text', text: 'a' },
+      { type: 'text', text: 'b' },
+    ]);
+    expect(drops).toEqual([]);
+  });
+
+  it('drops a non-text part (a tool-part in input-available) and reports its type', () => {
+    const drops: string[] = [];
+    const out = sanitizeUserParts(
+      [
+        { type: 'text', text: 'hi' },
+        {
+          type: 'tool-getPage',
+          toolCallId: 't1',
+          state: 'input-available',
+          input: { pageId: 'p' },
+        },
+      ] as unknown as UIMessage['parts'],
+      (t) => drops.push(t),
+    );
+    expect(out).toEqual([{ type: 'text', text: 'hi' }]);
+    expect(drops).toEqual(['tool-getPage']);
+  });
+
+  it('drops a null part (the shape that would poison convertToModelMessages)', () => {
+    const drops: string[] = [];
+    const out = sanitizeUserParts(
+      [{ type: 'text', text: 'hi' }, null] as unknown as UIMessage['parts'],
+      (t) => drops.push(t),
+    );
+    expect(out).toEqual([{ type: 'text', text: 'hi' }]);
+    expect(drops).toEqual(['(unknown)']);
+  });
+
+  it('returns undefined when nothing survives (so a null metadata is persisted)', () => {
+    const out = sanitizeUserParts(
+      [
+        { type: 'tool-x', toolCallId: 't', state: 'input-available' },
+      ] as unknown as UIMessage['parts'],
+      () => undefined,
+    );
+    expect(out).toBeUndefined();
+  });
+
+  it('returns undefined for a non-array input', () => {
+    expect(
+      sanitizeUserParts(undefined as unknown as UIMessage['parts'], () => undefined),
+    ).toBeUndefined();
+  });
+});
+
+describe('convertHistoryResilient (#489, branches: happy + per-row degradation)', () => {
+  it('happy path: healthy history converts identically to convertToModelMessages, no degrade', async () => {
+    const history: Row[] = [
+      { id: 'u1', role: 'user', parts: [{ type: 'text', text: 'hi' }] },
+      { id: 'a1', role: 'assistant', parts: [{ type: 'text', text: 'hello' }] },
+    ];
+    const degrades: number[] = [];
+    const out = await convertHistoryResilient(history, (i) => degrades.push(i));
+    const expected = await convertToModelMessages(history as UIMessage[]);
+    expect(out).toEqual(expected);
+    expect(degrades).toEqual([]);
+  });
+
+  it('REAL poison: a null part throws in the batch converter but is isolated and degraded to a marker', async () => {
+    // Sanity: the real converter genuinely throws on this shape.
+    const poisoned: Row = {
+      id: 'a1',
+      role: 'assistant',
+      parts: [
+        { type: 'text', text: 'earlier answer' },
+        null,
+      ] as unknown as UIMessage['parts'],
+    };
+    await expect(
+      convertToModelMessages([poisoned as UIMessage]),
+    ).rejects.toThrow();
+
+    const history: Row[] = [
+      { id: 'u1', role: 'user', parts: [{ type: 'text', text: 'first' }] },
+      poisoned,
+      { id: 'u2', role: 'user', parts: [{ type: 'text', text: 'second' }] },
+    ];
+    const degrades: number[] = [];
+    const out = await convertHistoryResilient(history, (i) => degrades.push(i));
+
+    // Only the poisoned row (index 1) is degraded.
+    expect(degrades).toEqual([1]);
+    // Healthy rows survive verbatim.
+    const flat = JSON.stringify(out);
+    expect(flat).toContain('first');
+    expect(flat).toContain('second');
+    // The degraded row carries its readable text AND the truncation marker so the
+    // model sees that tool context was omitted (never a silent loss).
+    expect(flat).toContain('earlier answer');
+    expect(flat).toContain(TOOL_CONTEXT_OMITTED_MARKER);
+    // The whole batch converted (3 model messages, none dropped).
+    expect(out).toHaveLength(3);
+  });
+
+  it('a fully-poisoned row (no readable text) still degrades to just the marker', async () => {
+    const history: Row[] = [
+      {
+        id: 'a1',
+        role: 'assistant',
+        parts: [null] as unknown as UIMessage['parts'],
+      },
+    ];
+    const out = await convertHistoryResilient(history, () => undefined);
+    expect(out).toHaveLength(1);
+    expect(JSON.stringify(out)).toContain(TOOL_CONTEXT_OMITTED_MARKER);
+  });
+});

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

@@ -101,6 +101,22 @@ const INTERRUPT_NOTE =
   'assume your previous response was complete, and do not silently restart the ' +
   'partial work — build on it or follow the new instruction.';
 
+/**
+ * #487: injected on a turn started by SUPERSEDING a previous run (the user hit
+ * "interrupt and send now" while a run was live). The previous run was Stopped,
+ * but there is NO side-effect quiescence — a write it had already committed, or
+ * one committing at the moment of Stop, may land with a small delay AFTER this new
+ * run starts. So the model is told its picture of the page/state may be a beat
+ * stale and to re-read before assuming an edit did or did not apply.
+ */
+const SUPERSEDE_NOTE =
+  'NOTE: A previous agent run in this conversation was just interrupted so this ' +
+  'new turn could start. That run was stopped, but any operation it had already ' +
+  'begun (e.g. a page edit) may still be applied with a short delay. Do not ' +
+  'assume the document/state is exactly as the interrupted run left it — if you ' +
+  'need to rely on the current content, RE-READ it with the page tools before ' +
+  'acting rather than trusting a cached view.';
+
 /**
  * Injected on a turn where the open page was hand-edited by the user (or anyone
  * else) AFTER the agent's previous response ended (#274). The server takes a
@@ -203,6 +219,14 @@ export interface BuildSystemPromptInput {
    * (partial) answer was cut off by the user's new message.
    */
   interrupted?: boolean;
+  /**
+   * #487: true when THIS turn was started by superseding a still-live previous run
+   * ("interrupt and send now"). Adds SUPERSEDE_NOTE so the model knows the previous
+   * run's last operations may still be applying and to re-read state it depends on.
+   * Distinct from `interrupted` (which is about a PARTIAL prior answer in history);
+   * both can be set together. Self-clears — set only for the superseding turn.
+   */
+  superseded?: boolean;
   /**
    * Set only when the open page was edited by the user AFTER the agent's previous
    * turn ended (#274), confirmed server-side by diffing the current page against
@@ -311,6 +335,7 @@ export function buildSystemPrompt({
   openedPage,
   mcpInstructions,
   interrupted,
+  superseded,
   pageChanged,
   deferredToolsEnabled,
   toolCatalog,
@@ -360,6 +385,13 @@ export function buildSystemPrompt({
     context += `\n${INTERRUPT_NOTE}`;
   }
 
+  // Supersede note (#487): present only for a turn that stopped and replaced a
+  // still-live previous run — warns the model the previous run's last operations
+  // may still be applying (no side-effect quiescence).
+  if (superseded) {
+    context += `\n${SUPERSEDE_NOTE}`;
+  }
+
   // Per-turn page-change note (#274). Added to the context section (inside the
   // safety sandwich), present only when the server detected that the open page
   // was edited by the user since the agent's last turn ended. The diff content is

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

@@ -89,11 +89,22 @@ describe('AiChatService.stream run-lifecycle safety net (#184)', () => {
     const runRepo = {
       insert: jest.fn().mockResolvedValue({ id: 'run-1', status: 'running' }),
       update: jest.fn().mockResolvedValue({ id: 'run-1' }),
+      // #487: the terminal settle now goes through the CONDITIONAL write.
+      finalizeIfActive: jest
+        .fn()
+        .mockResolvedValue({ id: 'run-1', status: 'failed' }),
+      findById: jest.fn().mockResolvedValue(undefined),
     };
     const runService = new AiChatRunService(runRepo as never, { isCloud: () => false } as never);
 
-    // The user-message insert (the first bare await after beginRun) throws.
+    // The user-message insert throws. #489 runs the history load + convert BEFORE
+    // the insert (convert-before-insert, so a retry cannot duplicate the user row),
+    // so `findAllByChat` (a real repo method) is now called first — stub it to an
+    // empty history so the flow reaches the insert. Both awaits are AFTER beginRun,
+    // so the "exception after beginRun -> settled to error" invariant is unchanged;
+    // the throw point simply moved from insert to a later insert after a no-op load.
     const aiChatMessageRepo = {
+      findAllByChat: jest.fn().mockResolvedValue([]),
       insert: jest.fn().mockRejectedValue(new Error('insert boom')),
     };
     const aiChatRepo = {
@@ -148,9 +159,10 @@ describe('AiChatService.stream run-lifecycle safety net (#184)', () => {
 
     // The run was begun...
     expect(runRepo.insert).toHaveBeenCalledTimes(1);
-    // ...then settled to a terminal FAILED status by the safety net...
-    expect(runRepo.update).toHaveBeenCalledTimes(1);
-    expect(runRepo.update).toHaveBeenCalledWith(
+    // ...then settled to a terminal FAILED status by the safety net (via the
+    // #487 conditional write)...
+    expect(runRepo.finalizeIfActive).toHaveBeenCalledTimes(1);
+    expect(runRepo.finalizeIfActive).toHaveBeenCalledWith(
       'run-1',
       'ws1',
       expect.objectContaining({ status: 'failed' }),

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
@@ -151,6 +155,8 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       insert: jest.fn(async () => ({ id: 'msg-1' })),
       findAllByChat: jest.fn(async () => []),
       update: jest.fn(async () => ({ id: 'msg-1' })),
+      finalizeOwner: jest.fn(async () => ({ id: 'msg-1' })),
+      findStreamingWithTerminalRun: jest.fn(async () => []),
     };
     const aiSettings = { resolve: jest.fn(async () => ({})) };
     const tools = { forUser: jest.fn(async () => ({})) };
@@ -175,7 +181,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       {} as never, // pageAccess
       { isAiChatDeferredToolsEnabled: () => false, isAiChatFinalStepLockdownEnabled: () => false } as never, // environment
     );
-    return { svc };
+    return { svc, aiChatMessageRepo };
   }
 
   const body = {
@@ -281,7 +287,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
   // Drive stream() to the point streamText is called, capturing the options object
   // (which carries onStepFinish/onFinish/onError/onAbort) and the run hooks.
   async function captureStreamCallbacks() {
-    const { svc } = makeService();
+    const { svc, aiChatMessageRepo } = makeService();
     let capturedOpts: any;
     streamTextMock.mockImplementation((opts: any) => {
       capturedOpts = opts;
@@ -308,7 +314,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       runHooks: runHooks as never,
     });
     expect(capturedOpts).toBeDefined();
-    return { capturedOpts, runHooks };
+    return { capturedOpts, runHooks, aiChatMessageRepo };
   }
 
   it('F9: onStepFinish bumps the run step count, onFinish settles the run "completed" (the dominant autonomous-run path)', async () => {
@@ -328,7 +334,13 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       usage: {},
       steps: [],
     });
-    expect(runHooks.onSettled).toHaveBeenCalledWith('run-1', 'completed');
+    // #487: onFinish passes the (undefined) error slot so a message-finalize
+    // failure could error-mark the run; on the success path it is undefined.
+    expect(runHooks.onSettled).toHaveBeenCalledWith(
+      'run-1',
+      'completed',
+      undefined,
+    );
   });
 
   it('F9: onAbort settles the run "aborted"', async () => {
@@ -357,25 +369,70 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       expect.stringContaining('provider exploded'),
     );
   });
+
+  // #490 reactive branch: a provider CONTEXT-OVERFLOW 400 in onError is classified,
+  // records a distinguishable cause, and stamps metadata.replayOverflow so the NEXT
+  // turn's budgeter trims aggressively (the recovery that un-bricks the chat).
+  it('#490: a context-overflow 400 stamps replayOverflow on the finalized row', async () => {
+    jest
+      .spyOn(Logger.prototype, 'error')
+      .mockImplementation(() => undefined as never);
+    jest
+      .spyOn(Logger.prototype, 'warn')
+      .mockImplementation(() => undefined as never);
+    const { capturedOpts, aiChatMessageRepo } = await captureStreamCallbacks();
+
+    const overflow = Object.assign(new Error('too large'), {
+      statusCode: 400,
+      message:
+        "This model's maximum context length is 128000 tokens. However, your messages resulted in 214000 tokens. Please reduce the length.",
+    });
+    await capturedOpts.onError({ error: overflow });
+
+    // The seed row exists (finalizeOwner is the owner-write path).
+    expect(aiChatMessageRepo.finalizeOwner).toHaveBeenCalled();
+    const calls = aiChatMessageRepo.finalizeOwner.mock.calls as any[][];
+    const patch = calls[calls.length - 1][2] as {
+      status: string;
+      metadata: Record<string, unknown>;
+    };
+    expect(patch.status).toBe('error');
+    expect(patch.metadata.replayOverflow).toBe(true);
+    expect(patch.metadata.error).toContain('контекстное окно');
+  });
+
+  it('#490: a non-overflow error does NOT stamp replayOverflow', async () => {
+    jest
+      .spyOn(Logger.prototype, 'error')
+      .mockImplementation(() => undefined as never);
+    const { capturedOpts, aiChatMessageRepo } = await captureStreamCallbacks();
+    await capturedOpts.onError({ error: new Error('network reset') });
+    const calls = aiChatMessageRepo.finalizeOwner.mock.calls as any[][];
+    const patch = calls[calls.length - 1][2] as {
+      status: string;
+      metadata: Record<string, unknown>;
+    };
+    expect('replayOverflow' in patch.metadata).toBe(false);
+  });
 });
 
 /**
- * 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
- *     (legacy fallback): it logs "...streaming without run tracking", leaves
- *     `effectiveSignal = signal` (runId undefined) and serves the turn anyway.
+ *   - ANY OTHER begin failure -> throw ServiceUnavailableException(A_RUN_BEGIN_FAILED)
+ *     BEFORE the first byte (#486, commit 4).
  *
- * The contract: a transient beginRun failure (e.g. a non-unique DB error inserting
- * the run row) must STILL serve the user's turn — it must NOT re-throw and must NOT
- * be misclassified as a 409. A regression that re-threw here would break EVERY turn
- * on a begin failure with nothing to catch it. This branch is otherwise undriven by
- * any spec, so it is pinned here SEPARATELY from the 409 path: a plain begin error
- * proceeds to streamText with the SOCKET signal and still persists the user turn.
+ * POLICY CHANGE (#486): the OLD contract here was "SWALLOW + stream the turn
+ * UNTRACKED on the socket signal". That was reversed: an untracked run is
+ * invisible to /stop, is not aborted on disconnect, and slips past the one-run
+ * gate — an unstoppable ghost run in autonomous mode. Now a plain begin failure
+ * FAILS the turn fast with a 503 A_RUN_BEGIN_FAILED, before any user row is
+ * 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() {
@@ -411,6 +468,8 @@ describe('AiChatService.stream — begin-failure resilience / legacy fallback (#
       insert: jest.fn(async () => ({ id: 'msg-1' })),
       findAllByChat: jest.fn(async () => []),
       update: jest.fn(async () => ({ id: 'msg-1' })),
+      finalizeOwner: jest.fn(async () => ({ id: 'msg-1' })),
+      findStreamingWithTerminalRun: jest.fn(async () => []),
     };
     const aiSettings = { resolve: jest.fn(async () => ({})) };
     const tools = { forUser: jest.fn(async () => ({})) };
@@ -455,7 +514,7 @@ 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 +546,26 @@ describe('AiChatService.stream — begin-failure resilience / legacy fallback (#
       } as never,
     });
 
-    // The turn proceeds: NO throw at all (in particular NOT a 409).
-    await expect(promise).resolves.toBeUndefined();
+    // NEW POLICY: the turn is REJECTED with a 503 A_RUN_BEGIN_FAILED (not a 409,
+    // 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.
-    expect(aiChatMessageRepo.insert).toHaveBeenCalled();
-    expect(streamTextMock).toHaveBeenCalledTimes(1);
-
-    // 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);
+    // Fail-fast: the turn NEVER streamed — no user row persisted, no streamText
+    // call, so no orphan/untracked run was left behind.
+    expect(aiChatMessageRepo.insert).not.toHaveBeenCalled();
+    expect(streamTextMock).not.toHaveBeenCalled();
   });
 });

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

@@ -13,6 +13,7 @@ import {
   compactToolOutput,
   assistantParts,
   serializeSteps,
+  type StepPartsCache,
   rowToUiMessage,
   prepareAgentStep,
   stepBudgetWarning,
@@ -28,6 +29,9 @@ import {
   FINAL_STEP_NUDGE,
   STEP_LIMIT_NO_ANSWER_MARKER,
   OUTPUT_DEGENERATION_ERROR,
+  lastAssistantContextTokens,
+  lastAssistantReplayOverflow,
+  seedActivatedTools,
 } from './ai-chat.service';
 import type { AiChatMessage, Workspace } from '@docmost/db/types/entity.types';
 import { buildSystemPrompt } from './ai-chat.prompt';
@@ -114,6 +118,54 @@ describe('compactToolOutput', () => {
 describe('assistantParts', () => {
   type AnyPart = Record<string, unknown>;
 
+  // #490 memoization: assistantParts builds each step's parts once and caches
+  // them by the step OBJECT's identity, so a mid-stream flush does not
+  // re-stringify every prior step's (large) output. Observable property: with a
+  // shared cache, the second call over the SAME step object returns the cached
+  // (identical) part array even if the step's underlying output was swapped —
+  // proving the work was memoized, not redone.
+  it('memoizes a step by identity (shared cache => one build per step)', () => {
+    const cache: StepPartsCache = new WeakMap();
+    const step = {
+      text: 'x',
+      toolCalls: [{ toolCallId: 'c1', toolName: 'getPage', input: {} }],
+      toolResults: [{ toolCallId: 'c1', toolName: 'getPage', output: { v: 1 } }],
+    };
+    const first = assistantParts([step], '', cache) as AnyPart[];
+    expect((first.find((p) => p.type === 'tool-getPage')!.output as any).v).toBe(
+      1,
+    );
+    // Swap the output for a NEW value; a re-build would pick it up, a cache hit
+    // keeps the first result.
+    step.toolResults[0] = {
+      toolCallId: 'c1',
+      toolName: 'getPage',
+      output: { v: 2 },
+    };
+    const second = assistantParts([step], '', cache) as AnyPart[];
+    expect((second.find((p) => p.type === 'tool-getPage')!.output as any).v).toBe(
+      1,
+    );
+    // Same cached part objects are reused.
+    expect(second.find((p) => p.type === 'tool-getPage')).toBe(
+      first.find((p) => p.type === 'tool-getPage'),
+    );
+  });
+
+  it('without a cache, each call rebuilds (no stale memo)', () => {
+    const step = {
+      text: 'x',
+      toolCalls: [{ toolCallId: 'c1', toolName: 'getPage', input: {} }],
+      toolResults: [{ toolCallId: 'c1', toolName: 'getPage', output: { v: 1 } }],
+    };
+    const first = assistantParts([step], '') as AnyPart[];
+    step.toolResults[0].output = { v: 2 };
+    const second = assistantParts([step], '') as AnyPart[];
+    expect((second.find((p) => p.type === 'tool-getPage')!.output as any).v).toBe(
+      2,
+    );
+  });
+
   it('emits output-available for a tool-call WITH a paired result', () => {
     const steps = [
       {
@@ -231,61 +283,299 @@ describe('assistantParts', () => {
   });
 });
 
-describe('serializeSteps', () => {
+// #490 trace format v2: per call the trace stores { input } for the call and an
+// OUTCOME element — { ok: true } on success, { error, kind: 'thrown' } on a
+// thrown tool-error, { error, kind: 'interrupted' } on a mid-step abort. The tool
+// OUTPUT is no longer duplicated here (it lives once in metadata.parts).
+describe('serializeSteps (trace v2)', () => {
   it('returns null when there are no calls or results', () => {
     expect(serializeSteps([])).toBeNull();
   });
 
-  it('flattens calls and results into a compact trace', () => {
+  it('pairs a successful call with an { ok: true } outcome and NO output', () => {
     const trace = serializeSteps([
       {
-        toolCalls: [{ toolName: 'getPage', input: { id: 'p1' } }],
-        toolResults: [{ toolName: 'getPage', output: { title: 'T' } }],
+        toolCalls: [{ toolCallId: 'c1', toolName: 'getPage', input: { id: 'p1' } }],
+        toolResults: [{ toolCallId: 'c1', toolName: 'getPage' }],
       },
     ]) as Array<Record<string, unknown>>;
     expect(trace).toHaveLength(2);
     expect(trace[0]).toEqual({ toolName: 'getPage', input: { id: 'p1' } });
-    expect(trace[1]).toEqual({ toolName: 'getPage', output: { title: 'T' } });
+    expect(trace[1]).toEqual({ toolName: 'getPage', ok: true });
+    // The output is NOT stored in the trace any more (dedup: it lives in parts).
+    expect(trace.some((e) => 'output' in e)).toBe(false);
   });
 
-  it('records a THROWN tool failure (tool-error part) with its error message', () => {
+  it('records a THROWN failure with { error, kind: "thrown" }', () => {
     const trace = serializeSteps([
       {
-        toolCalls: [{ toolName: 'editPageText', input: { id: 'p1' } }],
+        toolCalls: [
+          { toolCallId: 'c1', toolName: 'editPageText', input: { id: 'p1' } },
+        ],
         toolResults: [],
         content: [
           {
             type: 'tool-error',
+            toolCallId: 'c1',
             toolName: 'editPageText',
             error: new Error('page is locked'),
           },
         ],
       },
     ]) as Array<Record<string, unknown>>;
-    // The call element is followed by a paired error element (mirroring how a
-    // successful result is appended), so the failure survives in the trace.
     expect(trace).toHaveLength(2);
     expect(trace[0]).toEqual({ toolName: 'editPageText', input: { id: 'p1' } });
     expect(trace[1]).toEqual({
       toolName: 'editPageText',
       error: 'page is locked',
+      kind: 'thrown',
     });
   });
 
-  it('truncates a very long tool-error message to the tool-output limit', () => {
+  it('marks an interrupted call (no result, no throw) with kind "interrupted"', () => {
+    const trace = serializeSteps([
+      {
+        toolCalls: [
+          { toolCallId: 'c1', toolName: 'createComment', input: { x: 1 } },
+        ],
+        toolResults: [],
+        content: [],
+      },
+    ]) as Array<Record<string, unknown>>;
+    expect(trace).toHaveLength(2);
+    expect(trace[1]).toEqual({
+      toolName: 'createComment',
+      error: 'Tool call did not complete.',
+      kind: 'interrupted',
+    });
+    // Structurally distinct from a thrown hard-fail so it never inflates an
+    // error-rate scan.
+    expect((trace[1] as { kind: string }).kind).not.toBe('thrown');
+  });
+
+  it('truncates a very long thrown-error message to the tool-output limit', () => {
     const long = 'x'.repeat(5000);
     const trace = serializeSteps([
       {
-        toolCalls: [{ toolName: 'editPageText', input: {} }],
+        toolCalls: [{ toolCallId: 'c1', toolName: 'editPageText', input: {} }],
         toolResults: [],
-        content: [{ type: 'tool-error', toolName: 'editPageText', error: long }],
+        content: [
+          {
+            type: 'tool-error',
+            toolCallId: 'c1',
+            toolName: 'editPageText',
+            error: long,
+          },
+        ],
       },
     ]) as Array<Record<string, unknown>>;
     const errorText = trace[1].error as string;
-    // Truncated (not the full 5000 chars) and carries the omission marker.
     expect(errorText.length).toBeLessThan(long.length);
     expect(errorText).toContain('chars omitted');
   });
+
+  it('pairs parallel calls in one step with their outcomes by id', () => {
+    const trace = serializeSteps([
+      {
+        toolCalls: [
+          { toolCallId: 'a', toolName: 'getPage', input: {} },
+          { toolCallId: 'b', toolName: 'searchPages', input: {} },
+        ],
+        toolResults: [{ toolCallId: 'b', toolName: 'searchPages' }],
+        content: [
+          { type: 'tool-error', toolCallId: 'a', toolName: 'getPage', error: 'nope' },
+        ],
+      },
+    ]) as Array<Record<string, unknown>>;
+    // call a, outcome a (thrown), call b, outcome b (ok)
+    expect(trace).toHaveLength(4);
+    expect(trace[1]).toEqual({ toolName: 'getPage', error: 'nope', kind: 'thrown' });
+    expect(trace[3]).toEqual({ toolName: 'searchPages', ok: true });
+  });
+});
+
+// #490: every assistant row flushAssistant writes carries the v2 era marker so a
+// dual-shape diagnostic query can branch on the trace shape without inspecting it.
+describe('toolTraceVersion era marker (#490)', () => {
+  it('stamps metadata.toolTraceVersion = 2 on every flushed row', () => {
+    const seed = flushAssistant([], '', 'streaming');
+    expect(seed.metadata.toolTraceVersion).toBe(2);
+    const done = flushAssistant(
+      [
+        {
+          text: 'ok',
+          toolCalls: [{ toolCallId: 'c1', toolName: 'getPage', input: {} }],
+          toolResults: [{ toolCallId: 'c1', toolName: 'getPage' }],
+        },
+      ],
+      '',
+      'completed',
+      { finishReason: 'stop' },
+    );
+    expect(done.metadata.toolTraceVersion).toBe(2);
+  });
+});
+
+// #490 replay-budget signal helpers over persisted history.
+describe('lastAssistantContextTokens', () => {
+  const row = (
+    role: string,
+    metadata: Record<string, unknown> | null,
+  ): AiChatMessage => ({ role, metadata }) as unknown as AiChatMessage;
+
+  it('reads the most recent assistant turn contextTokens (provider fact)', () => {
+    const hist = [
+      row('user', null),
+      row('assistant', { contextTokens: 12000 }),
+      row('user', null),
+      row('assistant', { contextTokens: 41000 }),
+    ];
+    expect(lastAssistantContextTokens(hist)).toBe(41000);
+  });
+
+  it('returns undefined when the last assistant turn recorded no usage', () => {
+    const hist = [row('assistant', { error: 'boom' }), row('user', null)];
+    expect(lastAssistantContextTokens(hist)).toBeUndefined();
+    expect(lastAssistantContextTokens([])).toBeUndefined();
+  });
+});
+
+// #490 snapshotOpenPage fast-path: skip the full Markdown export + upsert when a
+// snapshot already exists at the page's CURRENT version (same updated_at instant).
+describe('snapshotOpenPage fast-path (#490)', () => {
+  function makeSvc(existingSnapshot: unknown, pageUpdatedAt: Date) {
+    const exportPageMarkdown = jest.fn(async () => '# md');
+    const upsert = jest.fn(async () => undefined);
+    const findByChatPage = jest.fn(async () => existingSnapshot);
+    const pageRepo = {
+      findById: jest.fn(async () => ({
+        id: 'p1',
+        workspaceId: 'ws1',
+        updatedAt: pageUpdatedAt,
+      })),
+    };
+    const svc = new AiChatService(
+      {} as never, // ai
+      {} as never, // aiChatRepo
+      {} as never, // aiChatMessageRepo
+      { findByChatPage, upsert } as never, // aiChatPageSnapshotRepo
+      {} as never, // aiSettings
+      { exportPageMarkdown } as never, // tools
+      {} as never, // mcpClients
+      {} as never, // aiAgentRoleRepo
+      pageRepo as never, // pageRepo
+      {} as never, // pageAccess
+      {} as never, // environment
+    );
+    return { svc, exportPageMarkdown, upsert, findByChatPage };
+  }
+
+  const args = () =>
+    [
+      'chat1',
+      'p1',
+      { id: 'ws1' } as never,
+      { id: 'u1' } as never,
+      'sess',
+    ] as const;
+
+  it('skips export + upsert when the snapshot is already at this page version', async () => {
+    const t = new Date('2026-07-07T10:00:00Z');
+    const { svc, exportPageMarkdown, upsert } = makeSvc(
+      { pageUpdatedAt: t, contentMd: '# md' },
+      t,
+    );
+    await (svc as unknown as { snapshotOpenPage: (...a: unknown[]) => Promise<void> })
+      .snapshotOpenPage(...args());
+    expect(exportPageMarkdown).not.toHaveBeenCalled();
+    expect(upsert).not.toHaveBeenCalled();
+  });
+
+  it('exports + upserts when the page advanced since the snapshot', async () => {
+    const { svc, exportPageMarkdown, upsert } = makeSvc(
+      { pageUpdatedAt: new Date('2026-07-07T10:00:00Z'), contentMd: 'old' },
+      new Date('2026-07-07T11:00:00Z'),
+    );
+    await (svc as unknown as { snapshotOpenPage: (...a: unknown[]) => Promise<void> })
+      .snapshotOpenPage(...args());
+    expect(exportPageMarkdown).toHaveBeenCalledTimes(1);
+    expect(upsert).toHaveBeenCalledTimes(1);
+  });
+
+  it('seeds (exports + upserts) on the first turn (no snapshot yet)', async () => {
+    const { svc, exportPageMarkdown, upsert } = makeSvc(
+      undefined,
+      new Date('2026-07-07T10:00:00Z'),
+    );
+    await (svc as unknown as { snapshotOpenPage: (...a: unknown[]) => Promise<void> })
+      .snapshotOpenPage(...args());
+    expect(exportPageMarkdown).toHaveBeenCalledTimes(1);
+    expect(upsert).toHaveBeenCalledTimes(1);
+  });
+});
+
+// #490 deferred-tool activation persisted across turns.
+describe('seedActivatedTools', () => {
+  const valid = new Set(['Search_web', 'getPageJson', 'diffPageVersions']);
+
+  it('seeds from persisted metadata, intersected with current valid names', () => {
+    expect(
+      seedActivatedTools(
+        { activatedTools: ['Search_web', 'getPageJson'] },
+        valid,
+      ),
+    ).toEqual(['Search_web', 'getPageJson']);
+  });
+
+  it('drops a stored tool that is no longer valid (allowlist/role changed)', () => {
+    // 'Habr_publish' was activated before but is not in the current allowlist.
+    expect(
+      seedActivatedTools({ activatedTools: ['Search_web', 'Habr_publish'] }, valid),
+    ).toEqual(['Search_web']);
+  });
+
+  it('is empty/robust for missing, non-array, or unknown-shaped metadata', () => {
+    expect(seedActivatedTools(undefined, valid)).toEqual([]);
+    expect(seedActivatedTools({}, valid)).toEqual([]);
+    expect(seedActivatedTools({ activatedTools: 'nope' }, valid)).toEqual([]);
+    expect(
+      seedActivatedTools({ activatedTools: [1, 'getPageJson', null] }, valid),
+    ).toEqual(['getPageJson']);
+  });
+
+  it('de-duplicates stored names', () => {
+    expect(
+      seedActivatedTools(
+        { activatedTools: ['getPageJson', 'getPageJson'] },
+        valid,
+      ),
+    ).toEqual(['getPageJson']);
+  });
+});
+
+describe('lastAssistantReplayOverflow', () => {
+  const row = (
+    role: string,
+    metadata: Record<string, unknown> | null,
+  ): AiChatMessage => ({ role, metadata }) as unknown as AiChatMessage;
+
+  it('is true only when the LAST assistant turn overflowed', () => {
+    expect(
+      lastAssistantReplayOverflow([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+      ]),
+    ).toBe(true);
+    // A recovered (later, non-overflow) assistant turn clears it.
+    expect(
+      lastAssistantReplayOverflow([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+        row('assistant', { contextTokens: 5 }),
+      ]),
+    ).toBe(false);
+    expect(lastAssistantReplayOverflow([])).toBe(false);
+  });
 });
 
 describe('rowToUiMessage', () => {
@@ -618,6 +908,23 @@ describe('flushAssistant', () => {
     expect(flushed.metadata.error).toBe('boom');
   });
 
+  // #490 observability: the replay budgeter's decision is stamped on the turn.
+  it('records replayTrimmedToTokens + replayOverflow when provided', () => {
+    const f = flushAssistant([], '', 'error', {
+      error: 'ctx',
+      replayTrimmedToTokens: 42_000,
+      replayOverflow: true,
+    });
+    expect(f.metadata.replayTrimmedToTokens).toBe(42_000);
+    expect(f.metadata.replayOverflow).toBe(true);
+  });
+
+  it('omits the replay metadata when not provided', () => {
+    const f = flushAssistant([], '', 'completed', { finishReason: 'stop' });
+    expect('replayTrimmedToTokens' in f.metadata).toBe(false);
+    expect('replayOverflow' in f.metadata).toBe(false);
+  });
+
   // #274 observability: the page-change diff the agent saw this turn is persisted
   // to metadata.pageChanged when a non-empty diff was injected, and omitted when
   // the diff is empty/whitespace or the arg is not supplied.
@@ -1315,8 +1622,12 @@ describe('AiChatService page-change lifecycle (#274)', () => {
 describe('isInterruptResume', () => {
   // history tail is the just-inserted user row; [len-2] is the previous turn.
   const withPrev = (
-    prev: { role: string; status?: string | null } | null,
-  ): Array<{ role: string; status?: string | null }> =>
+    prev: {
+      role: string;
+      status?: string | null;
+      metadata?: unknown;
+    } | null,
+  ): Array<{ role: string; status?: string | null; metadata?: unknown }> =>
     prev
       ? [prev, { role: 'user', status: null }]
       : [{ role: 'user', status: null }];
@@ -1357,6 +1668,33 @@ describe('isInterruptResume', () => {
   it('false when there is no preceding turn (only the new user row)', () => {
     expect(isInterruptResume(withPrev(null), true)).toBe(false);
   });
+
+  it('#487 EXCLUDES a reconcile stamp (finalizeFailed) — not a genuine interruption', () => {
+    // A row a reconcile settled to 'aborted' carries metadata.finalizeFailed. It
+    // must NOT be treated as an interrupt-resume (that would inject a false
+    // "you were interrupted" note), even though its status is 'aborted'.
+    expect(
+      isInterruptResume(
+        withPrev({
+          role: 'assistant',
+          status: 'aborted',
+          metadata: { finalizeFailed: true },
+        }),
+        true,
+      ),
+    ).toBe(false);
+    // A genuine abort (no finalizeFailed) still counts.
+    expect(
+      isInterruptResume(
+        withPrev({
+          role: 'assistant',
+          status: 'aborted',
+          metadata: { parts: [] },
+        }),
+        true,
+      ),
+    ).toBe(true);
+  });
 });
 
 /**
@@ -1409,7 +1747,7 @@ describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', ()
   }
 
   // Wire only the deps reached on the way to the pipe call, plus a spy registry.
-  function makeService(opts: { resumable: boolean }) {
+  function makeService(opts: { resumable: boolean; history?: unknown[] }) {
     const aiChatRepo = {
       findById: jest.fn(async () => ({ id: 'chat-1', workspaceId: 'ws-1' })),
       insert: jest.fn(),
@@ -1417,8 +1755,11 @@ describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', ()
     const aiChatMessageRepo = {
       // Both the user insert and the assistant seed return the same row id.
       insert: jest.fn(async () => ({ id: 'msg-1' })),
-      findAllByChat: jest.fn(async () => []),
+      findAllByChat: jest.fn(async () => opts.history ?? []),
       update: jest.fn(async () => ({ id: 'msg-1' })),
+      // #487: the terminal owner-write + the opportunistic reconcile query.
+      finalizeOwner: jest.fn(async () => ({ id: 'msg-1' })),
+      findStreamingWithTerminalRun: jest.fn(async () => []),
     };
     const aiSettings = { resolve: jest.fn(async () => ({})) };
     const tools = { forUser: jest.fn(async () => ({})) };
@@ -1453,7 +1794,7 @@ describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', ()
       } as never,
       streamRegistry as never,
     );
-    return { svc, streamRegistry };
+    return { svc, streamRegistry, aiChatMessageRepo };
   }
 
   const body = {
@@ -1536,6 +1877,86 @@ describe('AiChatService.stream — resumable pipe options (#184 phase 1.5)', ()
     await expect(drive(svc, makeRunHooks())).rejects.toThrow('boom');
     expect(streamRegistry.abortEntry).toHaveBeenCalledWith('chat-1', 'run-1');
   });
+
+  // #489 REGRESSION (against the REAL convertToModelMessages — not mocked here):
+  // a persisted history row whose parts contain a `null` element makes the real
+  // convertToModelMessages THROW ("Cannot read properties of null"). Pre-fix that
+  // 500-ed every turn forever and each retry appended a duplicate user row. The
+  // fix converts BEFORE the insert and isolates the poisoned row per-row, degrading
+  // it to text with a "[tool context omitted]" marker. Assert the turn still runs,
+  // the marker reaches the model, and exactly ONE user row is inserted.
+  it('#489: a poisoned OLD-history row keeps the chat working; the marker reaches the model; one user insert', async () => {
+    const { svc, aiChatMessageRepo } = makeService({
+      resumable: false,
+      history: [
+        {
+          id: 'old-1',
+          role: 'assistant',
+          content: 'earlier answer',
+          // A null part is the poison: rowToUiMessage keeps it (the array is
+          // non-empty) and the real convertToModelMessages throws on it.
+          metadata: { parts: [{ type: 'text', text: 'earlier answer' }, null] },
+          status: 'completed',
+        },
+      ],
+    });
+    // Must NOT throw — the poisoned row is degraded, not fatal.
+    await drive(svc, makeRunHooks());
+    expect(streamTextMock).toHaveBeenCalledTimes(1);
+    const passedMessages = streamTextMock.mock.calls[0][0].messages;
+    const serialized = JSON.stringify(passedMessages);
+    // The model sees the truncation marker (silent tool-context loss is not ok)
+    // AND the row's readable text is preserved alongside it.
+    expect(serialized).toContain('[tool context omitted]');
+    expect(serialized).toContain('earlier answer');
+    // Exactly ONE user row inserted (no duplicate), inserted AFTER conversion.
+    const userInserts = aiChatMessageRepo.insert.mock.calls
+      .map((c: unknown[]) => c[0] as { role?: string })
+      .filter((r) => r.role === 'user');
+    expect(userInserts).toHaveLength(1);
+  });
+
+  // #489: client-supplied non-text parts (a tool-part in `input-available`, the
+  // exact "bricking" payload) are dropped ON RECEIPT — never persisted — so they
+  // can never poison future turns. Only the text survives into metadata.parts.
+  it('#489: a non-text client part is stripped before persist (only text survives)', async () => {
+    const { svc, aiChatMessageRepo } = makeService({ resumable: false });
+    await svc.stream({
+      user: { id: 'u1' } as never,
+      workspace: { id: 'ws-1' } as never,
+      sessionId: 's1',
+      body: {
+        chatId: 'chat-1',
+        messages: [
+          {
+            id: 'm1',
+            role: 'user',
+            parts: [
+              { type: 'text', text: 'hello' },
+              // untrusted tool-part — must be dropped, never persisted
+              {
+                type: 'tool-getPage',
+                toolCallId: 't1',
+                state: 'input-available',
+                input: { pageId: 'p' },
+              },
+            ],
+          },
+        ],
+      } as never,
+      res: makeRes() as never,
+      signal: new AbortController().signal,
+      model: {} as never,
+      role: null,
+      runHooks: makeRunHooks() as never,
+    });
+    const userInsert = aiChatMessageRepo.insert.mock.calls
+      .map((c: unknown[]) => c[0] as { role?: string; metadata?: unknown })
+      .find((r) => r.role === 'user');
+    const parts = (userInsert?.metadata as { parts?: Array<{ type: string }> })
+      ?.parts;
+    expect(parts).toEqual([{ type: 'text', text: 'hello' }]);
+  });
 });
 
 /**
@@ -1623,6 +2044,19 @@ describe('AiChatService.stream — token-degeneration reaction (#444)', () => {
           return { id };
         },
       ),
+      // #487: the terminal owner-write records into the SAME `updated` recorder so
+      // assertions on the terminal 'completed'/'error'/'aborted' write still hold.
+      finalizeOwner: jest.fn(
+        async (
+          id: string,
+          workspaceId: string,
+          patch: Record<string, unknown>,
+        ) => {
+          updated.push({ id, workspaceId, patch });
+          return { id };
+        },
+      ),
+      findStreamingWithTerminalRun: jest.fn(async () => []),
     };
     const aiSettings = { resolve: jest.fn(async () => ({})) };
     const tools = { forUser: jest.fn(async () => ({})) };
@@ -1882,3 +2316,148 @@ describe('AiChatService.stream — token-degeneration reaction (#444)', () => {
     expect(patch.content).not.toContain(STEP_LIMIT_NO_ANSWER_MARKER);
   });
 });
+
+// #487 F3 — the reconcile() / reconcileChat() ORCHESTRATORS. The individual
+// clauses are exercised elsewhere; these pin the production orchestration the
+// per-clause specs do not: the clause ORDER, the per-clause try/catch ISOLATION
+// (one clause throwing must NOT abort the others), and reconcileChat() (which runs
+// at the start of every turn and was entirely uncovered).
+describe('AiChatService.reconcile / reconcileChat orchestrators (#487 F3)', () => {
+  let warnSpy: jest.SpyInstance;
+  beforeEach(() => {
+    // Silence the intentional clause-failure warnings (kept out of test output).
+    warnSpy = jest
+      .spyOn(Logger.prototype, 'warn')
+      .mockImplementation(() => undefined);
+  });
+  afterEach(() => {
+    warnSpy.mockRestore();
+  });
+
+  function makeService(opts: {
+    messageRepo?: Record<string, jest.Mock>;
+    runService?: Record<string, jest.Mock>;
+  }) {
+    const aiChatMessageRepo = {
+      findStreamingWithTerminalRun: jest.fn(async () => []),
+      stampTerminalIfStreaming: jest.fn(async () => undefined),
+      sweepStreamingWithoutActiveRun: jest.fn(async () => 0),
+      ...(opts.messageRepo ?? {}),
+    };
+    const aiChatRunService = opts.runService
+      ? {
+          zombieRunIds: jest.fn(() => []),
+          settleZombie: jest.fn(async () => true),
+          reconcileStaleRuns: jest.fn(async () => 0),
+          ...opts.runService,
+        }
+      : undefined;
+    const svc = new AiChatService(
+      {} as never, // ai
+      {} as never, // aiChatRepo
+      aiChatMessageRepo as never,
+      {} as never, // aiChatPageSnapshotRepo
+      {} as never, // aiSettings
+      {} as never, // tools
+      {} as never, // mcpClients
+      {} as never, // aiAgentRoleRepo
+      {} as never, // pageRepo
+      {} as never, // pageAccess
+      {} as never, // environment
+      {} as never, // streamRegistry
+      aiChatRunService as never, // aiChatRunService (#487)
+    );
+    return { svc, aiChatMessageRepo, aiChatRunService };
+  }
+
+  it('reconcile() fires all four clauses IN ORDER (a -> b -> c -> d)', async () => {
+    const order: string[] = [];
+    const { svc } = makeService({
+      messageRepo: {
+        findStreamingWithTerminalRun: jest.fn(async () => {
+          order.push('b:find');
+          return [
+            { messageId: 'm1', workspaceId: 'ws1', runStatus: 'succeeded' },
+          ];
+        }),
+        stampTerminalIfStreaming: jest.fn(async () => {
+          order.push('b:stamp');
+        }),
+        sweepStreamingWithoutActiveRun: jest.fn(async () => {
+          order.push('d');
+          return 0;
+        }),
+      },
+      runService: {
+        zombieRunIds: jest.fn(() => ['z1']),
+        settleZombie: jest.fn(async () => {
+          order.push('a');
+          return true;
+        }),
+        reconcileStaleRuns: jest.fn(async () => {
+          order.push('c');
+          return 0;
+        }),
+      },
+    });
+
+    await svc.reconcile();
+
+    expect(order).toEqual(['a', 'b:find', 'b:stamp', 'c', 'd']);
+  });
+
+  it('a clause that THROWS does not abort the remaining clauses (per-clause try/catch isolation)', async () => {
+    const { svc, aiChatMessageRepo, aiChatRunService } = makeService({
+      messageRepo: {
+        // Clause (b) blows up mid-reconcile.
+        findStreamingWithTerminalRun: jest.fn(async () => {
+          throw new Error('clause b DB blip');
+        }),
+      },
+      runService: {
+        zombieRunIds: jest.fn(() => ['z1']),
+      },
+    });
+
+    // reconcile() must SETTLE (the clause-b failure is swallowed), not reject.
+    await expect(svc.reconcile()).resolves.toBeUndefined();
+
+    // (a) ran before (b); crucially (c) and (d) STILL ran despite (b) throwing —
+    // the property a missing try/catch would break. MUTATION-VERIFY: drop clause
+    // (b)'s try/catch and this reddens (the throw propagates, skipping c + d).
+    expect(aiChatRunService!.settleZombie).toHaveBeenCalled(); // (a)
+    expect(aiChatRunService!.reconcileStaleRuns).toHaveBeenCalled(); // (c)
+    expect(
+      aiChatMessageRepo.sweepStreamingWithoutActiveRun,
+    ).toHaveBeenCalled(); // (d)
+  });
+
+  it('reconcileChat() settles THIS chat\'s stuck streaming rows by their run status', async () => {
+    const { svc, aiChatMessageRepo } = makeService({
+      messageRepo: {
+        findStreamingWithTerminalRun: jest.fn(async () => [
+          { messageId: 'm1', workspaceId: 'ws1', runStatus: 'failed' },
+          { messageId: 'm2', workspaceId: 'ws1', runStatus: 'succeeded' },
+        ]),
+      },
+    });
+
+    await svc.reconcileChat('chat-1', 'ws1');
+
+    // Scoped to THIS chat and bounded at 50 (the user-facing opportunistic path).
+    expect(
+      aiChatMessageRepo.findStreamingWithTerminalRun,
+    ).toHaveBeenCalledWith(50, { chatId: 'chat-1', workspaceId: 'ws1' });
+    // failed-run -> 'error'; every other terminal status -> 'aborted'.
+    expect(aiChatMessageRepo.stampTerminalIfStreaming).toHaveBeenCalledWith(
+      'm1',
+      'ws1',
+      'error',
+    );
+    expect(aiChatMessageRepo.stampTerminalIfStreaming).toHaveBeenCalledWith(
+      'm2',
+      'ws1',
+      'aborted',
+    );
+  });
+});

apps/server/src/core/ai-chat/ai-chat.write-volume.spec.ts

@@ -0,0 +1,209 @@
+import { randomBytes } from 'crypto';
+import { Client } from 'pg';
+import { flushAssistant, serializeSteps } from './ai-chat.service';
+
+/**
+ * #490 write-volume regression — an OBSERVABLE-PROPERTY test on a LIVE Postgres,
+ * not "bytes through a mock repo" (a mock measures exactly the thing that does not
+ * hurt). It drives a realistic 50-step run where each step returns a ~100 KB tool
+ * output and, at every `onStepFinish`, UPDATEs the assistant row the way the
+ * service does — then reads the REAL write volume via the `pg_current_wal_lsn()`
+ * delta around the run.
+ *
+ * The property proven: v2 stores each tool OUTPUT only in `metadata.parts`, no
+ * longer ALSO in the `tool_calls` trace. So:
+ *   1. the trace (`tool_calls`) column's write volume is now O(Σ steps) — tiny,
+ *      linear outcome flags — vs the pre-#490 O(N²) that re-persisted every prior
+ *      output on every step; and
+ *   2. the FULL-row write volume drops sharply (the duplicated output copy is gone).
+ *
+ * Connects to the local gitmost test Postgres (docker `gitmost-test-pg` on :5432);
+ * SKIPS cleanly when that DB is not reachable so it never breaks a DB-less CI.
+ */
+const CONN =
+  process.env.WAL_TEST_DATABASE_URL ??
+  'postgresql://docmost:docmost_dev_pw@localhost:5432/docmost';
+
+// A step whose tool output is ~100 KB (a page read), in the SDK StepLike shape.
+// The body is INCOMPRESSIBLE random text — a `'x'.repeat()` filler would TOAST-
+// compress to nothing and hide the real write volume (a page body does not).
+function makeStep(i: number, outputBytes = 100_000) {
+  const body = randomBytes(Math.ceil(outputBytes * 0.75)).toString('base64');
+  return {
+    text: `step ${i} reasoning`,
+    toolCalls: [{ toolCallId: `c${i}`, toolName: 'getPage', input: { id: `p${i}` } }],
+    toolResults: [
+      {
+        toolCallId: `c${i}`,
+        toolName: 'getPage',
+        output: { id: `p${i}`, title: `Page ${i}`, body },
+      },
+    ],
+  };
+}
+
+// The pre-#490 (v1) trace: outputs stored a SECOND time in `tool_calls`
+// (the duplication #490 removed). Mirrors the OLD serializeSteps shape.
+function v1Trace(steps: ReturnType<typeof makeStep>[]): unknown {
+  const calls: unknown[] = [];
+  for (const s of steps) {
+    for (const c of s.toolCalls) calls.push({ toolName: c.toolName, input: c.input });
+    for (const r of s.toolResults)
+      calls.push({ toolName: r.toolName, output: r.output });
+  }
+  return calls;
+}
+
+async function walDelta(
+  client: Client,
+  fn: () => Promise<void>,
+): Promise<number> {
+  const before = (await client.query('SELECT pg_current_wal_lsn() AS l')).rows[0]
+    .l as string;
+  await fn();
+  // NOTE: do NOT pg_switch_wal() here — a segment switch pads the LSN to the next
+  // 16 MB boundary and would swamp the actual write delta. The raw LSN advances by
+  // the bytes of WAL emitted, which is exactly what we want to measure.
+  const after = (await client.query('SELECT pg_current_wal_lsn() AS l')).rows[0]
+    .l as string;
+  return Number(
+    (await client.query('SELECT pg_wal_lsn_diff($1,$2) AS d', [after, before]))
+      .rows[0].d,
+  );
+}
+
+describe('#490 write-volume on a live Postgres (pg_current_wal_lsn delta)', () => {
+  let client: Client | undefined;
+  let available = false;
+
+  beforeAll(async () => {
+    try {
+      client = new Client(CONN);
+      await client.connect();
+      await client.query('SELECT pg_current_wal_lsn()');
+      available = true;
+    } catch {
+      available = false;
+      client = undefined;
+    }
+  });
+
+  afterAll(async () => {
+    await client?.end().catch(() => undefined);
+  });
+
+  const STEPS = 50;
+
+  it('v2 trace write volume is O(Σ steps) — a tiny fraction of the v1 duplicate', async () => {
+    if (!available || !client) {
+      console.warn('SKIP: gitmost-test-pg not reachable; skipping WAL test.');
+      return;
+    }
+    const c = client;
+    // Isolated table so we measure only the tool_calls (trace) column's writes.
+    await c.query('DROP TABLE IF EXISTS _wal_trace');
+    await c.query('CREATE TABLE _wal_trace(id int primary key, tool_calls jsonb)');
+    await c.query("INSERT INTO _wal_trace VALUES (1, '[]'::jsonb)");
+
+    const steps: ReturnType<typeof makeStep>[] = [];
+
+    // v1: each step re-persists ALL prior outputs into the trace (the O(N²) churn).
+    const v1 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await c.query('UPDATE _wal_trace SET tool_calls=$1 WHERE id=1', [
+          JSON.stringify(v1Trace(acc)),
+        ]);
+      }
+      steps.push(...acc);
+    });
+
+    await c.query("UPDATE _wal_trace SET tool_calls='[]'::jsonb WHERE id=1");
+
+    // v2: the REAL serializeSteps — outcome flags only, NO outputs.
+    const v2 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await c.query('UPDATE _wal_trace SET tool_calls=$1 WHERE id=1', [
+          JSON.stringify(serializeSteps(acc)),
+        ]);
+      }
+    });
+
+    await c.query('DROP TABLE IF EXISTS _wal_trace');
+
+    // eslint-disable-next-line no-console
+    console.log(
+      `[#490 WAL] trace column over ${STEPS} steps: v1=${(v1 / 1e6).toFixed(1)}MB ` +
+        `v2=${(v2 / 1e6).toFixed(2)}MB (${(v1 / v2).toFixed(0)}x smaller)`,
+    );
+
+    // The trace no longer carries outputs: v2 is a tiny fraction of v1's WAL.
+    expect(v2).toBeLessThan(v1 * 0.1);
+    // And v2's trace WAL is small in absolute terms — O(Σ steps) of flags, not
+    // O(N² × output). 50 steps of ~40-byte flags is well under a few MB of WAL.
+    expect(v2).toBeLessThan(5_000_000);
+    // v1's duplicate alone is huge (≈ the 100 KB output re-written N² times).
+    expect(v1).toBeGreaterThan(50_000_000);
+  }, 120_000);
+
+  it('the full assistant row write drops sharply once the duplicate is gone', async () => {
+    if (!available || !client) return;
+    const c = client;
+    await c.query('DROP TABLE IF EXISTS _wal_full');
+    await c.query(
+      'CREATE TABLE _wal_full(id int primary key, content text, tool_calls jsonb, metadata jsonb, status text)',
+    );
+    await c.query("INSERT INTO _wal_full VALUES (1, '', '[]'::jsonb, '{}'::jsonb, 'streaming')");
+
+    const writeRow = async (patch: {
+      content: string;
+      toolCalls: unknown;
+      metadata: unknown;
+      status: string;
+    }) =>
+      c.query(
+        'UPDATE _wal_full SET content=$1, tool_calls=$2, metadata=$3, status=$4 WHERE id=1',
+        [
+          patch.content,
+          JSON.stringify(patch.toolCalls ?? null),
+          JSON.stringify(patch.metadata),
+          patch.status,
+        ],
+      );
+
+    // v2 (real flushAssistant): outputs live once, in metadata.parts.
+    const v2 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await writeRow(flushAssistant(acc as never, '', 'streaming'));
+      }
+    });
+
+    await c.query("UPDATE _wal_full SET content='', tool_calls='[]'::jsonb, metadata='{}'::jsonb WHERE id=1");
+
+    // v1: same row PLUS the duplicated outputs in the trace column.
+    const v1 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        const f = flushAssistant(acc as never, '', 'streaming');
+        await writeRow({ ...f, toolCalls: v1Trace(acc) });
+      }
+    });
+
+    await c.query('DROP TABLE IF EXISTS _wal_full');
+
+    // eslint-disable-next-line no-console
+    console.log(
+      `[#490 WAL] full row over ${STEPS} steps: v1=${(v1 / 1e6).toFixed(1)}MB ` +
+        `v2=${(v2 / 1e6).toFixed(1)}MB (saved ${((1 - v2 / v1) * 100).toFixed(0)}%)`,
+    );
+
+    // Removing the duplicated trace copy is a large, real write-volume reduction.
+    expect(v2).toBeLessThan(v1 * 0.75);
+  }, 120_000);
+});

apps/server/src/core/ai-chat/external-mcp/mcp-clients.recovery.spec.ts

@@ -0,0 +1,261 @@
+import { errors } from 'undici';
+import {
+  McpClientsService,
+  isRetryableConnectError,
+} from './mcp-clients.service';
+
+/**
+ * #489 — external-MCP in-run transport recovery.
+ *
+ * The transport-error classification + retry gate are exercised against the REAL
+ * undici error CLASSES prod throws (`errors.SocketError` / `errors.BodyTimeoutError`,
+ * carrying the true `UND_ERR_*` codes and class names), wrapped EXACTLY as undici's
+ * `fetch` wraps them — a `TypeError('fetch failed'|'terminated')` whose `.cause` is
+ * the undici error. These are the real classes, not hand-rolled `{code:'...'}`
+ * mocks: constructing the genuine class is what makes this a faithful test of the
+ * prod predicate (epic root-cause #4 — a mock-shaped predicate would leave the
+ * evict/retry path silently dead in production while CI stays green). We construct
+ * rather than drive a live fetch because Jest's environment degrades the live-fetch
+ * error to a generic `Error` cause (no undici code), which would NOT be the prod
+ * shape.
+ */
+
+/** A REAL undici socket reset, wrapped as fetch wraps it. */
+function realSocketResetError(): unknown {
+  const err = new TypeError('fetch failed');
+  (err as { cause?: unknown }).cause = new errors.SocketError('other side closed');
+  return err;
+}
+
+/** A REAL undici body timeout, wrapped as fetch wraps it. */
+function realBodyTimeoutError(): unknown {
+  const err = new TypeError('terminated');
+  (err as { cause?: unknown }).cause = new errors.BodyTimeoutError();
+  return err;
+}
+
+type FakeServer = {
+  id: string;
+  name: string;
+  transport: 'http' | 'sse';
+  url: string;
+  headersEnc: string | null;
+  toolAllowlist: string[] | null;
+  instructions: string | null;
+};
+
+const server = (over: Partial<FakeServer> = {}): FakeServer => ({
+  id: 's1',
+  name: 'srv',
+  transport: 'http',
+  url: 'http://example.test/mcp',
+  headersEnc: null,
+  toolAllowlist: null,
+  instructions: null,
+  ...over,
+});
+
+function buildService(servers: FakeServer[], trusted = false) {
+  const repo = { listEnabled: jest.fn().mockResolvedValue(servers) };
+  const service = new McpClientsService(repo as never, {} as never);
+  // Seed a DETERMINISTIC write-class map so the retry gate is controlled here
+  // (the production map loads from @docmost/mcp via a dynamic ESM import). getPage
+  // is a read, patchNode is a write — the real classifications.
+  (
+    service as unknown as { writeClassMapPromise: Promise<unknown> }
+  ).writeClassMapPromise = Promise.resolve({
+    getPage: 'readOnly',
+    patchNode: 'write',
+  });
+  // The service only APPLIES that map to a TRUSTED internal Docmost server
+  // (isInternalDocmostServer, really false for every third-party row). A retry
+  // test needs a trusted server to exercise the readOnly-retry path at all, so it
+  // passes trusted=true to model a Docmost-origin server; the third-party
+  // double-apply test leaves it at the real value (false).
+  if (trusted) {
+    jest
+      .spyOn(
+        service as unknown as {
+          isInternalDocmostServer: (s: FakeServer) => boolean;
+        },
+        'isInternalDocmostServer',
+      )
+      .mockReturnValue(true);
+  }
+  return { service, repo };
+}
+
+/** Spy the private `connect` so each call yields a controlled fake client whose
+ *  single tool's execute is the supplied function. Returns the connect spy. */
+function stubConnect(
+  service: McpClientsService,
+  toolName: string,
+  execs: Array<(...a: unknown[]) => Promise<unknown>>,
+) {
+  let n = 0;
+  return jest
+    .spyOn(
+      service as unknown as { connect: (s: FakeServer) => Promise<unknown> },
+      'connect',
+    )
+    .mockImplementation(async () => {
+      const exec = execs[Math.min(n, execs.length - 1)];
+      n += 1;
+      return {
+        tools: async () => ({ [toolName]: { description: 'x', execute: exec } }),
+        close: jest.fn().mockResolvedValue(undefined),
+      };
+    });
+}
+
+const opts = (abortSignal?: AbortSignal) =>
+  ({ toolCallId: 't', messages: [], abortSignal }) as never;
+
+describe('isRetryableConnectError (#489, REAL error shapes)', () => {
+  it('classifies a real undici socket reset and body timeout as retryable', async () => {
+    const socketErr = await realSocketResetError();
+    const bodyErr = await realBodyTimeoutError();
+    expect(isRetryableConnectError(socketErr)).toBe(true);
+    expect(isRetryableConnectError(bodyErr)).toBe(true);
+    // Unwraps a wrapped cause chain (e.g. an MCPClientError around the socket err).
+    const wrapped = new Error('mcp call failed');
+    (wrapped as { cause?: unknown }).cause = socketErr;
+    expect(isRetryableConnectError(wrapped)).toBe(true);
+  });
+
+  it('does NOT classify an application-level error as a transport break', () => {
+    expect(isRetryableConnectError(new Error('validation failed'))).toBe(false);
+    expect(isRetryableConnectError({ name: 'HttpError', status: 400 })).toBe(false);
+    expect(isRetryableConnectError(undefined)).toBe(false);
+    expect(isRetryableConnectError('boom')).toBe(false);
+  });
+});
+
+describe('McpClientsService in-run transport recovery (#489)', () => {
+  afterEach(() => jest.restoreAllMocks());
+
+  it('a readOnly tool whose transport breaks reconnects and retries WITHIN the same run', async () => {
+    const realErr = await realSocketResetError();
+    const { service } = buildService([server()], true);
+    const first = jest.fn().mockRejectedValue(realErr);
+    const second = jest.fn().mockResolvedValue({ ok: true });
+    const connectSpy = stubConnect(service, 'getPage', [first, second]);
+
+    const toolset = await service.toolsFor('ws-1');
+    const tool = toolset.tools['srv_getPage'];
+    const result = await (tool.execute as (a: unknown, o: unknown) => Promise<unknown>)(
+      { pageId: 'p' },
+      opts(),
+    );
+
+    // The repeat call within the run got a LIVE client and succeeded.
+    expect(result).toEqual({ ok: true });
+    expect(first).toHaveBeenCalledTimes(1);
+    expect(second).toHaveBeenCalledTimes(1);
+    // Exactly one reconnect was minted (initial build connect + one recovery).
+    expect(connectSpy).toHaveBeenCalledTimes(2);
+    // The run accumulated BOTH leases (old + reconnected) — released together at end.
+    expect(toolset.clients).toHaveLength(2);
+    await Promise.all(toolset.clients.map((c) => c.close()));
+  });
+
+  it('a WRITE tool does NOT auto-retry on a transport error (indeterminate)', async () => {
+    const realErr = await realSocketResetError();
+    const { service } = buildService([server()], true);
+    const exec = jest.fn().mockRejectedValue(realErr);
+    const connectSpy = stubConnect(service, 'patchNode', [exec]);
+
+    const toolset = await service.toolsFor('ws-2');
+    const tool = toolset.tools['srv_patchNode'];
+    await expect(
+      (tool.execute as (a: unknown, o: unknown) => Promise<unknown>)(
+        { pageId: 'p' },
+        opts(),
+      ),
+    ).rejects.toThrow(/MAY have already applied/);
+
+    // Called exactly once — NO blind retry (avoids double-apply, the #435 class).
+    expect(exec).toHaveBeenCalledTimes(1);
+    // No fresh connection was minted for a write.
+    expect(connectSpy).toHaveBeenCalledTimes(1);
+    await Promise.all(toolset.clients.map((c) => c.close()));
+  });
+
+  it('does NOT retry (or reconnect) after the run is aborted (Stop)', async () => {
+    const realErr = await realSocketResetError();
+    const { service } = buildService([server()], true);
+    const controller = new AbortController();
+    // The transport error arrives, but the run was Stopped in the same tick.
+    const first = jest.fn().mockImplementation(async () => {
+      controller.abort();
+      throw realErr;
+    });
+    const second = jest.fn().mockResolvedValue({ ok: true });
+    const connectSpy = stubConnect(service, 'getPage', [first, second]);
+
+    const toolset = await service.toolsFor('ws-3');
+    const tool = toolset.tools['srv_getPage'];
+    await expect(
+      (tool.execute as (a: unknown, o: unknown) => Promise<unknown>)(
+        { pageId: 'p' },
+        opts(controller.signal),
+      ),
+    ).rejects.toBeDefined();
+
+    // getPage IS readOnly, but the Stop blocks the retry — no second call, no mint.
+    expect(second).not.toHaveBeenCalled();
+    expect(connectSpy).toHaveBeenCalledTimes(1);
+    await Promise.all(toolset.clients.map((c) => c.close()));
+  });
+
+  it('an app-level (non-transport) tool error is surfaced verbatim, never retried', async () => {
+    const { service } = buildService([server()], true);
+    const appErr = new Error('tool says: bad input');
+    const exec = jest.fn().mockRejectedValue(appErr);
+    const connectSpy = stubConnect(service, 'getPage', [exec]);
+
+    const toolset = await service.toolsFor('ws-4');
+    const tool = toolset.tools['srv_getPage'];
+    await expect(
+      (tool.execute as (a: unknown, o: unknown) => Promise<unknown>)(
+        { pageId: 'p' },
+        opts(),
+      ),
+    ).rejects.toThrow('tool says: bad input');
+    expect(exec).toHaveBeenCalledTimes(1);
+    expect(connectSpy).toHaveBeenCalledTimes(1); // no reconnect for an app error
+    await Promise.all(toolset.clients.map((c) => c.close()));
+  });
+
+  // #489 (review, MEDIUM) — the Docmost write-class map keys by DOCMOST tool
+  // names; a THIRD-PARTY server may name a WRITE tool `getPage` (a Docmost read
+  // name). It must NOT inherit readOnly and must NOT auto-retry on a transport
+  // error — a blind retry of that write is a double-apply (the #435 class). Here
+  // the server is UNTRUSTED (buildService default, isInternalDocmostServer=false),
+  // so the map is not applied and `getPage` classifies as a write.
+  //
+  // MUTATION-VERIFY: forcing the server "trusted" (buildService(..., true)) makes
+  // `getPage` inherit readOnly -> it WOULD reconnect+retry (connect twice) and the
+  // assertions below fail — i.e. removing the trust scope re-opens the bug.
+  it('a THIRD-PARTY WRITE tool named like a Docmost read does NOT auto-retry (no double-apply)', async () => {
+    const realErr = await realSocketResetError();
+    // Untrusted: default trusted=false — a real third-party server.
+    const { service } = buildService([server()]);
+    const exec = jest.fn().mockRejectedValue(realErr);
+    const connectSpy = stubConnect(service, 'getPage', [exec, exec]);
+
+    const toolset = await service.toolsFor('ws-5');
+    const tool = toolset.tools['srv_getPage'];
+    await expect(
+      (tool.execute as (a: unknown, o: unknown) => Promise<unknown>)(
+        { pageId: 'p' },
+        opts(),
+      ),
+    ).rejects.toThrow(/MAY have already applied/);
+
+    // Exactly one call, NO reconnect — the name collision granted no readOnly-retry.
+    expect(exec).toHaveBeenCalledTimes(1);
+    expect(connectSpy).toHaveBeenCalledTimes(1);
+    await Promise.all(toolset.clients.map((c) => c.close()));
+  });
+});

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

@@ -106,8 +106,11 @@ describe('McpClientsService.decryptHeaders', () => {
 
 describe('McpClientsService.guardedFetch (SSRF per-request guard)', () => {
   // The bound guardedFetch closure lives on the instance as a private field.
+  // #489 split it into per-transport HTTP/SSE bindings (they differ only in the
+  // dispatcher's bodyTimeout); the SSRF guard is identical, so testing the HTTP
+  // one is sufficient.
   const guardedFetchOf = (service: McpClientsService) =>
-    (service as unknown as { guardedFetch: typeof fetch }).guardedFetch;
+    (service as unknown as { guardedFetchHttp: typeof fetch }).guardedFetchHttp;
 
   let fetchSpy: jest.SpiedFunction<typeof fetch>;
 

apps/server/src/core/ai-chat/external-mcp/mcp-clients.service.ts

@@ -1,5 +1,6 @@
 import { isIP } from 'node:net';
 import { lookup as dnsLookup, type LookupAddress } from 'node:dns';
+import { pathToFileURL } from 'node:url';
 import { Injectable, Logger } from '@nestjs/common';
 import { type Tool, type ToolCallOptions } from 'ai';
 import { createMCPClient } from '@ai-sdk/mcp';
@@ -10,9 +11,29 @@ import {
   streamingDispatcherOptions,
   mcpStreamTimeoutMs,
   mcpCallTimeoutMs,
+  mcpSseBodyTimeoutMs,
 } from '../../../integrations/ai/ai-streaming-fetch';
 import { SecretBoxService } from '../../../integrations/crypto/secret-box';
 import { isUrlAllowed, isIpAllowed } from './ssrf-guard';
+// TYPE-ONLY (erased at compile): @docmost/mcp is ESM-only and cannot be a runtime
+// `require()` from this commonjs module (same constraint as docmost-client.loader).
+// The write-class MAP is loaded lazily via the dynamic-import trick below.
+import type { ToolWriteClass } from '@docmost/mcp';
+
+// TS(commonjs) downlevels a literal `import()` to `require()`, which cannot load
+// the ESM-only @docmost/mcp. Indirect through Function so the real dynamic
+// `import()` survives compilation (same trick as docmost-client.loader.ts).
+const esmImport = new Function(
+  'specifier',
+  'return import(specifier)',
+) as (specifier: string) => Promise<unknown>;
+
+/** Local read-only predicate — avoids a value import of the ESM-only package.
+ *  Only a pure read is retry-safe after a transport break (a write is
+ *  indeterminate). Kept in lockstep with @docmost/mcp's isRetryableWriteClass. */
+function isReadOnlyWriteClass(writeClass: ToolWriteClass | undefined): boolean {
+  return writeClass === 'readOnly';
+}
 
 /** A closable external MCP client handle. */
 export interface Closable {
@@ -81,12 +102,52 @@ const MAX_TOOL_NAME_LENGTH = 64;
  * close until the turn releases it, so a TTL expiry mid-turn never closes a
  * client a stream is still executing against.
  */
+/**
+ * Where a merged (namespaced) tool came from, so the per-run recovery wrapper
+ * (#489) can, on a transport error, reconnect THAT server and re-resolve the SAME
+ * underlying tool by its raw name. `writeClass` gates the single auto-retry (a
+ * read is retry-safe; a write is indeterminate). `serverIndex` indexes the
+ * entry's `servers` array (which server config to reconnect).
+ */
+interface ToolProvenance {
+  serverIndex: number;
+  rawName: string;
+  writeClass: ToolWriteClass | undefined;
+}
+
+/** A live reconnected server (its fresh client + raw call-timeout-wrapped tools). */
+interface RecoveredServerState {
+  client: McpClient;
+  tools: Record<string, Tool>;
+}
+
+/**
+ * Per-run, per-server recovery binding (#489). `current` is the server's LIVE
+ * target for this run: `null` means "use the ORIGINAL cached client/template";
+ * a non-null value is a reconnected throwaway client all this server's tools now
+ * call. `reconnecting` dedupes concurrent reconnects so only ONE fresh client is
+ * minted per death (a losing concurrent call awaits it and retries on the SAME
+ * new client — the CAS-by-identity rule).
+ */
+interface ServerBinding {
+  current: RecoveredServerState | null;
+  reconnecting?: Promise<RecoveredServerState>;
+}
+
 interface CacheEntry {
   tools: Record<string, Tool>;
   clients: McpClient[];
   outcomes: ServerOutcome[];
   /** Prompt guidance for qualifying servers (see McpServerInstruction). */
   instructions: McpServerInstruction[];
+  /**
+   * The enabled server configs used to build this entry (#489), so the per-run
+   * recovery wrapper can reconnect a specific server by index. Parallel to the
+   * indices referenced by {@link toolMeta}.
+   */
+  servers: AiMcpServer[];
+  /** merged-tool-key -> provenance (#489), for the per-run recovery wrapper. */
+  toolMeta: Record<string, ToolProvenance>;
   expiresAt: number;
   /** Active leases (turns currently using these clients). */
   refCount: number;
@@ -120,20 +181,82 @@ export class McpClientsService {
    */
   private readonly cache = new Map<string, Promise<CacheEntry>>();
   /**
-   * A single shared SSRF-pinned dispatcher for ALL outbound external-MCP fetches.
-   * Its custom connect.lookup runs per connection, so one instance safely guards
-   * every server's connections (we never connect to an unvalidated IP).
+   * SSRF-pinned dispatchers for outbound external-MCP fetches. Both use the SAME
+   * custom connect.lookup (so every connection is IP-validated), but carry a
+   * DIFFERENT `bodyTimeout` (#489): the HTTP (streamable) transport opens a fresh
+   * request per call, so it keeps the tight silence timeout; the SSE transport
+   * holds ONE long-lived body open across many calls, so a >1-min idle BETWEEN
+   * calls is LEGITIMATE and must not break the socket — it gets a much larger
+   * bodyTimeout. (headersTimeout stays tight on both.)
    */
-  private readonly dispatcher: Dispatcher = buildPinnedDispatcher();
-  /** guardedFetch bound to the pinned dispatcher; reused by every transport. */
-  private readonly guardedFetch: typeof fetch = (input, init) =>
-    guardedFetch(this.dispatcher, input, init);
+  private readonly dispatcherHttp: Dispatcher = buildPinnedDispatcher(
+    mcpStreamTimeoutMs(),
+  );
+  private readonly dispatcherSse: Dispatcher = buildPinnedDispatcher(
+    mcpSseBodyTimeoutMs(),
+  );
+  /** guardedFetch bound to each dispatcher; picked by transport type in connect(). */
+  private readonly guardedFetchHttp: typeof fetch = (input, init) =>
+    guardedFetch(this.dispatcherHttp, input, init);
+  private readonly guardedFetchSse: typeof fetch = (input, init) =>
+    guardedFetch(this.dispatcherSse, input, init);
+
+  /**
+   * Memoized write-class map (#489), loaded lazily from @docmost/mcp via the
+   * dynamic-import trick. Keyed by tool name (=== mcpName). A tool NOT in the map
+   * (any third-party external MCP tool) classifies as `undefined` -> treated as a
+   * write by the retry gate (the safe default: never blind-retry an unknown tool).
+   * On any load failure the map is `{}` (every tool -> no auto-retry), so a
+   * missing/older @docmost/mcp build only DISABLES retries, never mis-retries.
+   */
+  private writeClassMapPromise: Promise<Record<string, ToolWriteClass>> | null =
+    null;
 
   constructor(
     private readonly repo: AiMcpServerRepo,
     private readonly secretBox: SecretBoxService,
   ) {}
 
+  /**
+   * Whether an external MCP server is the TRUSTED internal Docmost MCP server —
+   * the only server whose tools may be classified by the Docmost write-class map
+   * (#489 review). Today this is ALWAYS false: every `ai_mcp_servers` row is an
+   * admin-configured THIRD-PARTY endpoint (there is no builtin/self flag, sentinel
+   * URL, or synthetic server in this path — Docmost's OWN tools are exposed via the
+   * separate in-app tools path, never through this external-MCP client). So no
+   * third-party tool can inherit `readOnly` by a name collision with a Docmost read
+   * tool, and none is ever auto-retried on a transport error (which would risk a
+   * double-apply — the #435 class). Flip this (an explicit `kind`/`isBuiltin`
+   * column, or a configured self-MCP URL) if a trusted internal server is ever
+   * introduced. A method (not a free function) so it is a single, mockable seam.
+   */
+  private isInternalDocmostServer(_server: AiMcpServer): boolean {
+    return false;
+  }
+
+  /** Lazily load + memoize the shared write-class map (see the field doc). */
+  private getWriteClassMap(): Promise<Record<string, ToolWriteClass>> {
+    if (!this.writeClassMapPromise) {
+      this.writeClassMapPromise = (async () => {
+        try {
+          const entry = require.resolve('@docmost/mcp');
+          const mod = (await esmImport(pathToFileURL(entry).href)) as {
+            SHARED_TOOL_WRITE_CLASS?: Record<string, ToolWriteClass>;
+          };
+          return mod.SHARED_TOOL_WRITE_CLASS ?? {};
+        } catch (err) {
+          this.logger.warn(
+            `Could not load MCP write-class map (auto-retry disabled): ${shortError(
+              err,
+            )}`,
+          );
+          return {};
+        }
+      })();
+    }
+    return this.writeClassMapPromise;
+  }
+
   /**
    * Build (or reuse a cached) external toolset for a workspace. Returns the
    * merged tools, the open client handles to release, and per-server outcomes.
@@ -162,11 +285,37 @@ export class McpClientsService {
         }
       },
     };
-    // One release handle drives the whole leased entry; closing it releases all
-    // underlying clients together (they share the same lease lifecycle).
+
+    // #489: the run accumulates a SET of leases — the primary cache lease PLUS any
+    // throwaway client minted by an in-run transport-recovery reconnect. They are
+    // NEVER released mid-run (releasing a swapped-out client while a concurrent
+    // in-flight call still holds it would INDUCE a second failure); the caller
+    // releases the WHOLE set together at turn-end. A recovery reconnect pushes its
+    // lease onto this live array, which the consumer closes over.
+    const leaseSet: Closable[] = [release];
+
+    // #489: per-RUN transport-recovery binding, one per server, SHARED by all of
+    // that server's tools so a swap by one call is seen by the next (CAS by
+    // identity). Kept per-run (here, not in the cached entry) because the binding
+    // + lease-set state is per-run.
+    const bindings = new Map<number, ServerBinding>();
+    const capMs = mcpCallTimeoutMs();
+
+    // Wrap each cached tool with the recovery layer. On a transport error a
+    // declared readOnly tool reconnects its server and retries ONCE; a write is
+    // never blind-retried (indeterminate — may have applied before the reset). A
+    // tool without provenance (a minimal stub entry in a test) passes through raw.
+    const tools: Record<string, Tool> = {};
+    for (const [key, tool] of Object.entries(entry.tools)) {
+      const meta = entry.toolMeta?.[key];
+      tools[key] = meta
+        ? this.wrapWithTransportRecovery(entry, meta, tool, leaseSet, bindings, capMs)
+        : tool;
+    }
+
     return {
-      tools: entry.tools,
-      clients: [release],
+      tools,
+      clients: leaseSet,
       outcomes: entry.outcomes,
       instructions: entry.instructions,
     };
@@ -254,6 +403,16 @@ export class McpClientsService {
     // Per-call total wall-clock cap, read once for this build (env-overridable).
     const callTimeoutMs = mcpCallTimeoutMs();
     const instructions: McpServerInstruction[] = [];
+    // merged-key -> provenance for the per-run recovery wrapper (#489).
+    const toolMeta: Record<string, ToolProvenance> = {};
+    // Shared Docmost write-class map (#489) — classifies a tool by its raw name.
+    // Loaded ONLY when at least one server is a TRUSTED internal Docmost server
+    // (see isInternalDocmostServer): for third-party servers the map is never
+    // applied (a name collision must not grant readOnly-retry), so we skip the
+    // dynamic ESM load entirely in that (currently universal) case.
+    const writeClassMap = servers.some((s) => this.isInternalDocmostServer(s))
+      ? await this.getWriteClassMap()
+      : null;
 
     // Per-server connect+tools result, still tagged with its server so the merge
     // below can be applied in the SAME order as `servers` (see the parallel note).
@@ -327,11 +486,23 @@ export class McpClientsService {
       // against names already merged from earlier servers, so no external
       // tool is silently overwritten on collision. The returned count drives
       // whether this server's prompt guidance is included (≥1 tool merged).
+      // #489 (review): the Docmost write-class map keys by DOCMOST tool names and
+      // may ONLY be trusted for a server KNOWN to be the internal Docmost MCP
+      // server. Every row here is an admin-configured THIRD-PARTY endpoint, so a
+      // third-party WRITE tool that happens to be named like a Docmost read
+      // (getPage, listPages, ...) must NOT inherit readOnly — that would auto-retry
+      // a mutation on a transport error (double-apply, the #435 class). Gate the
+      // map on the trust check; untrusted servers get writeClass=undefined -> the
+      // recovery wrapper treats them as writes and never auto-retries.
+      const trustWriteClass = this.isInternalDocmostServer(server);
       const merged = this.mergeNamespaced(
         tools,
         result.guarded,
         server.name,
         server.id,
+        toolMeta,
+        i,
+        trustWriteClass ? writeClassMap : null,
       );
       outcomes.push({ name: server.name, ok: true });
       // Include this server's guidance ONLY when it actually contributed at
@@ -353,6 +524,8 @@ export class McpClientsService {
       clients,
       outcomes,
       instructions,
+      servers,
+      toolMeta,
       expiresAt: Date.now() + CACHE_TTL_MS,
       refCount: 0,
       evicted: false,
@@ -379,18 +552,33 @@ export class McpClientsService {
     picked: Record<string, Tool>,
     serverName: string,
     serverId: string,
+    toolMeta: Record<string, ToolProvenance>,
+    serverIndex: number,
+    // The Docmost write-class map, or `null` for an UNTRUSTED (third-party)
+    // server whose tools must all default to write (never auto-retried).
+    writeClassMap: Record<string, ToolWriteClass> | null,
   ): { count: number; prefix: string } {
     let count = 0;
-    for (const [name, tool] of Object.entries(namespace(picked, serverName))) {
-      let key = name;
+    for (const { full, raw, tool } of namespace(picked, serverName)) {
+      let key = full;
       if (key in target) {
         const original = key;
-        key = disambiguate(name, serverId, (candidate) => candidate in target);
+        key = disambiguate(full, serverId, (candidate) => candidate in target);
         this.logger.debug(
           `External MCP tool name "${original}" collided; renamed to "${key}"`,
         );
       }
       target[key] = tool;
+      // Record provenance so the per-run recovery wrapper (#489) can reconnect
+      // this tool's server and re-resolve it by its raw name. writeClass is set
+      // ONLY from a TRUSTED (internal-Docmost) map; for a third-party server the
+      // map is null -> writeClass stays undefined -> the wrapper treats the tool
+      // as a write and never auto-retries it (no double-apply on name collision).
+      toolMeta[key] = {
+        serverIndex,
+        rawName: raw,
+        writeClass: writeClassMap ? writeClassMap[raw] : undefined,
+      };
       count += 1;
     }
     return { count, prefix: namespacePrefix(serverName) };
@@ -424,7 +612,10 @@ export class McpClientsService {
         // Defense in depth: re-validate the actual request host on EVERY fetch
         // AND pin the socket to a validated IP via the dispatcher's connect
         // lookup, closing the DNS-rebinding TOCTOU between check and connect.
-        fetch: this.guardedFetch,
+        // #489: the SSE transport uses the raised-bodyTimeout dispatcher (idle
+        // between calls is legit); HTTP uses the tight one.
+        fetch:
+          transportType === 'sse' ? this.guardedFetchSse : this.guardedFetchHttp,
       },
     })) as unknown as McpClient;
     return client;
@@ -505,6 +696,176 @@ export class McpClientsService {
     }
   }
 
+  /**
+   * Wrap one merged external tool with the per-run transport-recovery layer (#489).
+   *
+   * attempt 1 runs on the server's CURRENT binding (the cached client, or a client
+   * a sibling tool already reconnected this run). On a REAL transport error
+   * (undici/@ai-sdk socket/body-timeout shapes — {@link isRetryableConnectError},
+   * NOT a mock) and ONLY for a declared readOnly tool, it reconnects the server
+   * and retries EXACTLY ONCE on the fresh client; a write is surfaced as an
+   * indeterminate error (it may have applied before the reset — never
+   * blind-retried). A single per-call cap bounds BOTH attempts + the reconnect,
+   * and the run's abort signal is checked before the retry AND before minting a
+   * fresh connection (no connection is opened for a stopped run).
+   */
+  private wrapWithTransportRecovery(
+    entry: CacheEntry,
+    meta: ToolProvenance,
+    template: Tool,
+    leaseSet: Closable[],
+    bindings: Map<number, ServerBinding>,
+    capMs: number,
+  ): Tool {
+    const original = template.execute;
+    if (typeof original !== 'function') return template;
+    const service = this;
+    const { serverIndex, rawName, writeClass } = meta;
+
+    let binding = bindings.get(serverIndex);
+    if (!binding) {
+      binding = { current: null };
+      bindings.set(serverIndex, binding);
+    }
+    const boundBinding = binding;
+
+    const execute = async (args: unknown, options: ToolCallOptions) => {
+      // The per-call cap governs the WHOLE sequence (attempt1 + reconnect +
+      // attempt2). Compose it with the run's abort signal so a Stop or the cap
+      // ends any awaited call — @ai-sdk/mcp does not settle on abort, so we RACE.
+      const capController = new AbortController();
+      const capTimer = setTimeout(() => {
+        capController.abort(new Error(`MCP tool call timed out after ${capMs}ms`));
+      }, capMs);
+      capTimer.unref?.();
+      const runSignal = options?.abortSignal;
+      const composed = runSignal
+        ? AbortSignal.any([runSignal, capController.signal])
+        : capController.signal;
+      const stopped = () => runSignal?.aborted === true || capController.signal.aborted;
+
+      const callOn = async (
+        exec: NonNullable<Tool['execute']>,
+      ): Promise<unknown> => {
+        const aborted = new Promise<never>((_, reject) => {
+          const fail = () => reject(abortReason(composed));
+          if (composed.aborted) fail();
+          else composed.addEventListener('abort', fail, { once: true });
+        });
+        return Promise.race([exec(args, { ...options, abortSignal: composed }), aborted]);
+      };
+
+      const execFor = (
+        state: RecoveredServerState | null,
+      ): NonNullable<Tool['execute']> | undefined =>
+        state ? (state.tools[rawName]?.execute as NonNullable<Tool['execute']>) : original;
+
+      try {
+        // Snapshot the target BEFORE the call so a swap by a concurrent call is
+        // detected by identity in the catch.
+        const attemptState = boundBinding.current;
+        const attemptExec = execFor(attemptState);
+        if (typeof attemptExec !== 'function') {
+          throw new Error(`external MCP tool "${rawName}" is not callable`);
+        }
+        try {
+          return await callOn(attemptExec);
+        } catch (err) {
+          // Never retry on a Stop or an exhausted cap.
+          if (stopped()) throw err;
+          // Only a genuine transport break is a recovery candidate.
+          if (!isRetryableConnectError(err)) throw err;
+          // A write tool is INDETERMINATE on a transport error (may have applied
+          // before the reset) — surface that; do NOT auto-retry (double-apply is
+          // the #435 incident class).
+          if (!isReadOnlyWriteClass(writeClass)) {
+            throw new Error(
+              `external MCP tool "${rawName}" hit a transport error and MAY have already ` +
+                `applied on the server — not retried automatically; verify state before ` +
+                `retrying. (${shortError(err)})`,
+            );
+          }
+          // Abort check BEFORE minting a fresh connection (no socket for a
+          // stopped run). LIMITATION (#489, LOW): the reconnect's own connect is
+          // bounded by CONNECT_TIMEOUT_MS but does NOT itself observe `composed`,
+          // so a Stop that lands DURING the handshake is only honored at the next
+          // `stopped()` gate (before the retry) — a bounded ≤5s late-abort window;
+          // the throwaway client is closed at turn-end regardless. Threading
+          // `composed` into the SHARED (CAS-deduped) reconnect is deliberately
+          // avoided: it would let the first caller's abort tear down a reconnect a
+          // concurrent still-live caller depends on.
+          if (stopped()) throw err;
+          // CAS-swap by IDENTITY: mint+swap only if nobody swapped since this
+          // call's snapshot; a losing concurrent call awaits the same reconnect
+          // and retries on the SAME fresh client.
+          let target: RecoveredServerState;
+          if (boundBinding.current === attemptState) {
+            if (!boundBinding.reconnecting) {
+              boundBinding.reconnecting = (async () => {
+                const server = entry.servers[serverIndex];
+                const fresh = await service.reconnectServer(server, capMs);
+                leaseSet.push(fresh.lease); // accumulate; released at turn-end
+                boundBinding.current = fresh.state;
+                return fresh.state;
+              })();
+              // Clear the in-flight marker once it settles (success or failure) so
+              // a LATER death of the new client can reconnect again.
+              void boundBinding.reconnecting.then(
+                () => (boundBinding.reconnecting = undefined),
+                () => (boundBinding.reconnecting = undefined),
+              );
+            }
+            target = await boundBinding.reconnecting;
+          } else {
+            target = boundBinding.current as RecoveredServerState;
+          }
+          // Abort check BEFORE the retry.
+          if (stopped()) throw err;
+          const retryExec = execFor(target);
+          if (typeof retryExec !== 'function') throw err;
+          return await callOn(retryExec);
+        }
+      } finally {
+        clearTimeout(capTimer);
+      }
+    };
+    return { ...template, execute } as unknown as Tool;
+  }
+
+  /**
+   * Reconnect ONE server for an in-run recovery (#489): open a fresh client and
+   * list+wrap its tools. The throwaway client is NOT cached — it is owned by the
+   * RUN via the returned lease (closed at turn-end), independent of the shared
+   * cache entry (whose TTL rebuild heals future turns). On a failure the fresh
+   * client is closed so its socket never leaks.
+   */
+  private async reconnectServer(
+    server: AiMcpServer,
+    capMs: number,
+  ): Promise<{ state: RecoveredServerState; lease: Closable }> {
+    const client = await this.connectWithTimeout(server, CONNECT_TIMEOUT_MS);
+    let tools: Record<string, Tool>;
+    try {
+      const raw = await withTimeout(client.tools(), CONNECT_TIMEOUT_MS);
+      const allow = server.toolAllowlist;
+      const picked =
+        Array.isArray(allow) && allow.length > 0 ? pick(raw, allow) : raw;
+      tools = wrapToolsWithCallTimeout(picked, capMs);
+    } catch (err) {
+      void client.close().catch(() => undefined);
+      throw err;
+    }
+    let released = false;
+    const lease: Closable = {
+      close: async () => {
+        if (released) return;
+        released = true;
+        await client.close().catch(() => undefined);
+      },
+    };
+    return { state: { client, tools }, lease };
+  }
+
   /** Mark an entry evicted; close its clients now if nothing is leasing them. */
   private evict(entry: CacheEntry): void {
     clearTimeout(entry.timer);
@@ -554,22 +915,21 @@ export function validateResolvedAddresses(addrs: readonly LookupAddress[]): {
  * certificate validation still uses the real hostname (we never rewrite the URL
  * to an IP literal).
  */
-function buildPinnedDispatcher(): Agent {
-  // External-MCP traffic uses a DEDICATED, shorter silence timeout
+function buildPinnedDispatcher(bodyTimeoutMs: number): Agent {
+  // External-MCP traffic uses a DEDICATED, shorter HEADERS silence timeout
   // (`AI_MCP_STREAM_TIMEOUT_MS`, default 1 min) — deliberately tighter than the
   // chat provider's 15-min `streamTimeoutMs()` — so a byte-silent/hung MCP
   // upstream is broken in ~1 min instead of 15. We keep the keep-alive options
-  // from `streamingDispatcherOptions()` but OVERRIDE headers/body timeouts.
-  // Accepted trade-off: a legitimately long but byte-silent single tool call,
-  // and an SSE transport idling >1 min BETWEEN tool calls, are also cut here; the
-  // per-call total cap (wrapToolsWithCallTimeout, `AI_MCP_CALL_TIMEOUT_MS`) is the
-  // complementary guard for chatty-but-stuck calls that keep the socket warm yet
-  // never return.
-  const mcpSilenceMs = mcpStreamTimeoutMs();
+  // from `streamingDispatcherOptions()` but OVERRIDE the timeouts. `bodyTimeout`
+  // is passed in per-transport (#489): tight for HTTP (fresh request per call),
+  // raised for SSE (one long-lived body across calls — idle BETWEEN calls is
+  // legit). The per-call total cap (`AI_MCP_CALL_TIMEOUT_MS`) is the complementary
+  // guard for chatty-but-stuck calls that keep the socket warm yet never return.
+  const headersMs = mcpStreamTimeoutMs();
   return new Agent({
     ...streamingDispatcherOptions(),
-    headersTimeout: mcpSilenceMs,
-    bodyTimeout: mcpSilenceMs,
+    headersTimeout: headersMs,
+    bodyTimeout: bodyTimeoutMs,
     connect: {
       lookup: (hostname, _options, callback) => {
         // Always resolve ALL addresses ourselves; do not trust the caller's
@@ -669,18 +1029,22 @@ function pick(
 function namespace(
   tools: Record<string, Tool>,
   serverName: string,
-): Record<string, Tool> {
+): Array<{ full: string; raw: string; tool: Tool }> {
   const prefix = namespacePrefix(serverName);
-  const out: Record<string, Tool> = {};
+  const out: Array<{ full: string; raw: string; tool: Tool }> = [];
+  const taken: Record<string, true> = {};
   for (const [name, t] of Object.entries(tools)) {
     const safe = sanitizeName(name);
     let full = capName(`${prefix}_${safe}`);
     // Duplicate names within ONE server can still collide after sanitize/
     // truncate — suffix-disambiguate so the second tool is not overwritten.
-    if (full in out) {
-      full = disambiguate(full, '', (candidate) => candidate in out);
+    if (full in taken) {
+      full = disambiguate(full, '', (candidate) => candidate in taken);
     }
-    out[full] = t;
+    taken[full] = true;
+    // Keep the RAW (un-namespaced) name alongside the merged key so the per-run
+    // recovery wrapper (#489) can re-resolve the same tool on a fresh client.
+    out.push({ full, raw: name, tool: t });
   }
   return out;
 }
@@ -804,6 +1168,69 @@ export function wrapToolWithCallTimeout(tool: Tool, ms: number): Tool {
   return { ...tool, execute } as unknown as Tool;
 }
 
+/**
+ * undici / Node network error CODES that mean the connection broke (not an
+ * application-level error) — a transient transport failure a readOnly call may
+ * safely retry after reconnecting. Matched against the REAL error shapes (#489):
+ * a socket reset surfaces as `TypeError: fetch failed` whose `.cause` is an
+ * undici `SocketError { code:'UND_ERR_SOCKET' }`; a body-timeout as
+ * `TypeError: terminated` whose `.cause` is `BodyTimeoutError`. Classifying by
+ * these real codes/names (not by mock errors) is essential — a mock-shaped
+ * predicate would leave eviction silently dead in production while CI is green.
+ */
+const RETRYABLE_TRANSPORT_ERROR_CODES: ReadonlySet<string> = new Set([
+  'ECONNRESET',
+  'ECONNREFUSED',
+  'ECONNABORTED',
+  'EPIPE',
+  'ETIMEDOUT',
+  'EAI_AGAIN',
+  'ENETUNREACH',
+  'EHOSTUNREACH',
+  'UND_ERR_SOCKET',
+  'UND_ERR_CONNECT_TIMEOUT',
+  'UND_ERR_HEADERS_TIMEOUT',
+  'UND_ERR_BODY_TIMEOUT',
+  'UND_ERR_CLOSED',
+  'UND_ERR_DESTROYED',
+]);
+
+/** undici error CLASS names for the same transport-break conditions. */
+const RETRYABLE_TRANSPORT_ERROR_NAMES: ReadonlySet<string> = new Set([
+  'SocketError',
+  'BodyTimeoutError',
+  'HeadersTimeoutError',
+  'ConnectTimeoutError',
+  'ClientClosedError',
+  'ClientDestroyedError',
+]);
+
+/**
+ * Whether `err` is a retryable TRANSPORT break (a broken socket / body timeout),
+ * classified by the REAL undici/@ai-sdk error shapes (#489). undici surfaces a
+ * reset as `TypeError('fetch failed'|'terminated')` with the real error in
+ * `.cause`, and @ai-sdk/mcp may wrap it again in an `MCPClientError` (cause
+ * chain), so we walk `.cause` (bounded depth) checking `.code` and `.name`. An
+ * app-level tool error (a 4xx, a validation failure) is NOT retryable and returns
+ * false — only a connection-level failure heals with a reconnect.
+ */
+export function isRetryableConnectError(err: unknown, depth = 0): boolean {
+  if (!err || typeof err !== 'object' || depth > 6) return false;
+  const e = err as {
+    code?: unknown;
+    name?: unknown;
+    cause?: unknown;
+  };
+  if (typeof e.code === 'string' && RETRYABLE_TRANSPORT_ERROR_CODES.has(e.code)) {
+    return true;
+  }
+  if (typeof e.name === 'string' && RETRYABLE_TRANSPORT_ERROR_NAMES.has(e.name)) {
+    return true;
+  }
+  if (e.cause != null) return isRetryableConnectError(e.cause, depth + 1);
+  return false;
+}
+
 /** The signal's reason as an Error (informative thrown value on abort/timeout). */
 function abortReason(signal: AbortSignal): Error {
   const r = signal.reason;

apps/server/src/core/ai-chat/history-budget.spec.ts

@@ -0,0 +1,266 @@
+import type { ModelMessage } from 'ai';
+import {
+  resolveReplayBudget,
+  isContextOverflowError,
+  estimateMessagesTokens,
+  trimHistoryForReplay,
+  REPLAY_BUDGET_DEFAULT_TOKENS,
+  REPLAY_TRUNCATION_MARKER,
+  REPLAY_TURN_COLLAPSED_MARKER,
+} from './history-budget';
+
+describe('resolveReplayBudget', () => {
+  it('uses min(default, 0.7 x window) for a configured window', () => {
+    // 0.7 x 60k = 42k < 100k
+    expect(resolveReplayBudget(60_000)).toEqual({
+      thresholdTokens: 42_000,
+      usedDefault: false,
+    });
+    // 0.7 x 1M = 700k, capped to the 100k default
+    expect(resolveReplayBudget(1_000_000)).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: false,
+    });
+  });
+
+  it('accepts the raw ::text stored form', () => {
+    expect(resolveReplayBudget('60000').thresholdTokens).toBe(42_000);
+  });
+
+  // The crux (#490): a chat with NO context window configured must STILL be
+  // budgeted — those are exactly the installs that hit terminal overflow.
+  it('applies the flat default when the window is unset/empty', () => {
+    expect(resolveReplayBudget(undefined)).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+    expect(resolveReplayBudget('')).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+    expect(resolveReplayBudget('   ')).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+  });
+
+  it('treats an explicit 0 as the off-switch (distinct from unset)', () => {
+    expect(resolveReplayBudget(0)).toEqual({
+      thresholdTokens: null,
+      usedDefault: false,
+    });
+    expect(resolveReplayBudget('0')).toEqual({
+      thresholdTokens: null,
+      usedDefault: false,
+    });
+  });
+
+  it('falls back to the default on a negative/garbage value', () => {
+    expect(resolveReplayBudget(-5).usedDefault).toBe(true);
+    expect(resolveReplayBudget('abc').usedDefault).toBe(true);
+  });
+});
+
+describe('isContextOverflowError', () => {
+  it('classifies a real provider 400 context-overflow shape', () => {
+    // OpenAI-compatible shape.
+    expect(
+      isContextOverflowError({
+        statusCode: 400,
+        message:
+          "This model's maximum context length is 128000 tokens. However, your messages resulted in 214000 tokens. Please reduce the length of the messages.",
+      }),
+    ).toBe(true);
+    // Anthropic-style wording.
+    expect(
+      isContextOverflowError({
+        status: 400,
+        message: 'prompt is too long: 250000 tokens > 200000 maximum',
+      }),
+    ).toBe(true);
+    // Nested body + string status.
+    expect(
+      isContextOverflowError({
+        response: { status: '400' },
+        message: 'input is too long for the requested model',
+      }),
+    ).toBe(true);
+    // Error instance with the cause carrying the body.
+    const e = new Error('Bad request');
+    (e as any).statusCode = 400;
+    (e as any).cause = new Error('maximum context window exceeded');
+    expect(isContextOverflowError(e)).toBe(true);
+  });
+
+  it('does NOT classify unrelated 400s or auth/rate-limit errors', () => {
+    expect(
+      isContextOverflowError({ statusCode: 400, message: 'invalid tool schema' }),
+    ).toBe(false);
+    expect(
+      isContextOverflowError({
+        statusCode: 429,
+        message: 'context length exceeded but rate limited',
+      }),
+    ).toBe(false);
+    expect(isContextOverflowError({ statusCode: 500, message: 'server error' })).toBe(
+      false,
+    );
+    expect(isContextOverflowError(undefined)).toBe(false);
+    expect(isContextOverflowError('some random string')).toBe(false);
+  });
+});
+
+// Helpers to build ModelMessage fixtures in the ai@6 shape.
+const userMsg = (text: string): ModelMessage =>
+  ({ role: 'user', content: [{ type: 'text', text }] }) as ModelMessage;
+const assistantMsg = (
+  text: string,
+  toolCallId?: string,
+  toolName?: string,
+): ModelMessage =>
+  ({
+    role: 'assistant',
+    content: [
+      { type: 'text', text },
+      ...(toolCallId
+        ? [{ type: 'tool-call', toolCallId, toolName, input: {} }]
+        : []),
+    ],
+  }) as ModelMessage;
+const toolMsg = (
+  toolCallId: string,
+  toolName: string,
+  value: unknown,
+): ModelMessage =>
+  ({
+    role: 'tool',
+    content: [
+      { type: 'tool-result', toolCallId, toolName, output: { type: 'json', value } },
+    ],
+  }) as ModelMessage;
+
+describe('trimHistoryForReplay', () => {
+  it('null budget disables trimming (returns the same reference)', () => {
+    const msgs = [userMsg('hi'), assistantMsg('yo')];
+    const r = trimHistoryForReplay(msgs, null);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+
+  it('leaves history under budget untouched (same reference)', () => {
+    const msgs = [userMsg('hi'), assistantMsg('a short answer')];
+    const r = trimHistoryForReplay(msgs, 100_000);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+
+  it('truncates OLD tool outputs but keeps recent turns full', () => {
+    const big = 'X'.repeat(40_000); // ~16k tokens on its own
+    const msgs: ModelMessage[] = [];
+    // 6 OLD turns (indices 0..5), each with a huge tool output.
+    for (let i = 0; i < 6; i++) {
+      msgs.push(userMsg(`old q${i}`));
+      msgs.push(assistantMsg('looking', `c${i}`, 'getPage'));
+      msgs.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+      msgs.push(assistantMsg(`old a${i}`));
+    }
+    // 3 small recent turns, then the CURRENT turn with its own huge tool output.
+    // With REPLAY_KEEP_RECENT_TURNS=4 the last 4 user-turns stay full, so only
+    // these small recent turns + the current big one are kept full; the 6 old
+    // turns above fall in the trim region.
+    for (let i = 0; i < 3; i++) {
+      msgs.push(userMsg(`recent q${i}`));
+      msgs.push(assistantMsg(`recent a${i}`));
+    }
+    msgs.push(userMsg('current q'));
+    msgs.push(assistantMsg('looking', 'cR', 'getPage'));
+    msgs.push(toolMsg('cR', 'getPage', { body: big }));
+    msgs.push(assistantMsg('current a'));
+
+    // Budget large enough that phase-1 tool truncation alone brings it under.
+    const r = trimHistoryForReplay(msgs, 30_000);
+    expect(r.trimmed).toBe(true);
+    const flat = JSON.stringify(r.messages);
+    // The CURRENT turn's tool output survives in full.
+    expect(flat).toContain(big);
+    // Old outputs were truncated with the marker.
+    expect(flat).toContain(REPLAY_TRUNCATION_MARKER);
+    // Phase 1 sufficed: the oldest turns were NOT collapsed.
+    expect(flat).not.toContain(REPLAY_TURN_COLLAPSED_MARKER);
+    expect(estimateMessagesTokens(r.messages)).toBeLessThan(
+      estimateMessagesTokens(msgs),
+    );
+  });
+
+  it('collapses the oldest turns when tool truncation is not enough', () => {
+    // Many turns with LARGE assistant TEXT (not tool output) so phase 1 can't help.
+    const bigText = 'слово '.repeat(8_000); // large Cyrillic text per turn
+    const msgs: ModelMessage[] = [];
+    for (let i = 0; i < 12; i++) {
+      msgs.push(userMsg(`q${i}`));
+      msgs.push(assistantMsg(bigText));
+    }
+    const r = trimHistoryForReplay(msgs, 30_000);
+    expect(r.trimmed).toBe(true);
+    // Oldest turns collapsed; result fits (best-effort) and is much smaller.
+    expect(estimateMessagesTokens(r.messages)).toBeLessThan(
+      estimateMessagesTokens(msgs),
+    );
+    // The LAST turn's text is preserved in full (recent turns stay full).
+    expect(JSON.stringify(r.messages[r.messages.length - 1])).toContain(bigText);
+  });
+
+  it('is deterministic / byte-stable for identical inputs', () => {
+    const big = 'Y'.repeat(30_000);
+    const build = (): ModelMessage[] => {
+      const m: ModelMessage[] = [];
+      for (let i = 0; i < 10; i++) {
+        m.push(userMsg(`q${i}`));
+        m.push(assistantMsg('t', `c${i}`, 'getPage'));
+        m.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+      }
+      return m;
+    };
+    const a = trimHistoryForReplay(build(), 15_000);
+    const b = trimHistoryForReplay(build(), 15_000);
+    expect(JSON.stringify(a.messages)).toBe(JSON.stringify(b.messages));
+  });
+
+  it('never leaves an unpaired tool-call after collapsing (balanced history)', () => {
+    const big = 'Z'.repeat(40_000);
+    const msgs: ModelMessage[] = [];
+    for (let i = 0; i < 10; i++) {
+      msgs.push(userMsg(`q${i}`));
+      msgs.push(assistantMsg('t', `c${i}`, 'getPage'));
+      msgs.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+    }
+    const r = trimHistoryForReplay(msgs, 8_000);
+    // Count tool-call vs tool-result parts in the trimmed output.
+    let calls = 0;
+    let results = 0;
+    for (const m of r.messages) {
+      if (!Array.isArray(m.content)) continue;
+      for (const p of m.content as Array<{ type?: string }>) {
+        if (p.type === 'tool-call') calls++;
+        if (p.type === 'tool-result' || p.type === 'tool-error') results++;
+      }
+    }
+    // Every surviving tool-call has a surviving result (collapsing drops BOTH).
+    expect(calls).toBe(results);
+    // Collapsed turns carry the marker.
+    expect(JSON.stringify(r.messages)).toContain(REPLAY_TURN_COLLAPSED_MARKER);
+  });
+
+  it('respects the provider fact: under-budget contextTokens skips trimming', () => {
+    const big = 'W'.repeat(60_000);
+    const msgs = [
+      userMsg('q'),
+      assistantMsg('t', 'c1', 'getPage'),
+      toolMsg('c1', 'getPage', { body: big }),
+    ];
+    // char-estimate is high, but the provider says we are well under budget.
+    const r = trimHistoryForReplay(msgs, 100_000, 5_000);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+});

apps/server/src/core/ai-chat/history-budget.ts

@@ -0,0 +1,356 @@
+/**
+ * History-replay token budget (#490).
+ *
+ * The whole persisted conversation is replayed to the provider on EVERY turn, so
+ * a long chat eventually exceeds the model's context window and the provider 400s
+ * on every turn — terminally (the chat "bricks"). This module bounds the replayed
+ * history at REPLAY TIME only: it never mutates what is persisted (the DB stays
+ * the full record), and its output is a deterministic, byte-stable function of its
+ * input so the trimmed prefix is identical turn to turn (provider prompt-cache
+ * friendliness — real money on long chats).
+ *
+ * The PRIMARY signal is the provider's own fact: `metadata.contextTokens` from the
+ * last turn. The chars-based {@link estimateTokens} (shared with the client) is
+ * used only for the DELTA of not-yet-sent messages, to decide WHAT to trim, and as
+ * the fallback for chats with no usage yet.
+ */
+import type { ModelMessage } from 'ai';
+import { estimateTokens } from '@docmost/token-estimate';
+
+/** Flat default budget when no context window is configured (tokens). */
+export const REPLAY_BUDGET_DEFAULT_TOKENS = 100_000;
+/** Fraction of a configured context window used as the budget. */
+export const REPLAY_BUDGET_WINDOW_FRACTION = 0.7;
+/**
+ * Fraction of the normal budget used for the REACTIVE re-trim after a provider
+ * context-overflow 400 — the preventive estimate under-counted, so cut harder.
+ */
+export const REPLAY_AGGRESSIVE_FRACTION = 0.5;
+/**
+ * Turns (a user message + its assistant/tool replies) kept FULL at the tail,
+ * including the current one — never trimmed. Older turns are compacted first.
+ */
+export const REPLAY_KEEP_RECENT_TURNS = 4;
+/** Leading chars kept from a truncated old tool output. */
+export const REPLAY_TOOL_OUTPUT_HEAD = 800;
+/** Trailing chars kept from a truncated old tool output. */
+export const REPLAY_TOOL_OUTPUT_TAIL = 300;
+/** Marker inserted where an old tool output was truncated for replay. */
+export const REPLAY_TRUNCATION_MARKER =
+  '[…truncated for replay; call the tool again to read the full output]';
+/** Marker for a whole old turn collapsed to its text. */
+export const REPLAY_TURN_COLLAPSED_MARKER =
+  '[earlier tool activity omitted for replay]';
+
+export interface ReplayBudget {
+  /** Token threshold above which replay history is trimmed; `null` = OFF. */
+  thresholdTokens: number | null;
+  /** True when the flat default was used (no context window configured). */
+  usedDefault: boolean;
+}
+
+/**
+ * Resolve the replay budget from the RAW stored `chatContextWindow` (text/number).
+ *  - a positive value  -> `min(default, floor(fraction × window))`
+ *  - explicit `0`       -> OFF (admin opt-out; `null` threshold)
+ *  - unset/empty/invalid-> the flat default (still protects — the installations
+ *    that hit terminal overflow are exactly the ones that never set a window)
+ *
+ * Note the raw value is needed because the parsed `chatContextWindow` collapses
+ * both `0` and unset to `undefined`, which would erase the explicit off-switch.
+ */
+export function resolveReplayBudget(rawContextWindow: unknown): ReplayBudget {
+  let n: number | undefined;
+  if (typeof rawContextWindow === 'number') {
+    n = rawContextWindow;
+  } else if (typeof rawContextWindow === 'string') {
+    const t = rawContextWindow.trim();
+    n = t === '' ? undefined : Number(t);
+  }
+  // Unset / empty / non-numeric / negative -> flat default (the protective case).
+  if (n === undefined || !Number.isFinite(n) || n < 0) {
+    return { thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS, usedDefault: true };
+  }
+  // Explicit 0 -> off-switch.
+  if (n === 0) {
+    return { thresholdTokens: null, usedDefault: false };
+  }
+  return {
+    thresholdTokens: Math.min(
+      REPLAY_BUDGET_DEFAULT_TOKENS,
+      Math.floor(REPLAY_BUDGET_WINDOW_FRACTION * n),
+    ),
+    usedDefault: false,
+  };
+}
+
+/**
+ * True when a provider error is a CONTEXT-OVERFLOW rejection (the prompt exceeds
+ * the model's window). Providers surface this as an HTTP 400 with a recognizable
+ * message; match both the status and the message patterns robustly across
+ * OpenAI-compatible / Anthropic / Gemini wordings, since the exact shape varies.
+ */
+export function isContextOverflowError(error: unknown): boolean {
+  const status = extractStatus(error);
+  const msg = extractMessage(error).toLowerCase();
+  // Message patterns seen across providers for "prompt too long".
+  const overflowPattern =
+    /context (?:length|window)|maximum context|too many tokens|too large for|reduce the length|prompt is too long|input (?:is )?too long|exceeds? the (?:maximum )?(?:context|token)|maximum.*tokens|string too long/;
+  if (!overflowPattern.test(msg)) return false;
+  // A 400/413 with an overflow-shaped message is an overflow. Some providers
+  // omit/rewrite the status, so accept the message match when the status is
+  // unknown, but reject it for auth/rate-limit statuses that never mean overflow.
+  if (status === 400 || status === 413) return true;
+  if (status === 401 || status === 403 || status === 429) return false;
+  return true;
+}
+
+function extractStatus(error: unknown): number | undefined {
+  if (!error || typeof error !== 'object') return undefined;
+  const e = error as Record<string, unknown>;
+  for (const k of ['statusCode', 'status']) {
+    const v = e[k];
+    if (typeof v === 'number') return v;
+    if (typeof v === 'string' && /^\d+$/.test(v)) return Number(v);
+  }
+  // Nested (e.g. { response: { status } } / { cause: { statusCode } }).
+  for (const k of ['response', 'cause', 'data']) {
+    const nested = e[k];
+    if (nested && typeof nested === 'object') {
+      const s = extractStatus(nested);
+      if (s !== undefined) return s;
+    }
+  }
+  return undefined;
+}
+
+function extractMessage(error: unknown): string {
+  if (error == null) return '';
+  if (typeof error === 'string') return error;
+  if (error instanceof Error) {
+    // Include nested causes (provider libs wrap the real body in `cause`).
+    const cause = (error as { cause?: unknown }).cause;
+    return `${error.message} ${cause ? extractMessage(cause) : ''}`;
+  }
+  if (typeof error === 'object') {
+    const e = error as Record<string, unknown>;
+    const parts: string[] = [];
+    for (const k of ['message', 'error', 'body', 'responseBody', 'data']) {
+      const v = e[k];
+      if (typeof v === 'string') parts.push(v);
+      else if (v && typeof v === 'object') parts.push(extractMessage(v));
+    }
+    return parts.join(' ');
+  }
+  return String(error);
+}
+
+/** Rough token size of a ModelMessage array via the shared chars estimator. */
+export function estimateMessagesTokens(
+  messages: ReadonlyArray<ModelMessage>,
+): number {
+  let total = 0;
+  for (const m of messages) {
+    total += estimateTokens(serializeContent(m.content));
+  }
+  return total;
+}
+
+function serializeContent(content: unknown): string {
+  if (typeof content === 'string') return content;
+  try {
+    return JSON.stringify(content) ?? '';
+  } catch {
+    return '';
+  }
+}
+
+/** Deep JSON string of an arbitrary value, bounded so estimation never throws. */
+function stringifyValue(value: unknown): string {
+  if (typeof value === 'string') return value;
+  try {
+    return JSON.stringify(value) ?? String(value);
+  } catch {
+    return String(value);
+  }
+}
+
+export interface TrimResult {
+  messages: ModelMessage[];
+  /** Whether any trimming was applied. */
+  trimmed: boolean;
+  /** Estimated tokens of the returned messages (chars-based). */
+  estimatedTokens: number;
+}
+
+/**
+ * Bound the replayed history to `budgetTokens`, deterministically. Returns the
+ * SAME array reference (no copy) when nothing needs trimming, so the common case
+ * is free and byte-identical. Trimming order (spec #490):
+ *   1. truncate OLD turns' tool outputs (head+tail + marker) — the bulk of the size
+ *   2. mechanically collapse the OLDEST turns to their text (concatenation, no LLM)
+ *   3. the current + last {@link REPLAY_KEEP_RECENT_TURNS} turns stay FULL
+ *
+ * `budgetTokens === null` disables trimming. `priorContextTokens` (the provider's
+ * fact from last turn) short-circuits the decision: when it is known and already
+ * under budget we skip trimming even if the char-estimate is higher (the provider
+ * count is authoritative). The char-estimate drives WHAT to cut.
+ */
+export function trimHistoryForReplay(
+  messages: ModelMessage[],
+  budgetTokens: number | null,
+  priorContextTokens?: number,
+): TrimResult {
+  if (budgetTokens == null) {
+    return { messages, trimmed: false, estimatedTokens: 0 };
+  }
+  const estimated = estimateMessagesTokens(messages);
+  // Decision signal: prefer the provider's fact (last turn's contextTokens) plus
+  // the estimated delta of the messages appended since; fall back to the pure
+  // char-estimate for a chat with no usage yet.
+  const projected =
+    priorContextTokens != null
+      ? Math.max(priorContextTokens, estimated)
+      : estimated;
+  if (projected <= budgetTokens) {
+    return { messages, trimmed: false, estimatedTokens: estimated };
+  }
+
+  // The tail we always keep full: from the Nth-from-last user message onward.
+  const boundary = recentBoundaryIndex(messages, REPLAY_KEEP_RECENT_TURNS);
+  const tail = messages.slice(boundary);
+  let head = messages.slice(0, boundary).map(cloneMessage);
+
+  // Phase 1: truncate old tool outputs.
+  for (const m of head) {
+    if (m.role === 'tool') truncateToolMessage(m);
+  }
+  let out = [...head, ...tail];
+  let est = estimateMessagesTokens(out);
+  if (est <= budgetTokens) {
+    return { messages: out, trimmed: true, estimatedTokens: est };
+  }
+
+  // Phase 2: collapse the oldest turns (in `head`) to their text, one at a time,
+  // from the oldest, until we fit or the whole head is collapsed.
+  const turns = splitTurns(head);
+  const collapsed: ModelMessage[] = [];
+  let i = 0;
+  for (; i < turns.length; i++) {
+    if (est <= budgetTokens) break;
+    collapsed.push(...collapseTurn(turns[i]));
+    // Re-estimate the whole prospective output.
+    const remaining = turns.slice(i + 1).flat();
+    out = [...collapsed, ...remaining, ...tail];
+    est = estimateMessagesTokens(out);
+  }
+  // Include any turns we didn't need to collapse.
+  const remaining = turns.slice(i).flat();
+  out = [...collapsed, ...remaining, ...tail];
+  est = estimateMessagesTokens(out);
+  return { messages: out, trimmed: true, estimatedTokens: est };
+}
+
+/** Index of the first message of the Nth-from-last user turn (0 if fewer). */
+function recentBoundaryIndex(
+  messages: ReadonlyArray<ModelMessage>,
+  keepTurns: number,
+): number {
+  const userIdx: number[] = [];
+  for (let i = 0; i < messages.length; i++) {
+    if (messages[i].role === 'user') userIdx.push(i);
+  }
+  if (userIdx.length <= keepTurns) return 0;
+  return userIdx[userIdx.length - keepTurns];
+}
+
+/** Split a message list into turns; each turn starts at a `user` message. */
+function splitTurns(messages: ModelMessage[]): ModelMessage[][] {
+  const turns: ModelMessage[][] = [];
+  for (const m of messages) {
+    if (m.role === 'user' || turns.length === 0) turns.push([m]);
+    else turns[turns.length - 1].push(m);
+  }
+  return turns;
+}
+
+/**
+ * Collapse a whole turn to its plain text (mechanical concatenation, not an LLM
+ * summary). Keeps the user message; replaces the assistant/tool messages with a
+ * single assistant text message = the assistant's concatenated text + a marker
+ * when tool activity was dropped. Dropping BOTH the tool-call and tool-result
+ * parts together keeps the rebuilt history balanced (no unpaired calls).
+ */
+function collapseTurn(turn: ModelMessage[]): ModelMessage[] {
+  const out: ModelMessage[] = [];
+  let assistantText = '';
+  let hadTools = false;
+  for (const m of turn) {
+    if (m.role === 'user') {
+      out.push(m);
+    } else if (m.role === 'assistant') {
+      const { text, tools } = extractAssistantText(m.content);
+      assistantText += text;
+      hadTools = hadTools || tools;
+    } else if (m.role === 'tool') {
+      hadTools = true;
+    } else {
+      out.push(m);
+    }
+  }
+  const note =
+    (assistantText ? assistantText.trimEnd() : '') +
+    (hadTools
+      ? `${assistantText ? '\n\n' : ''}${REPLAY_TURN_COLLAPSED_MARKER}`
+      : '');
+  if (note) out.push({ role: 'assistant', content: note } as ModelMessage);
+  return out;
+}
+
+function extractAssistantText(content: unknown): {
+  text: string;
+  tools: boolean;
+} {
+  if (typeof content === 'string') return { text: content, tools: false };
+  if (!Array.isArray(content)) return { text: '', tools: false };
+  let text = '';
+  let tools = false;
+  for (const part of content) {
+    const type = (part as { type?: string })?.type;
+    if (type === 'text') text += (part as { text?: string }).text ?? '';
+    else if (type === 'tool-call') tools = true;
+  }
+  return { text, tools };
+}
+
+/** Truncate every tool-result output in a `tool` message to head+tail+marker. */
+function truncateToolMessage(message: ModelMessage): void {
+  const content = message.content;
+  if (!Array.isArray(content)) return;
+  for (const part of content) {
+    const p = part as { type?: string; output?: { type?: string; value?: unknown } };
+    if (p.type !== 'tool-result' && p.type !== 'tool-error') continue;
+    if (!p.output) continue;
+    const raw = stringifyValue(p.output.value);
+    const budget = REPLAY_TOOL_OUTPUT_HEAD + REPLAY_TOOL_OUTPUT_TAIL;
+    if (raw.length <= budget + REPLAY_TRUNCATION_MARKER.length) continue;
+    const truncated =
+      raw.slice(0, REPLAY_TOOL_OUTPUT_HEAD) +
+      `\n${REPLAY_TRUNCATION_MARKER}\n` +
+      raw.slice(raw.length - REPLAY_TOOL_OUTPUT_TAIL);
+    // Represent the shrunk output as a text output (a valid tool-result output).
+    p.output = { type: 'text', value: truncated };
+  }
+}
+
+/** Shallow-ish clone so trimming never mutates the caller's (persisted-derived)
+ *  message objects — only the OLD region is cloned before it is edited. */
+function cloneMessage(m: ModelMessage): ModelMessage {
+  if (typeof m.content === 'string') return { ...m };
+  return {
+    ...m,
+    content: (m.content as unknown[]).map((p) =>
+      p && typeof p === 'object' ? { ...(p as object) } : p,
+    ),
+  } as ModelMessage;
+}

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
-          // unified between the log line and the streamed error message — a
-          // share reader sees 402/429/503 causes consistently with the
-          // authenticated path.
-          return describeProviderError(error, 'AI stream error');
+          // SECURITY (#394): the string this returns is written verbatim into the
+          // SSE error frame delivered to an ANONYMOUS reader (for a tool failure
+          // it becomes the atomic `tool-output-error` frame's errorText; for a
+          // stream/provider failure, the terminal error frame).
+          //
+          // 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/ai-chat-tools.cap.spec.ts

@@ -0,0 +1,160 @@
+import {
+  wrapInAppToolWithCap,
+  inAppToolCallCapMs,
+  type ToolAbortSignalSink,
+} from './ai-chat-tools.service';
+import type { Tool, ToolCallOptions } from 'ai';
+
+/**
+ * #487 commit 1 — in-app tool race-on-abort + safe-points + per-call cap.
+ *
+ * Tests assert the HONEST observable property the spec names — "after Stop, NO
+ * new HTTP/WS call STARTS; an already-started single call may take either
+ * outcome" — against the REAL wrapper mechanism (the composite abort signal it
+ * publishes on the client + the RACE it runs), NOT a timing-dependent proxy like
+ * "the write didn't land".
+ */
+
+// A minimal stand-in for the client's `toolAbortSignal` field. In production the
+// wrapper publishes the composite here and the client's paginateAll /
+// mutatePageContent safe-points read it; the fake "tool" below reads it the same
+// way, so this exercises the real contract without a live DB / collab socket.
+class FakeClient implements ToolAbortSignalSink {
+  private signal: AbortSignal | null = null;
+  setToolAbortSignal(signal: AbortSignal | null): void {
+    this.signal = signal;
+  }
+  getToolAbortSignal(): AbortSignal | null {
+    return this.signal;
+  }
+}
+
+// A ToolCallOptions with just the field the wrapper reads (abortSignal). The AI
+// SDK passes a fuller object; the wrapper only spreads it and reads abortSignal.
+const opts = (abortSignal?: AbortSignal): ToolCallOptions =>
+  ({ toolCallId: 't1', messages: [], abortSignal }) as unknown as ToolCallOptions;
+
+const tick = (ms = 5) => new Promise((r) => setTimeout(r, ms));
+
+describe('#487 wrapInAppToolWithCap — race-on-abort + safe-points', () => {
+  it('after Stop, no NEW simulated call starts (multi-call tool)', async () => {
+    const client = new FakeClient();
+    const started: number[] = [];
+    // A multi-call tool that mirrors paginateAll: it consults the client signal
+    // at a safe-point BEFORE starting each simulated network call.
+    const multiCall: Tool = {
+      execute: (async (_args: unknown) => {
+        for (let i = 0; i < 6; i++) {
+          // Safe-point: exactly what paginateAll / mutatePageContent do.
+          client.getToolAbortSignal()?.throwIfAborted();
+          started.push(i);
+          await tick(10);
+        }
+        return 'done';
+      }) as unknown as Tool['execute'],
+    } as Tool;
+
+    const wrapped = wrapInAppToolWithCap(multiCall, client, 10_000);
+    const ac = new AbortController();
+    const call = (
+      wrapped.execute as (a: unknown, o: ToolCallOptions) => Promise<unknown>
+    )({}, opts(ac.signal));
+
+    // Let one or two calls start, then Stop.
+    await tick(12);
+    ac.abort(new Error('user stop'));
+
+    await expect(call).rejects.toThrow(); // wrapper rejects promptly
+    const startedAtStop = started.length;
+
+    // Give the abandoned loser ample time; its next safe-point must throw because
+    // the (aborted) composite is still published on the client.
+    await tick(60);
+    expect(started.length).toBe(startedAtStop);
+    // It must NOT have run the whole sequence (that would mean Stop did nothing).
+    expect(started.length).toBeLessThan(6);
+  });
+
+  it('rejects immediately on Stop even if the call never settles (discard loser)', async () => {
+    const client = new FakeClient();
+    let settled = false;
+    const hang: Tool = {
+      execute: (async () => {
+        await new Promise(() => undefined); // never resolves
+        settled = true;
+      }) as unknown as Tool['execute'],
+    } as Tool;
+    const wrapped = wrapInAppToolWithCap(hang, client, 10_000);
+    const ac = new AbortController();
+    const call = (
+      wrapped.execute as (a: unknown, o: ToolCallOptions) => Promise<unknown>
+    )({}, opts(ac.signal));
+    await tick(5);
+    ac.abort();
+    await expect(call).rejects.toThrow();
+    expect(settled).toBe(false);
+  });
+
+  it('per-call cap rejects a hung call with no Stop signal', async () => {
+    const client = new FakeClient();
+    const hang: Tool = {
+      execute: (async () => {
+        await new Promise(() => undefined);
+      }) as unknown as Tool['execute'],
+    } as Tool;
+    // Tiny cap; no options.abortSignal at all (composite = cap only).
+    const wrapped = wrapInAppToolWithCap(hang, client, 20);
+    const start = Date.now();
+    await expect(
+      (wrapped.execute as (a: unknown, o: ToolCallOptions) => Promise<unknown>)(
+        {},
+        opts(undefined),
+      ),
+    ).rejects.toThrow(/per-call cap/);
+    expect(Date.now() - start).toBeLessThan(2000);
+  });
+
+  it('publishes a composite signal on the client for the duration of the call', async () => {
+    const client = new FakeClient();
+    let seenDuringCall: AbortSignal | null = null;
+    const probe: Tool = {
+      execute: (async () => {
+        seenDuringCall = client.getToolAbortSignal();
+        return 'ok';
+      }) as unknown as Tool['execute'],
+    } as Tool;
+    const wrapped = wrapInAppToolWithCap(probe, client, 10_000);
+    const ac = new AbortController();
+    await (
+      wrapped.execute as (a: unknown, o: ToolCallOptions) => Promise<unknown>
+    )({}, opts(ac.signal));
+    expect(seenDuringCall).not.toBeNull();
+    // The published composite must reflect the turn's Stop signal.
+    ac.abort();
+    expect((seenDuringCall as unknown as AbortSignal).aborted).toBe(true);
+  });
+
+  it('a completed call returns its raw result unchanged', async () => {
+    const client = new FakeClient();
+    const ok: Tool = {
+      execute: (async () => ({ items: [1, 2, 3] })) as unknown as Tool['execute'],
+    } as Tool;
+    const wrapped = wrapInAppToolWithCap(ok, client, 10_000);
+    const res = await (
+      wrapped.execute as (a: unknown, o: ToolCallOptions) => Promise<unknown>
+    )({}, opts(new AbortController().signal));
+    expect(res).toEqual({ items: [1, 2, 3] });
+  });
+
+  it('cap is env-tunable with a 2-minute default', () => {
+    const prev = process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS;
+    delete process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS;
+    expect(inAppToolCallCapMs()).toBe(120_000);
+    process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS = '5000';
+    expect(inAppToolCallCapMs()).toBe(5000);
+    process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS = 'not-a-number';
+    expect(inAppToolCallCapMs()).toBe(120_000);
+    if (prev === undefined) delete process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS;
+    else process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS = prev;
+  });
+});

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

@@ -1,5 +1,5 @@
 import { Injectable, Logger } from '@nestjs/common';
-import { tool, type Tool } from 'ai';
+import { tool, type Tool, type ToolCallOptions } from 'ai';
 import { z } from 'zod';
 import { User } from '@docmost/db/types/entity.types';
 import { TokenService } from '../../auth/services/token.service';
@@ -159,6 +159,129 @@ function __assertClientCallContract(client: DocmostClientLike): void {
  * existing service-account `/mcp` path already calls loopback successfully, so
  * this works for single-workspace self-host.
  */
+/**
+ * #487: wall-clock cap for a SINGLE in-app tool call, env-tunable via
+ * `AI_CHAT_INAPP_TOOL_CALL_CAP_MS`. Bounds a read tool that would otherwise
+ * paginate for minutes and a content write whose collab commit hangs, and is the
+ * per-call CAP half of the composite abort signal every in-app tool is wrapped
+ * with (the other half is the turn's Stop signal). Default 2 minutes: generous
+ * for a legitimate long read/write, tight enough that a stuck call cannot pin the
+ * turn. The reconcile staleness floor (#487 commit 4) is derived as
+ * max(2 x this cap, 15 min), so keep this well under that.
+ */
+export function inAppToolCallCapMs(): number {
+  const raw = Number(process.env.AI_CHAT_INAPP_TOOL_CALL_CAP_MS);
+  return Number.isFinite(raw) && raw > 0 ? raw : 120_000;
+}
+
+/** #487: the composite signal's reason as an Error (informative thrown value). */
+function inAppAbortReason(signal: AbortSignal): Error {
+  const r = signal.reason;
+  return r instanceof Error
+    ? r
+    : new Error(typeof r === 'string' ? r : 'In-app tool call aborted');
+}
+
+/**
+ * The client surface {@link wrapInAppToolWithCap} drives (#487). Both methods are
+ * OPTIONAL: the real loopback DocmostClient implements them (so a Stop/cap reaches
+ * its pagination / pre-commit safe-points), but a client that omits them still
+ * gets the OUTER guarantee — the race rejects on abort regardless. This keeps the
+ * wrapper decoupled from the exact client shape (unit-test doubles need not stub
+ * the plumbing).
+ */
+export interface ToolAbortSignalSink {
+  setToolAbortSignal?(signal: AbortSignal | null): void;
+  getToolAbortSignal?(): AbortSignal | null;
+}
+
+/**
+ * #487: wrap an in-app tool so a Stop (the turn's `options.abortSignal`) OR the
+ * per-call wall-clock cap REJECTS the call immediately, and so that SAME
+ * composite signal reaches the client's pagination / pre-commit safe-points (via
+ * `client.setToolAbortSignal`) — making a Stop stop the NEXT HTTP/WS call from
+ * starting.
+ *
+ * Reuses the RACE pattern of `wrapToolWithCallTimeout` (mcp-clients.service.ts):
+ * the call is raced against the composite signal, so on abort we reject in the
+ * SAME tick and DISCARD the loser promise. Its network / collab teardown latency
+ * therefore never blocks the turn — the supersede timeout W=10s (#487 commit 3)
+ * relies on this abort->settle latency being milliseconds, not a socket teardown.
+ * Awaiting the client's own signal-into-write path alone would NOT satisfy this
+ * (the loser could still be tearing down a collab socket).
+ *
+ * The composite is SET on the client at entry and deliberately NOT restored on
+ * unwind: after this wrapper rejects on abort, the ABANDONED loser promise keeps
+ * running, and its safe-points read the client field — leaving the (aborted)
+ * composite there is exactly what makes the loser's NEXT call throw and stop. The
+ * next in-app tool call overwrites the field with its own fresh composite before
+ * any of its safe-points run, so a stale settled signal never leaks forward.
+ * SINGLE-WRITER by phase-1 assumption — see DocmostClientContext.toolAbortSignal
+ * for the parallel-call caveat (#487).
+ *
+ * KNOWN LIMITATION (#487): a write tool that issues SEVERAL sequential collab
+ * commits can be aborted BETWEEN commits, leaving a partially-applied operation.
+ * Cancel guarantees "no NEW call starts", NOT "the write didn't land".
+ */
+export function wrapInAppToolWithCap(
+  toolDef: Tool,
+  client: ToolAbortSignalSink,
+  capMs: number,
+): Tool {
+  const original = toolDef.execute;
+  if (typeof original !== 'function') return toolDef;
+  const execute = async (args: unknown, options: ToolCallOptions) => {
+    const capController = new AbortController();
+    const timer = setTimeout(() => {
+      capController.abort(
+        new Error(`In-app tool call exceeded the ${capMs}ms per-call cap`),
+      );
+    }, capMs);
+    timer.unref?.();
+    const composite = options?.abortSignal
+      ? AbortSignal.any([options.abortSignal, capController.signal])
+      : capController.signal;
+    // Reject the MOMENT the composite fires, independent of whether `original`
+    // ever settles (a hung collab write / read would otherwise pin the turn). The
+    // losing `original` is left pending; Promise.race attaches a rejection
+    // handler to both inputs so a late rejection is never unhandled.
+    const aborted = new Promise<never>((_, reject) => {
+      const fail = () => reject(inAppAbortReason(composite));
+      if (composite.aborted) fail();
+      else composite.addEventListener('abort', fail, { once: true });
+    });
+    // Publish the composite so the client's pagination / pre-commit safe-points
+    // observe it (see the "not restored on unwind" rationale above). Guarded: a
+    // client without the plumbing still gets the OUTER race guarantee below.
+    client.setToolAbortSignal?.(composite);
+    try {
+      return await Promise.race([
+        (original as (a: unknown, o: ToolCallOptions) => Promise<unknown>)(
+          args,
+          { ...options, abortSignal: composite },
+        ),
+        aborted,
+      ]);
+    } finally {
+      clearTimeout(timer);
+    }
+  };
+  return { ...toolDef, execute } as unknown as Tool;
+}
+
+/** #487: apply {@link wrapInAppToolWithCap} to every tool in a set. */
+export function wrapInAppToolsWithCap(
+  tools: Record<string, Tool>,
+  client: ToolAbortSignalSink,
+  capMs: number,
+): Record<string, Tool> {
+  const out: Record<string, Tool> = {};
+  for (const [name, t] of Object.entries(tools)) {
+    out[name] = wrapInAppToolWithCap(t, client, capMs);
+  }
+  return out;
+}
+
 @Injectable()
 export class AiChatToolsService {
   private readonly logger = new Logger(AiChatToolsService.name);
@@ -186,7 +309,12 @@ export class AiChatToolsService {
     sessionId: string,
     workspaceId: string,
     aiChatId: string,
-  ): Promise<DocmostClientLike> {
+    // #487: the returned client also carries the tool-cancellation plumbing
+    // (setToolAbortSignal/getToolAbortSignal). These are host plumbing, NOT part
+    // of the tool-execute surface (DocmostClientMethod), so they are surfaced here
+    // as an intersection rather than by widening that Pick — keeping the
+    // positional-call drift-guard (#446) scoped to the actual tool methods.
+  ): Promise<DocmostClientLike & ToolAbortSignalSink> {
     const apiUrl =
       process.env.MCP_DOCMOST_API_URL ||
       `http://127.0.0.1:${process.env.PORT || 3000}/api`;
@@ -630,7 +758,15 @@ export class AiChatToolsService {
     // dependency and reuses the CASL enforcement already on `client`. When the
     // loaded package predates #417 (factory undefined) or the loader is mocked in
     // a unit test, signalling is a pure no-op and results are byte-identical.
-    if (!createCommentSignalTracker) return tools;
+    // #487: wrap every in-app tool with the race-on-abort + per-call cap guard so
+    // a Stop / cap rejects immediately AND reaches the client's write/pagination
+    // safe-points. Applied as the OUTERMOST wrapper (over the comment-signal
+    // wrapper below) so the race governs the whole call. The client carries the
+    // per-call composite signal via setToolAbortSignal.
+    const capMs = inAppToolCallCapMs();
+    if (!createCommentSignalTracker) {
+      return wrapInAppToolsWithCap(tools, client, capMs);
+    }
 
     const tracker = createCommentSignalTracker({
       probe: async (pageId: string, sinceMs: number) => {
@@ -659,7 +795,11 @@ export class AiChatToolsService {
       },
     });
 
-    return wrapToolsWithCommentSignal(tools, tracker);
+    return wrapInAppToolsWithCap(
+      wrapToolsWithCommentSignal(tools, tracker),
+      client,
+      capMs,
+    );
   }
 }
 

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`
-// layout so `computeSrcRegistryStamp(<pkg>/build/index.js)` resolves src the same
-// way the loader does (dirname(dirname(entry))/src/tool-specs.ts).
-function makeFakePackage(toolSpecsSource: string | null): {
+// Build a throwaway `<pkg>/build/index.js` + optional `<pkg>/src/` tree so
+// `computeSrcRegistryStamp(<pkg>/build/index.js)` resolves src the same way the
+// loader does (dirname(dirname(entry))/src). Since #486 the stamp hashes the WHOLE
+// 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 });
-    writeFileSync(join(srcDir, 'tool-specs.ts'), toolSpecsSource, 'utf8');
+    for (const [rel, content] of Object.entries(files)) {
+      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 FIXED_INPUT = 'line1\r\nline2\n';
-    const EXPECTED =
-      '683376e290829b482c2655745caffa7a1dccfa10afaa62dac2b42dd6c68d0f83';
-    const { entry, cleanup } = makeFakePackage(FIXED_INPUT);
+  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';
+
+  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', () => {
-    // Proves EXPECTED is not a magic constant but the documented computation.
-    const FIXED_INPUT = 'line1\r\nline2\n';
-    const normalized = FIXED_INPUT.replace(/\r\n/g, '\n').replace(/\n$/, '');
-    const expected = createHash('sha256')
-      .update(normalized, 'utf8')
-      .digest('hex');
-    const { entry, cleanup } = makeFakePackage(FIXED_INPUT);
+  it('the documented EXPECTED is the enumerate+normalize+sha256 of the tree', () => {
+    // Proves EXPECTED is not a magic constant but the documented computation — a
+    // local re-implementation of the loader's tree walk.
+    const { entry, cleanup } = makeFakePackage(CROSS_IMPL_TREE);
     try {
-      expect(computeSrcRegistryStamp(entry)).toBe(expected);
+      const srcDir = join(dirname(dirname(entry)), 'src');
+      const collect = (dir: string): string[] => {
+        const out: string[] = [];
+        for (const e of readdirSync(dir)) {
+          const f = join(dir, e);
+          if (statSync(f).isDirectory()) out.push(...collect(f));
+          else if (e.endsWith('.ts') && !e.endsWith('.generated.ts'))
+            out.push(f);
+        }
+        return out;
+      };
+      const files = collect(srcDir)
+        .map((abs) => ({ rel: relative(srcDir, abs).split(sep).join('/'), abs }))
+        .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+      const h = createHash('sha256');
+      for (const { rel, abs } of files) {
+        const n = readFileSync(abs, 'utf8')
+          .replace(/\r\n/g, '\n')
+          .replace(/\n$/, '');
+        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 { dirname, join } from 'node:path';
+import { existsSync, readdirSync, readFileSync, statSync } from 'node:fs';
+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
- * normalization (CRLF -> LF, strip a single trailing newline), same sha256.
+ * build-time and src-time hashes agree: same file set (every src/**\/*.ts except
+ * *.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
- * worktree that file exists; in a prod image (build/ only, src/ stripped) it does
- * not, so this returns null and the caller skips the check. Any error (ENOENT, a
- * bad resolve) is swallowed to null — the stale-check must NEVER break startup.
+ * build/index.js) and look for ../src next to it. In a dev/test worktree that
+ * directory exists; in a prod image (build/ only, src/ stripped) it does not, so
+ * this returns null and the caller skips the check. Any error (ENOENT, a bad
+ * 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(
-      dirname(dirname(packageEntry)),
-      'src',
-      'tool-specs.ts',
-    );
-    if (!existsSync(toolSpecsPath)) return null; // prod: no src tree -> skip.
-    const source = readFileSync(toolSpecsPath, 'utf8');
-    const normalized = source.replace(/\r\n/g, '\n').replace(/\n$/, '');
-    return createHash('sha256').update(normalized, 'utf8').digest('hex');
+    const srcDir = join(dirname(dirname(packageEntry)), 'src');
+    if (!existsSync(srcDir)) return null; // prod: no src tree -> skip.
+    // Enumerate every src/**\/*.ts except the codegen's own *.generated.ts
+    // output (including it would be a fixed-point cycle). Sort by POSIX-relative
+    // path so ordering is platform-independent, then fold each file's relative
+    // path + normalized content into one hash — identical to the codegen.
+    const files = collectStampFiles(srcDir)
+      .map((abs) => ({
+        rel: relative(srcDir, abs).split(sep).join('/'),
+        abs,
+      }))
+      .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+    const hash = createHash('sha256');
+    for (const { rel, abs } of files) {
+      const normalized = readFileSync(abs, 'utf8')
+        .replace(/\r\n/g, '\n')
+        .replace(/\n$/, '');
+      hash.update(rel, 'utf8');
+      hash.update('\0', 'utf8');
+      hash.update(normalized, 'utf8');
+      hash.update('\0', 'utf8');
+    }
+    return hash.digest('hex');
   } 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;
-    let isApiKeyModuleReady = false;
+    const apiKeyService = this.resolveApiKeyService();
+    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');
-      isApiKeyModuleReady = true;
+      const ApiKeyModule = require('./../../../ee/api-key/api-key.service');
+      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/database/migrations/20260707T120000-ai-chat-metadata.ts

@@ -0,0 +1,24 @@
+import { type Kysely, sql } from 'kysely';
+
+export async function up(db: Kysely<any>): Promise<void> {
+  // Chat-level metadata bag (#490). First use: the deferred-tool ACTIVATION set
+  // (`activatedTools`) is persisted here so it survives across turns — previously
+  // the set was reset every turn, forcing the model to re-run loadTools and pay a
+  // fresh round-trip to re-activate the same tools each turn. On load the stored
+  // set is intersected with the current valid deferred names, so an allowlist /
+  // role change can never inject a now-nonexistent tool.
+  //
+  // jsonb, defaulted to '{}' so every row (incl. pre-migration ones, backfilled
+  // by the default) is a readable object — the app never has to null-guard the
+  // bag itself, only individual keys.
+  await db.schema
+    .alterTable('ai_chats')
+    .addColumn('metadata', 'jsonb', (col) =>
+      col.notNull().defaultTo(sql`'{}'::jsonb`),
+    )
+    .execute();
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema.alterTable('ai_chats').dropColumn('metadata').execute();
+}

apps/server/src/database/repos/ai-chat/ai-chat-message.repo.ts

@@ -1,5 +1,6 @@
 import { Injectable, Logger } from '@nestjs/common';
 import { InjectKysely } from 'nestjs-kysely';
+import { sql } from 'kysely';
 import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
 import { dbOrTx } from '../../utils';
 import {
@@ -188,6 +189,144 @@ export class AiChatMessageRepo {
     return query.returning(this.baseFields).executeTakeFirst();
   }
 
+  /**
+   * #487 OWNER terminal write — the streamText terminal callback's finalize. Like
+   * `update` but CONDITIONAL on `status='streaming' OR metadata.finalizeFailed`:
+   * the owner writes its real content EITHER when the row is still streaming (the
+   * normal case) OR when a reconcile stamp already flipped it to a terminal status
+   * but marked `finalizeFailed:true` — the owner's real content OVERWRITES that
+   * placeholder stamp (owner-write priority, #487). A row that is properly terminal
+   * (no finalizeFailed) is left untouched (undefined) — idempotent. The `patch`
+   * carries the real metadata WITHOUT finalizeFailed, so a successful write CLEARS
+   * the flag. Returns the updated row, or undefined when nothing matched.
+   */
+  async finalizeOwner(
+    id: string,
+    workspaceId: string,
+    patch: Partial<{
+      content: string | null;
+      toolCalls: unknown;
+      metadata: unknown;
+      status: string | null;
+    }>,
+    trx?: KyselyTransaction,
+  ): Promise<AiChatMessage | undefined> {
+    const db = dbOrTx(this.db, trx);
+    return db
+      .updateTable('aiChatMessages')
+      .set({ ...(patch as Record<string, unknown>), updatedAt: new Date() })
+      .where('id', '=', id)
+      .where('workspaceId', '=', workspaceId)
+      .where((eb) =>
+        eb.or([
+          eb('status', '=', 'streaming'),
+          eb(sql<string>`(metadata->>'finalizeFailed')`, '=', 'true'),
+        ]),
+      )
+      .returning(this.baseFields)
+      .executeTakeFirst();
+  }
+
+  /**
+   * #487 RECONCILE status-only stamp — settle a stuck 'streaming' row to a
+   * terminal status WITHOUT the owner's real content (which lived only in the
+   * dead process's memory — a documented loss). CONDITIONAL on `status='streaming'`
+   * (never touches an already-terminal row) AND it MERGES `finalizeFailed:true`
+   * into metadata (preserving the partial `parts` already persisted) so a LATER
+   * owner-write (finalizeOwner) can still OVERWRITE this placeholder with real
+   * content, and so `isInterruptResume` can EXCLUDE this row (a reconcile stamp is
+   * not a genuine user interruption). Returns the updated row, or undefined.
+   */
+  async stampTerminalIfStreaming(
+    id: string,
+    workspaceId: string,
+    status: 'aborted' | 'error' | 'completed',
+    trx?: KyselyTransaction,
+  ): Promise<AiChatMessage | undefined> {
+    const db = dbOrTx(this.db, trx);
+    return db
+      .updateTable('aiChatMessages')
+      .set({
+        status,
+        metadata: sql`coalesce(metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
+        updatedAt: new Date(),
+      })
+      .where('id', '=', id)
+      .where('workspaceId', '=', workspaceId)
+      .where('status', '=', 'streaming')
+      .returning(this.baseFields)
+      .executeTakeFirst();
+  }
+
+  /**
+   * #487 reconcile clause (b): streaming assistant rows whose linked RUN has
+   * already reached a terminal status — an asymmetry ("run settled / message
+   * streaming forever") the periodic reconcile heals by stamping the message.
+   * Returns the message id + its run's terminal status, bounded.
+   */
+  async findStreamingWithTerminalRun(
+    limit = 200,
+    // #487: scope to ONE chat for the opportunistic per-turn reconcile (removes
+    // reconcile latency from the user-visible path); omit for the periodic sweep.
+    chat?: { chatId: string; workspaceId: string },
+  ): Promise<
+    Array<{ messageId: string; workspaceId: string; runStatus: string }>
+  > {
+    let query = this.db
+      .selectFrom('aiChatMessages as m')
+      .innerJoin('aiChatRuns as r', 'r.assistantMessageId', 'm.id')
+      .select([
+        'm.id as messageId',
+        'm.workspaceId as workspaceId',
+        'r.status as runStatus',
+      ])
+      .where('m.status', '=', 'streaming')
+      .where('r.status', 'in', ['succeeded', 'failed', 'aborted']);
+    if (chat) {
+      query = query
+        .where('m.chatId', '=', chat.chatId)
+        .where('m.workspaceId', '=', chat.workspaceId);
+    }
+    return query.limit(limit).execute();
+  }
+
+  /**
+   * #487 reconcile clause (d) — historical-row safety: streaming rows older than
+   * `staleMs` whose chat has NO active run row (double-gated). Settle them to
+   * 'aborted' + finalizeFailed (so a late owner-write could still overwrite).
+   * Returns the count. Used ONLY by the periodic reconcile, never at boot.
+   */
+  async sweepStreamingWithoutActiveRun(
+    staleMs: number,
+    trx?: KyselyTransaction,
+  ): Promise<number> {
+    const db = dbOrTx(this.db, trx);
+    const staleBefore = new Date(Date.now() - staleMs);
+    const rows = await db
+      .updateTable('aiChatMessages as m')
+      .set({
+        status: 'aborted',
+        metadata: sql`coalesce(m.metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
+        updatedAt: new Date(),
+      })
+      .where('m.status', '=', 'streaming')
+      .where('m.updatedAt', '<', staleBefore)
+      .where((eb) =>
+        eb.not(
+          eb.exists(
+            eb
+              .selectFrom('aiChatRuns as r')
+              .select('r.id')
+              .whereRef('r.chatId', '=', 'm.chatId')
+              .where('r.status', 'in', ['pending', 'running']),
+          ),
+        ),
+      )
+      .returning('m.id')
+      .execute();
+    return rows.length;
+  }
+
   /**
    * Crash-recovery sweep (#183): flip every assistant row still left in the
    * 'streaming' state (a turn that died mid-write before reaching a terminal
@@ -200,13 +339,20 @@ export class AiChatMessageRepo {
    * step, so an actively-streaming row never matches; this prevents a fresh
    * replica's boot-sweep from aborting a turn another replica is still streaming
    * in a multi-instance deploy.
+   *
+   * #487: the sweep now ALSO marks `finalizeFailed:true` so a late owner-write can
+   * overwrite this placeholder with real content (owner-write priority).
    */
   async sweepStreaming(trx?: KyselyTransaction): Promise<number> {
     const db = dbOrTx(this.db, trx);
     const staleBefore = new Date(Date.now() - SWEEP_STREAMING_STALE_MS);
     const rows = await db
       .updateTable('aiChatMessages')
-      .set({ status: 'aborted', updatedAt: new Date() })
+      .set({
+        status: 'aborted',
+        metadata: sql`coalesce(metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
+        updatedAt: new Date(),
+      })
       .where('status', '=', 'streaming')
       .where('updatedAt', '<', staleBefore)
       .returning('id')

apps/server/src/database/repos/ai-chat/ai-chat-run.repo.ts

@@ -143,6 +143,41 @@ export class AiChatRunRepo {
       .executeTakeFirst();
   }
 
+  /**
+   * #487: CONDITIONAL terminal finalize — flip a run to a terminal status and
+   * stamp `finished_at` ONLY while it is still active (pending|running), mirroring
+   * the assistant message's `onlyIfStreaming` guard. A double-settle (a late or
+   * second writer, a supersede applying a zombie's intended, a reconcile stamp)
+   * matches NOTHING once the row is terminal and is a benign no-op — so a terminal
+   * status can never be clobbered by a later writer (last-writer-wins is gone).
+   *
+   * Returns the updated row when it WAS active (this call wrote it), else
+   * undefined (the row was already terminal — another writer won). The caller
+   * distinguishes the two to resolve the correct settle outcome.
+   */
+  async finalizeIfActive(
+    id: string,
+    workspaceId: string,
+    patch: { status: string; error: string | null },
+    trx?: KyselyTransaction,
+  ): Promise<AiChatRun | undefined> {
+    const db = dbOrTx(this.db, trx);
+    const now = new Date();
+    return db
+      .updateTable('aiChatRuns')
+      .set({
+        status: patch.status,
+        error: patch.error,
+        finishedAt: now,
+        updatedAt: now,
+      })
+      .where('id', '=', id)
+      .where('workspaceId', '=', workspaceId)
+      .where('status', 'in', ACTIVE_RUN_STATUSES as unknown as string[])
+      .returning(this.baseFields)
+      .executeTakeFirst();
+  }
+
   /**
    * Mark an EXPLICIT stop request on an active run (distinct from a browser
    * disconnect, which never stops a run). Stamps `stop_requested_at` ONLY while
@@ -184,6 +219,31 @@ export class AiChatRunRepo {
    * sweeps only runs UNTOUCHED past the window. Phase 1 is single-process, so the
    * boot path supplies no window.
    */
+  /**
+   * #487 reconcile clause (c): active (pending|running) runs UNTOUCHED past
+   * `staleMs` — candidates for "no live runner" abort. Staleness is measured from
+   * `updated_at` (the LAST-PROGRESS timestamp — recordStep bumps it), NOT
+   * `started_at`, so a legitimate long-running marathon (11–25 min of steady
+   * progress) is never a candidate. The caller filters these against its in-memory
+   * `active` / zombie maps ("no entry" is the PRIMARY gate — a live entry is never
+   * aborted) before settling any of them. Bounded.
+   */
+  async findStaleActive(
+    staleMs: number,
+    limit = 200,
+    trx?: KyselyTransaction,
+  ): Promise<Array<{ id: string; workspaceId: string; chatId: string }>> {
+    const db = dbOrTx(this.db, trx);
+    const staleBefore = new Date(Date.now() - staleMs);
+    return db
+      .selectFrom('aiChatRuns')
+      .select(['id', 'workspaceId', 'chatId'])
+      .where('status', 'in', ACTIVE_RUN_STATUSES as unknown as string[])
+      .where('updatedAt', '<', staleBefore)
+      .limit(limit)
+      .execute();
+  }
+
   async sweepRunning(
     opts: { staleMs?: number } = {},
     trx?: KyselyTransaction,

apps/server/src/database/types/db.d.ts

@@ -606,6 +606,9 @@ export interface AiChats {
   // The document the chat was created in (open page at first message). NULL =>
   // started outside any document. ON DELETE SET NULL on the page FK.
   pageId: string | null;
+  // Chat-level metadata bag (#490). jsonb, defaulted to '{}'. First key:
+  // `activatedTools` — the deferred-tool activation set persisted across turns.
+  metadata: Generated<Json>;
   createdAt: Generated<Timestamp>;
   updatedAt: Generated<Timestamp>;
   deletedAt: Timestamp | null;

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/ai/ai-settings.service.ts

@@ -245,6 +245,9 @@ export class AiSettingsService {
       // Max context window for the chat header badge denominator. Stored as
       // ::text; 0/unset/invalid = no limit (undefined).
       chatContextWindow: parsePositiveInt(provider.chatContextWindow),
+      // RAW stored value (#490): the replay budgeter reads this to distinguish an
+      // explicit `0` (off-switch) from unset, which parsePositiveInt cannot.
+      chatContextWindowRaw: provider.chatContextWindow,
       // Plain passthrough; getChatModel defaults unset to 'openai-compatible'.
       chatApiStyle: provider.chatApiStyle,
       // Cheap model id for the anonymous public-share assistant; reuses the chat

apps/server/src/integrations/ai/ai-streaming-fetch.ts

@@ -129,6 +129,12 @@ const DEFAULT_MCP_STREAM_TIMEOUT_MS = 60_000;
 /** Default total wall-clock cap for ONE external MCP tool call (2 min). */
 const DEFAULT_MCP_CALL_TIMEOUT_MS = 120_000;
 
+/**
+ * Default `bodyTimeout` for the EXTERNAL-MCP SSE transport (10 min) — #489.
+ * Deliberately much LARGER than {@link DEFAULT_MCP_STREAM_TIMEOUT_MS}.
+ */
+const DEFAULT_MCP_SSE_BODY_TIMEOUT_MS = 600_000;
+
 /**
  * SILENCE timeout (ms) for EXTERNAL-MCP transport ONLY. Override with
  * `AI_MCP_STREAM_TIMEOUT_MS`; a missing/invalid/non-positive value falls back to
@@ -164,6 +170,26 @@ export function mcpCallTimeoutMs(): number {
   return positiveEnv('AI_MCP_CALL_TIMEOUT_MS', DEFAULT_MCP_CALL_TIMEOUT_MS);
 }
 
+/**
+ * `bodyTimeout` (ms) for the EXTERNAL-MCP **SSE** transport ONLY — #489. Override
+ * with `AI_MCP_SSE_BODY_TIMEOUT_MS`; a missing/invalid/non-positive value falls
+ * back to {@link DEFAULT_MCP_SSE_BODY_TIMEOUT_MS} (10 min).
+ *
+ * The SSE transport holds ONE long-lived response body open across many tool
+ * calls, so undici's `bodyTimeout` (time between body bytes) counts the LEGITIMATE
+ * silence BETWEEN calls, not just a hung single call. At the tight HTTP silence
+ * timeout ({@link mcpStreamTimeoutMs}, 1 min) a normal >1-min gap between the
+ * model's tool calls would break the SSE socket, and the cache would then serve a
+ * dead client until TTL. So the SSE transport gets its OWN, RAISED bodyTimeout;
+ * the per-call total cap ({@link mcpCallTimeoutMs}) still bounds a single stuck
+ * call, and the app-level transport-error retry heals a socket that does break.
+ * The HTTP (streamable) transport keeps the tight timeout — it opens a fresh
+ * request per call, so idle-between-calls does not apply there.
+ */
+export function mcpSseBodyTimeoutMs(): number {
+  return positiveEnv('AI_MCP_SSE_BODY_TIMEOUT_MS', DEFAULT_MCP_SSE_BODY_TIMEOUT_MS);
+}
+
 /**
  * undici `Agent` options for streaming AI traffic — the (generous, finite)
  * silence timeouts plus the keep-alive recycle window. Shared by the chat

apps/server/src/integrations/ai/ai.types.ts

@@ -105,6 +105,10 @@ export interface ResolvedAiConfig extends Partial<AiProviderSettings> {
   // Max context window in tokens; surfaced to the chat header badge as the
   // "current / max" denominator. 0/unset = no limit.
   chatContextWindow?: number;
+  // RAW stored context window (::text), BEFORE parsePositiveInt collapses `0` and
+  // unset to `undefined`. The #490 replay budgeter needs the raw value to honor an
+  // explicit `0` off-switch distinctly from "unset -> flat default".
+  chatContextWindowRaw?: string | number;
   // Cheap model id for the public-share assistant; reuses the chat creds.
   publicShareChatModel?: string;
   // Agent-role id whose persona the public-share assistant adopts (empty/unset

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(), {
-        maxRetriesPerRequest: 15,
-      });
-
+      const redis = this.getProbeClient();
+      // 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

@@ -119,10 +119,42 @@ 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:
   // it is only a fallback when no per-user Basic/Bearer credentials are sent.

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
-  // another container as docmost:9464. The port is not published to the host.
-  server.listen(port, '0.0.0.0', () => {
-    logger.log(`Metrics endpoint listening on :${port}/metrics`);
+  // Bind to loopback by default so the auth-less endpoint is not exposed on all
+  // interfaces. Set METRICS_BIND=0.0.0.0 (ideally with METRICS_TOKEN) when the
+  // scraper runs in a separate container and reaches this as docmost:9464.
+  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',

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

@@ -0,0 +1,305 @@
+import { Kysely } from 'kysely';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunRepo } from '@docmost/db/repos/ai-chat/ai-chat-run.repo';
+import { AiChatRunService } from '../../src/core/ai-chat/ai-chat-run.service';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+  createMessage,
+} from './db';
+
+/**
+ * #487 commit 4 — bidirectional reconcile + owner-write priority, real SQL.
+ *
+ * Proves the OBSERVABLE recovery properties against docmost_test:
+ *  - the CONDITIONAL owner-write beats a reconcile stamp, and a stamp never
+ *    clobbers a proper terminal row;
+ *  - a LATE owner-finalize with real content OVERWRITES a reconcile 'aborted'
+ *    stamp (finalizeFailed);
+ *  - each reconcile clause (b message<-run, c stale-run, d historical row) settles
+ *    the stuck row/run, and a LIVE run entry is never touched;
+ *  - the "kill DB on finish" recovery: after the DB comes back, neither the
+ *    message row nor the run row stays stuck.
+ */
+describe('#487 reconcile + owner-write priority [integration]', () => {
+  let db: Kysely<any>;
+  let messageRepo: AiChatMessageRepo;
+  let runRepo: AiChatRunRepo;
+  let runService: AiChatRunService;
+  let workspaceId: string;
+  let userId: string;
+
+  beforeAll(async () => {
+    db = getTestDb();
+    messageRepo = new AiChatMessageRepo(db as any);
+    runRepo = new AiChatRunRepo(db as any);
+    runService = new AiChatRunService(runRepo, { isCloud: () => false } as never);
+    workspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  const newChat = async () =>
+    (await createChat(db, { workspaceId, creatorId: userId })).id;
+
+  const metaOf = async (id: string): Promise<Record<string, unknown> | null> => {
+    const row = await messageRepo.findById(id, workspaceId);
+    return (row?.metadata as Record<string, unknown> | null) ?? null;
+  };
+
+  it('owner finalizeOwner writes a streaming row and CLEARS finalizeFailed', async () => {
+    const chatId = await newChat();
+    const m = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [] },
+    });
+    const wrote = await messageRepo.finalizeOwner(m.id, workspaceId, {
+      content: 'final answer',
+      status: 'completed',
+      metadata: { parts: [{ type: 'text', text: 'final answer' }] },
+    } as never);
+    expect(wrote!.status).toBe('completed');
+    expect((await metaOf(m.id))?.finalizeFailed).toBeUndefined();
+  });
+
+  it('a reconcile stamp NEVER clobbers a proper terminal row (finalizeOwner is a no-op there)', async () => {
+    const chatId = await newChat();
+    const m = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'completed',
+      content: 'real',
+      metadata: { parts: [] },
+    });
+    // The reconcile stamp is onlyIfStreaming -> no-op on a completed row.
+    const stamped = await messageRepo.stampTerminalIfStreaming(
+      m.id,
+      workspaceId,
+      'aborted',
+    );
+    expect(stamped).toBeUndefined();
+    expect((await messageRepo.findById(m.id, workspaceId))!.status).toBe(
+      'completed',
+    );
+  });
+
+  it('LATE owner-finalize with real content OVERWRITES a reconcile aborted stamp', async () => {
+    const chatId = await newChat();
+    const m = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [{ type: 'text', text: 'partial' }] },
+    });
+    // Reconcile stamps it aborted + finalizeFailed (final text lived only in mem).
+    const stamped = await messageRepo.stampTerminalIfStreaming(
+      m.id,
+      workspaceId,
+      'aborted',
+    );
+    expect(stamped!.status).toBe('aborted');
+    expect((await metaOf(m.id))?.finalizeFailed).toBe(true);
+
+    // A LATE owner-write (finalizeFailed=true satisfies the OR) overwrites it with
+    // real content, clearing the flag — owner-write priority.
+    const wrote = await messageRepo.finalizeOwner(m.id, workspaceId, {
+      content: 'the real final answer',
+      status: 'completed',
+      metadata: { parts: [{ type: 'text', text: 'the real final answer' }] },
+    } as never);
+    expect(wrote!.status).toBe('completed');
+    expect(wrote!.content).toBe('the real final answer');
+    expect((await metaOf(m.id))?.finalizeFailed).toBeUndefined();
+  });
+
+  it('clause (c): a stale active run with NO live entry -> aborted; a LIVE entry is untouched', async () => {
+    // Stale run, NOT owned by this replica (no entry) -> reconcile aborts it.
+    const staleChat = await newChat();
+    const stale = await runRepo.insert({
+      chatId: staleChat,
+      workspaceId,
+      createdBy: userId,
+      status: 'running',
+    });
+    await db
+      .updateTable('aiChatRuns')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', stale.id)
+      .execute();
+
+    // A live run OWNED by this replica (beginRun registers an in-memory entry),
+    // ALSO backdated stale — the "no entry" primary gate must protect it.
+    const liveChat = await newChat();
+    const live = await runService.beginRun({
+      chatId: liveChat,
+      workspaceId,
+      userId,
+    });
+    await db
+      .updateTable('aiChatRuns')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', live.runId)
+      .execute();
+
+    const aborted = await runService.reconcileStaleRuns(15 * 60 * 1000);
+    expect(aborted).toBeGreaterThanOrEqual(1);
+    expect((await runRepo.findById(stale.id, workspaceId))!.status).toBe(
+      'aborted',
+    );
+    // The live entry is NEVER aborted, however stale its row looks.
+    expect((await runRepo.findById(live.runId, workspaceId))!.status).toBe(
+      'running',
+    );
+    expect(runService.isLocallyActive(live.runId)).toBe(true);
+
+    // cleanup the live run
+    await runService.finalizeRun(live.runId, workspaceId, 'aborted');
+  });
+
+  it('clause (b): a streaming message whose RUN is terminal is stamped by run status (succeeded -> aborted, NOT completed-empty)', async () => {
+    const chatId = await newChat();
+    const msg = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [] },
+    });
+    // A SUCCEEDED run linked to the still-streaming message (the asymmetry).
+    const run = await runRepo.insert({
+      chatId,
+      workspaceId,
+      createdBy: userId,
+      status: 'running',
+      assistantMessageId: msg.id,
+    });
+    await runRepo.finalizeIfActive(run.id, workspaceId, {
+      status: 'succeeded',
+      error: null,
+    });
+
+    const stuck = await messageRepo.findStreamingWithTerminalRun();
+    const mine = stuck.find((s) => s.messageId === msg.id);
+    expect(mine?.runStatus).toBe('succeeded');
+    // Reconcile clause (b): succeeded run -> message 'aborted' (NOT 'completed'),
+    // the final text lived only in memory (documented loss), +finalizeFailed.
+    const status = mine!.runStatus === 'failed' ? 'error' : 'aborted';
+    await messageRepo.stampTerminalIfStreaming(msg.id, workspaceId, status);
+    const row = await messageRepo.findById(msg.id, workspaceId);
+    expect(row!.status).toBe('aborted');
+    expect((row!.metadata as Record<string, unknown>).finalizeFailed).toBe(true);
+  });
+
+  it('clause (d): a stale streaming row with NO active run on the chat -> aborted+finalizeFailed', async () => {
+    const chatId = await newChat();
+    const msg = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [] },
+    });
+    await db
+      .updateTable('aiChatMessages')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', msg.id)
+      .execute();
+
+    const swept = await messageRepo.sweepStreamingWithoutActiveRun(
+      15 * 60 * 1000,
+    );
+    expect(swept).toBeGreaterThanOrEqual(1);
+    const row = await messageRepo.findById(msg.id, workspaceId);
+    expect(row!.status).toBe('aborted');
+    expect((row!.metadata as Record<string, unknown>).finalizeFailed).toBe(true);
+  });
+
+  it('clause (d) is DOUBLE-GATED: a stale streaming row WITH an active run on the chat is left alone', async () => {
+    const chatId = await newChat();
+    const msg = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [] },
+    });
+    await db
+      .updateTable('aiChatMessages')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', msg.id)
+      .execute();
+    // An ACTIVE run on the same chat -> clause (d) must NOT touch the message.
+    const run = await runRepo.insert({
+      chatId,
+      workspaceId,
+      createdBy: userId,
+      status: 'running',
+    });
+
+    await messageRepo.sweepStreamingWithoutActiveRun(15 * 60 * 1000);
+    expect((await messageRepo.findById(msg.id, workspaceId))!.status).toBe(
+      'streaming',
+    );
+    await runRepo.finalizeIfActive(run.id, workspaceId, {
+      status: 'aborted',
+      error: null,
+    });
+  });
+
+  it('"kill DB on finish" recovery: after the DB is back, reconcile leaves NEITHER the row nor the run stuck', async () => {
+    // Simulate a process that seeded the assistant row + run, then died before
+    // finalizing EITHER (a mid-turn crash): a streaming message + a running run,
+    // both stale, with no in-memory entry (fresh service = fresh maps).
+    const chatId = await newChat();
+    const msg = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      status: 'streaming',
+      metadata: { parts: [{ type: 'text', text: 'partial' }] },
+    });
+    const run = await runRepo.insert({
+      chatId,
+      workspaceId,
+      createdBy: userId,
+      status: 'running',
+      assistantMessageId: msg.id,
+    });
+    await db
+      .updateTable('aiChatRuns')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', run.id)
+      .execute();
+    await db
+      .updateTable('aiChatMessages')
+      .set({ updatedAt: new Date(Date.now() - 60 * 60 * 1000) })
+      .where('id', '=', msg.id)
+      .execute();
+
+    // Reconcile (as the periodic job would): (c) aborts the orphan run, then
+    // (b) settles the message from the now-terminal run.
+    await runService.reconcileStaleRuns(15 * 60 * 1000);
+    const stuck = await messageRepo.findStreamingWithTerminalRun();
+    for (const s of stuck) {
+      const status = s.runStatus === 'failed' ? 'error' : 'aborted';
+      await messageRepo.stampTerminalIfStreaming(s.messageId, s.workspaceId, status);
+    }
+
+    // Neither is stuck: the run is terminal AND the message is terminal.
+    expect((await runRepo.findById(run.id, workspaceId))!.status).toBe('aborted');
+    const row = await messageRepo.findById(msg.id, workspaceId);
+    expect(row!.status).toBe('aborted');
+    expect((row!.metadata as Record<string, unknown>).finalizeFailed).toBe(true);
+  });
+});

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

@@ -281,6 +281,52 @@ describe('AiChatRun durable lifecycle [integration]', () => {
     });
   });
 
+  it('#487 finalizeIfActive is CONDITIONAL: a late terminal write cannot clobber the settled status (real SQL)', async () => {
+    const c = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const run = await runRepo.insert({
+      chatId: c,
+      workspaceId,
+      createdBy: userId,
+      status: 'running',
+    });
+
+    // First terminal write: the run IS active, so it flips + returns the row.
+    const first = await runRepo.finalizeIfActive(run.id, workspaceId, {
+      status: 'succeeded',
+      error: null,
+    });
+    expect(first!.status).toBe('succeeded');
+    expect(first!.finishedAt).toBeTruthy();
+
+    // A late/second writer tries to flip it to 'aborted' — the WHERE status IN
+    // ('pending','running') guard matches NOTHING now, so it is a benign no-op.
+    const second = await runRepo.finalizeIfActive(run.id, workspaceId, {
+      status: 'aborted',
+      error: 'late clobber attempt',
+    });
+    expect(second).toBeUndefined();
+
+    // The persisted terminal status is UNCHANGED — last-writer-wins is gone.
+    const row = await runRepo.findById(run.id, workspaceId);
+    expect(row!.status).toBe('succeeded');
+    expect(row!.error).toBeNull();
+  });
+
+  it('#487 double-settle through the service collapses to one write at the SQL gate', async () => {
+    const c = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const handle = await service.beginRun({ chatId: c, workspaceId, userId });
+
+    // First settle writes 'aborted' via the conditional write.
+    await service.finalizeRun(handle.runId, workspaceId, 'aborted');
+    // A late safety-net settle to 'error' is a no-op (row already terminal).
+    await service.finalizeRun(handle.runId, workspaceId, 'error', 'late');
+
+    const row = await runRepo.findById(handle.runId, workspaceId);
+    expect(row!.status).toBe('aborted');
+    expect(service.isLocallyActive(handle.runId)).toBe(false);
+    expect(service.hasZombie(handle.runId)).toBe(false);
+  });
+
   it('sweepRunning() with NO args (boot sweep / variant C) aborts even a FRESH running run', async () => {
     // F1/DECISION C at the SQL level: the unconditional boot sweep has NO
     // staleness window, so a run updated just now (a fast restart) is settled too

docs/reading-ai-logs.md

@@ -8,19 +8,35 @@ real pain (a "which tools fail most?" analysis that confidently answered
 
 Read the **Gotchas** section before you trust any error count.
 
+> **TWO ERAS — check the marker first.** The `tool_calls` shape changed in **#490
+> (trace v2)**. A row written by v2 carries `metadata.toolTraceVersion = 2`; older
+> rows have no such key. The two shapes store DIFFERENT things (v2 dropped the tool
+> OUTPUT from the trace), so **every query below is dual-shape** — branch on the
+> marker. **Never compare an aggregate or trend across the era boundary**: a metric
+> jump on the cut-over week is an artifact of the shape change, not a behavior
+> change.
+
 ## TL;DR
 
 - Agent chats live in Postgres, DB `docmost`, tables `ai_chat_*`.
-- Each tool invocation is stored as **two** array elements (a `tool-call` part and
-  a `tool-result` part), so naive counting double-counts.
-- **A tool that *throws* writes no result part.** Since the #407 fix its error is
-  persisted as a dedicated `{toolName, error}` element in `tool_calls` (queryable +
-  replayed to the model). **Rows written before #407 still drop it** — the error is
-  nowhere in the DB and shows only in the live UI. So `isError` / `success=false`
-  scans under-report by design, and pre-#407 thrown errors are invisible.
-- To find where agents fail: (1) soft-failure markers in `tool_calls`, (2) the new
-  `error` field for thrown errors (new rows) / the orphan-gap proxy (old rows),
-  (3) server logs / the live UI for full stack traces beyond the truncated message.
+- **Era marker:** `metadata.toolTraceVersion = 2` ⇒ v2 (#490) row; absent ⇒ legacy row.
+- Each tool invocation is stored as **two** consecutive array elements — a
+  `tool-call` part then an OUTCOME part — so naive counting double-counts.
+  - **v2 (#490):** outcome is `{toolName, ok: true}` on success, or
+    `{toolName, error, kind: 'thrown'|'interrupted'}` on failure. The tool **OUTPUT
+    is NOT in `tool_calls`** any more — it lives once in `metadata.parts` (this
+    removed a hundreds-of-MB-per-run write duplication). Soft-failure analysis
+    therefore reads `metadata.parts`, not `tool_calls`.
+  - **legacy:** outcome is `{toolName, output}` on success; a **thrown** failure is
+    a `{toolName, error}` element **only on rows after #407**, and is dropped
+    entirely (silent orphan) on pre-#407 rows.
+- **A tool that *throws* writes no result part.** In v2 it is a
+  `{error, kind:'thrown'}` element; an interrupted/aborted call is a distinct
+  `{error, kind:'interrupted'}`. `isError`/`success=false` scans read the *output*
+  and so under-report thrown failures in every era.
+- To find where agents fail: (1) soft-failure markers in `metadata.parts` outputs
+  (v2) / `tool_calls` outputs (legacy), (2) the `error`/`kind` fields for thrown
+  failures (v2 + post-#407), (3) server logs / the live UI for full stack traces.
 
 ## Where the data lives
 
@@ -53,33 +69,67 @@ are rows in `workspaces`, not separate deployments.
 separate `tool` role), `content` (text), `tool_calls` (jsonb array), `metadata`
 (jsonb, holds run `error` + rendered `parts`), `status`, `tsv` (full-text index).
 
+## Era marker — check this before every query
+
+```sql
+-- how many rows are in each era?
+SELECT COALESCE((metadata->>'toolTraceVersion'), 'legacy') AS era, count(*)
+FROM ai_chat_messages
+WHERE role = 'assistant' AND jsonb_typeof(tool_calls) = 'array'
+GROUP BY 1 ORDER BY 2 DESC;
+```
+
+- `toolTraceVersion = '2'` → **v2** (#490): outcome flags, **no output in the trace**.
+- `NULL` (`'legacy'`) → pre-#490: outcome carries the tool `output` inline.
+
+**Do not trend a metric across the cut-over.** The shape change alone shifts counts
+(e.g. "elements with `output`" collapses to zero for v2), so a week that straddles
+the boundary shows an artifact, not a behavior change. Segment by era, or restrict to
+one era, before comparing.
+
 ## How tool calls are stored — READ THIS
 
 Tool calls are **not** one-object-per-call. Each logical invocation is split into
-two consecutive elements of the `tool_calls` array:
+two consecutive elements of the `tool_calls` array — a **call** then an **outcome**.
+The outcome shape is era-dependent:
 
 ```text
-index 0: { "toolName": "getPage",  "input":  { "pageId": "…" } }   ← tool-call   (has input, NO output)
-index 1: { "toolName": "getPage",  "output": { … } }                ← tool-result (has output, NO input)
+# v2 (#490) — metadata.toolTraceVersion = 2
+index 0: { "toolName":"getPage", "input":{...} }                       ← call     (has input)
+index 1: { "toolName":"getPage", "ok":true }                          ← success  (NO output here)
+   or  : { "toolName":"getPage", "error":"…", "kind":"thrown" }       ← threw
+   or  : { "toolName":"getPage", "error":"…", "kind":"interrupted" }  ← aborted mid-step
+
+# legacy — no toolTraceVersion
+index 0: { "toolName":"getPage", "input":{...} }                       ← call     (has input, NO output)
+index 1: { "toolName":"getPage", "output":{...} }                     ← success  (has output)
+   or  : { "toolName":"getPage", "error":"…" }                        ← threw (post-#407 only)
 ```
 
-The keys that appear on an element are `toolName`, `input`, `output`, and — for a
-**thrown** failure on rows written after the #407 fix — `error` (the tool's error
-message; see the "Hard failures" section below). There is no `state`, no `errorText`,
-no `type`. On pre-#407 rows a thrown failure has NO paired result element at all
-(silent orphan). Consequences:
+The keys that can appear: `toolName`, `input` (call), and on the outcome — **v2:**
+`ok` **or** `error`+`kind`; **legacy:** `output` **or** (post-#407) `error`. There is
+no `state`, no `errorText`, no `type` in `tool_calls` (those live on `metadata.parts`).
+Consequences:
 
-1. **Real invocation count = elements that have `output` or `error`.** Counting every
-   element double-counts (you get ~2× and a spurious "~50% of every tool has no output").
-2. **Pairing:** a call = a `tool-call` part followed by its result part. A success
-   carries `output`; a thrown failure (post-#407) carries `error` instead. Both carry
-   `toolName`, so you can group by tool on either.
+1. **Real invocation count** — count the OUTCOME elements, not every element (else you
+   double-count): **v2** = elements with `ok` or `error`; **legacy** = elements with
+   `output` or `error`.
+2. **Pairing:** a call (`input`) is followed by its outcome. `toolName` is on both, so
+   you can group by tool on either. In v2 the `kind` field separates a real hard-fail
+   (`thrown`) from an aborted call (`interrupted`) — a distinction legacy rows cannot
+   make (both are orphans; see below).
+3. **The tool OUTPUT is only in `metadata.parts` on v2 rows.** To inspect what a tool
+   returned (soft-error markers, page bodies) on a v2 row, read the parts
+   (`part->>'type' LIKE 'tool-%'`, `part->>'state' = 'output-available'`, `part->'output'`),
+   not `tool_calls`.
 
 ## The two classes of failure (and which the DB can see)
 
 ### 1. Soft failures — tool RAN and returned an error-shaped result → PERSISTED ✅
 
-These are visible in the `tool-result` `output`. The marker differs per tool:
+These are visible in the tool `output` — **on v2 rows in `metadata.parts`** (the
+`output-available` part's `output`), on **legacy rows in the `tool_calls` outcome
+element's `output`**. The marker differs per tool:
 
 | Tool(s) | Error marker in `output` |
 | --- | --- |
@@ -91,37 +141,32 @@ These are visible in the `tool-result` `output`. The marker differs per tool:
 Note `editPageText` returns `failed: []` on success — filtering on the *presence*
 of the key gives false positives; filter on **non-empty**.
 
-### 2. Hard failures — tool THREW → NOW PERSISTED ✅ (since the #407 fix)
+### 2. Hard failures — tool THREW → PERSISTED ✅
 
 When a tool throws (the classic one is `patchNode` / `insertNode` / `tableUpdateCell`
 → `Failed to encode document to Yjs (fromJSON): Unknown node type: undefined`), the
-runtime still writes **no `tool-result` part** — the failure is an ai@6 `tool-error`
-content part instead. **Since the #407 fix, that error is persisted**: `serializeSteps`
-appends a dedicated element `{toolName, error: "<message>"}` right after the failed
-call, mirroring how a successful `{toolName, output}` element is appended. So a thrown
-error now leaves a queryable `error` field carrying its (truncated) reason, and the
-same real text is replayed to the model on the next turn (an `output-error` part with
-the real `errorText`, no longer the `'Tool call did not complete.'` placeholder).
+runtime writes **no `tool-result` part** — the failure is an ai@6 `tool-error` content
+part. How that lands in `tool_calls` depends on the era:
 
-**Cutover caveat — old rows keep the old blind shape.** Rows written **before** this
-change have the two-part shape (`call` + `output` only) and simply **drop** thrown
-errors, leaving a silent **orphan** (a `call` with no `output` *and* no `error`). Rows
-written **after** the fix additionally carry the `error` element. So:
+- **v2 (#490):** a `{toolName, error, kind:'thrown'}` outcome element. An interrupted /
+  aborted mid-step call is a **distinct** `{toolName, error:'Tool call did not
+  complete.', kind:'interrupted'}` element — so you can tell a real hard-fail from an
+  abort **directly, without the orphan heuristic**. Query `kind = 'thrown'`.
+- **post-#407 legacy:** a `{toolName, error}` element (no `kind`) right after the call.
+- **pre-#407 legacy:** the error is **dropped** — a silent **orphan** (a `call` with no
+  `output` *and* no `error`).
 
-- **New rows:** query the `error` field directly (see the hard-error query below) — no
-  orphan heuristic needed for thrown failures.
-- **Old rows (pre-#407):** the only DB-side proxy is still an **orphan**: a `tool-call`
-  part with no matching `tool-result` *and* no `error`. Orphans also appear when a run
-  is **aborted** mid-flight (server restart), so a high-volume tool (`createComment`,
-  `searchInPage`, `Search_web_search`) shows orphans from aborts, not real errors on
-  old rows. Treat the orphan gap as an *upper bound*, and cross-check the tool: a gap on
-  a structural editor (`patchNode`, `insertNode`, `updatePageJson`, `transformPage`) is
-  almost certainly a thrown Yjs-encode error; a gap on `createComment` is mostly aborts.
+The same real error text is replayed to the model on the next turn (an `output-error`
+part with the real `errorText`, from `metadata.parts`), in every era.
 
-A note on the aborted-call fallback: a call with **neither** a result **nor** a
-`tool-error` (genuinely interrupted mid-step) still replays with the
-`'Tool call did not complete.'` placeholder and persists as an orphan — that path is
-unchanged, and is distinct from a real thrown error, which now carries `error`.
+**Cutover caveat.** Only pre-#407 legacy rows need the orphan proxy: an orphan is a
+`tool-call` with no matching outcome. Orphans there also appear when a run is **aborted**
+mid-flight (server restart), so a high-volume tool (`createComment`, `searchInPage`,
+`Search_web_search`) shows orphans from aborts, not real errors. Treat the orphan gap as
+an *upper bound* and cross-check the tool: a gap on a structural editor (`patchNode`,
+`insertNode`, `updatePageJson`, `transformPage`) is almost certainly a thrown Yjs-encode
+error; a gap on `createComment` is mostly aborts. **On v2 rows this ambiguity is gone**
+— `kind` labels each outcome.
 
 ### 3. Run-level failures → `ai_chat_runs`
 
@@ -134,22 +179,34 @@ the wild: `Run interrupted by a server restart.` (aborts) and
 
 Run all of these via `docker exec gitmost-postgresql psql -U docmost -d docmost -P pager=off -c "…"`.
 
-**Real invocation count per tool** (result parts only — the correct denominator):
+**Real invocation count per tool** (outcome parts only — the correct denominator).
+Dual-shape: a v2 outcome has `ok` or `error`; a legacy outcome has `output` or `error`:
 
 ```sql
 SELECT elem->>'toolName' AS tool, count(*) AS calls
 FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
-WHERE jsonb_typeof(m.tool_calls) = 'array' AND elem ? 'output'
+WHERE jsonb_typeof(m.tool_calls) = 'array'
+  AND (elem ? 'ok' OR elem ? 'output' OR elem ? 'error')
 GROUP BY 1 ORDER BY 2 DESC;
 ```
 
-**Soft errors per tool** (everything the DB can honestly see):
+**Soft errors per tool.** The soft-error marker lives in the tool OUTPUT — which on
+**v2 rows is in `metadata.parts`**, on **legacy rows is in the `tool_calls` outcome
+element**. This query UNIONs both eras, projecting each output as `o`:
 
 ```sql
 WITH res AS (
+  -- v2 (#490): output is in metadata.parts (output-available tool parts)
+  SELECT part->>'type' AS tool, part->'output' AS o
+  FROM ai_chat_messages m, jsonb_array_elements(m.metadata->'parts') part
+  WHERE (m.metadata->>'toolTraceVersion') = '2'
+    AND part->>'type' LIKE 'tool-%' AND part->>'state' = 'output-available'
+  UNION ALL
+  -- legacy: output is inline in the tool_calls outcome element
   SELECT elem->>'toolName' AS tool, elem->'output' AS o
   FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
-  WHERE jsonb_typeof(m.tool_calls) = 'array' AND elem ? 'output'
+  WHERE (m.metadata->>'toolTraceVersion') IS NULL
+    AND jsonb_typeof(m.tool_calls) = 'array' AND elem ? 'output'
 )
 SELECT tool, count(*) AS calls,
        sum(COALESCE(
@@ -167,13 +224,23 @@ FROM res GROUP BY tool HAVING sum(COALESCE(
 ORDER BY soft_errors DESC;
 ```
 
-**`editPageText` failure reasons** (the most common real agent mistake — bad `find`):
+Note the v2 `tool` label is the part type (`tool-editPageText`); strip the `tool-`
+prefix if you join it against the legacy `toolName`.
+
+**`editPageText` failure reasons** (the most common real agent mistake — bad `find`).
+Same dual-shape output source:
 
 ```sql
 WITH res AS (
+  SELECT part->'output' AS o
+  FROM ai_chat_messages m, jsonb_array_elements(m.metadata->'parts') part
+  WHERE (m.metadata->>'toolTraceVersion') = '2'
+    AND part->>'type' = 'tool-editPageText' AND part->>'state' = 'output-available'
+  UNION ALL
   SELECT elem->'output' AS o
   FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
-  WHERE jsonb_typeof(m.tool_calls) = 'array'
+  WHERE (m.metadata->>'toolTraceVersion') IS NULL
+    AND jsonb_typeof(m.tool_calls) = 'array'
     AND elem->>'toolName' = 'editPageText' AND elem ? 'output'
 )
 SELECT f->>'reason' AS reason, count(*)
@@ -182,30 +249,43 @@ WHERE jsonb_typeof(o->'failed') = 'array'
 GROUP BY 1 ORDER BY 2 DESC;
 ```
 
-**Hard errors — persisted `error` field per tool (NEW rows, since #407)** — thrown
-tool failures now carry their real reason, so query them directly:
+**Hard errors — persisted `error` field per tool (v2 + post-#407 rows)** — thrown tool
+failures carry their real reason, so query them directly. On **v2** rows exclude the
+`interrupted` kind so an aborted call is not counted as a hard-fail:
 
 ```sql
 SELECT elem->>'toolName' AS tool, count(*) AS thrown_errors,
        min(elem->>'error') AS sample_error
 FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
 WHERE jsonb_typeof(m.tool_calls) = 'array' AND elem ? 'error'
+  -- v2 rows label the kind; a legacy error element has no kind (count it).
+  AND COALESCE(elem->>'kind', 'thrown') = 'thrown'
+GROUP BY 1 ORDER BY 2 DESC;
+```
+
+Aborted mid-step calls on v2 rows are a distinct, directly countable population:
+
+```sql
+SELECT elem->>'toolName' AS tool, count(*) AS interrupted
+FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
+WHERE jsonb_typeof(m.tool_calls) = 'array' AND elem->>'kind' = 'interrupted'
 GROUP BY 1 ORDER BY 2 DESC;
 ```
 
 **Hard-error proxy for OLD rows (pre-#407) — orphan gap per tool, WITH a spread column**
-(call parts minus result parts, plus how many distinct chats the gap is spread across).
-This covers rows written before thrown errors were persisted; on new rows a thrown
-failure now has its own `error` element (use the query above) and an orphan means only
-a genuinely aborted mid-step call:
+(call parts minus outcome parts, plus how many distinct chats the gap is spread across).
+This is needed ONLY for pre-#407 legacy rows (v2 and post-#407 rows carry the error /
+`kind` directly — use the queries above). The `WHERE` restricts to the legacy era so v2
+rows (where an `ok` outcome is not an `output`) never produce phantom orphans:
 
 ```sql
 WITH parts AS (
   SELECT m.chat_id, elem->>'toolName' AS tool,
-         (elem ? 'input' AND NOT (elem ? 'output')) AS is_call,
-         (elem ? 'output' OR elem ? 'error')        AS is_result
+         (elem ? 'input' AND NOT (elem ? 'output') AND NOT (elem ? 'ok')) AS is_call,
+         (elem ? 'output' OR elem ? 'error' OR elem ? 'ok')               AS is_result
   FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
   WHERE jsonb_typeof(m.tool_calls) = 'array' AND m.role = 'assistant'
+    AND (m.metadata->>'toolTraceVersion') IS NULL
 ),
 per_chat AS (
   SELECT tool, chat_id, sum(is_call::int) - sum(is_result::int) AS gap
@@ -261,11 +341,21 @@ WHERE tsv @@ websearch_to_tsquery('english', 'some phrase') LIMIT 20;
 
 ## Don't blow up your context
 
-A single `tool_calls` row can be **300–400 KB** (results embed full page content and
-search payloads). Never `SELECT tool_calls` (or `jsonb_pretty(tool_calls)`) raw.
-Always project just the keys you need and truncate:
+Tool outputs embed full page content and search payloads (hundreds of KB per row).
+On **legacy** rows they are in `tool_calls`; on **v2** rows they moved to
+`metadata->'parts'` (the `tool_calls` trace itself is now small). Never `SELECT
+tool_calls` / `metadata` (or `jsonb_pretty(...)`) raw — project just the keys you need
+and truncate:
 
 ```sql
+-- v2: outputs live in metadata.parts
+SELECT part->>'type',
+       left(regexp_replace((part->'output')::text, '\s+', ' ', 'g'), 200)
+FROM ai_chat_messages m, jsonb_array_elements(m.metadata->'parts') part
+WHERE (m.metadata->>'toolTraceVersion') = '2'
+  AND part->>'state' = 'output-available' LIMIT 5;
+
+-- legacy: outputs live in tool_calls
 SELECT elem->>'toolName',
        left(regexp_replace((elem->'output')::text, '\s+', ' ', 'g'), 200)
 FROM ai_chat_messages m, jsonb_array_elements(m.tool_calls) elem
@@ -280,26 +370,32 @@ docker compose -p gitmost logs -f --tail=100 # whole stack
 ```
 
 Logging is `json-file`, `max-size=10m max-file=5` → ~50 MB retained, then rotated,
-and **wiped on container recreate**. Since the #407 fix, thrown-tool error text is
-**persisted in the `error` field** of `tool_calls` (see the hard-error query above), so
-you no longer depend on live logs for it. Logs/live UI remain useful for **pre-#407
-rows** (whose thrown errors were dropped) and for full stack traces beyond the
-truncated stored message. A per-tool `tool_calls_total{tool,status}` metric to
-VictoriaMetrics is still a possible future add for aggregate dashboards.
+and **wiped on container recreate**. Thrown-tool error text is **persisted** — in the
+`error` field of `tool_calls` (v2 `kind:'thrown'` / post-#407 legacy) — so you no longer
+depend on live logs for it. Logs/live UI remain useful for **pre-#407 rows** (whose
+thrown errors were dropped) and for full stack traces beyond the truncated stored
+message. A per-tool `tool_calls_total{tool,status}` metric to VictoriaMetrics is still a
+possible future add for aggregate dashboards.
 
 ## Gotchas checklist
 
-- [ ] Counting every `tool_calls` element → **overcount**. Count `output` elements; add `error` elements for thrown failures (new rows), but don't count both as invocations.
-- [ ] `isError` / `success=false` ≈ 0 does **not** mean "no errors" — thrown errors are a separate `error` element (new rows) or dropped entirely (pre-#407 rows).
-- [ ] Thrown errors persist only on rows written **after the #407 fix** — pre-#407 rows still drop them (orphan only). Mind the cutover when trending over time.
+- [ ] **Check `metadata.toolTraceVersion` first.** v2 (`= 2`) has no output in `tool_calls`; legacy has it inline. Never trend a metric across the era boundary.
+- [ ] Counting every `tool_calls` element → **overcount**. Count OUTCOME elements — v2: `ok` or `error`; legacy: `output` or `error` — never both call+outcome as invocations.
+- [ ] `isError` / `success=false` ≈ 0 does **not** mean "no errors" — thrown errors are an `error` element (v2 `kind:'thrown'` / post-#407), not in the output.
+- [ ] **v2:** soft-error markers (the tool output) are in `metadata.parts`, NOT `tool_calls`. Legacy: they are in the `tool_calls` outcome `output`.
+- [ ] **v2:** `kind` splits a real hard-fail (`thrown`) from an aborted call (`interrupted`) directly — no orphan heuristic needed. The orphan gap is a pre-#407-legacy-only proxy.
 - [ ] `editPageText.failed` is `[]` on success — test for **non-empty**, not presence.
-- [ ] Orphan gap on OLD rows mixes thrown errors **and** aborted runs — split by tool. On NEW rows a thrown error is its own `error` element, so a gap ≈ aborted call.
 - [ ] `aborted` runs = server restarts, `failed` runs = provider overload — not agent mistakes.
-- [ ] Never dump a raw `tool_calls` cell — it can be hundreds of KB.
-- [ ] Logs are ephemeral (≤50 MB, wiped on recreate) — grab hard-error text live.
+- [ ] Never dump a raw `tool_calls` **or** `metadata.parts` cell — outputs are hundreds of KB.
+- [ ] Logs are ephemeral (≤50 MB, wiped on recreate) — grab pre-#407 hard-error text live.
 
 ## Snapshot (2026-07-07, illustrative — rerun the queries for current numbers)
 
+> All rows in this snapshot predate #490, so they are **legacy-era** (outputs inline in
+> `tool_calls`, orphan proxy for thrown errors). Do not trend these numbers against v2
+> rows — segment by `toolTraceVersion` first.
+
+
 - 226 chats, 732 messages, 46 runs; ~4 400 real tool invocations.
 - Soft errors (persisted): `editPageText` 4/79 (bad/non-unique `find`) + 9 markdown-in-`find` warnings; `semanticSearch` 3/4 (`unavailable`); `Habr_update_draft_from_docmost` 1/2 (`doc` sent as object, not string).
 - Missing-result proxy, read WITH the spread column:

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
-// detectable at runtime.
+// the ENTIRE src/ tree, so a build/ vs src/ skew (issue #447) is detectable at
+// runtime for ANY source file — not just tool-specs.ts.
 //
-// WHY hash the raw source text (not extracted structured data):
-// SHARED_TOOL_SPECS carries `buildShape` functions (the input SCHEMAS) which are
-// NOT serializable. The input schema is exactly one of the things that MUST stay
-// in sync between build/ and src/, so we cannot drop it from the hash. Rather
-// than probe zod with a fragile shim to reconstruct the schema shape, we hash the
-// STABLE, deterministic source TEXT of tool-specs.ts. That text fully captures
-// every field that must stay in sync — mcpName, inAppKey, description, tier,
-// 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.
+// WHY hash the whole src tree (not just tool-specs.ts): the runtime tools are
+// assembled from far more than the spec registry — client.ts, the client/*
+// domain modules, comment-signal.ts and the drawio-* helpers all ship in build/
+// and are loaded by the in-app server. Hashing ONLY tool-specs.ts meant an edit
+// to any of those (e.g. a behavioural fix in client.ts) left the stamp unchanged,
+// so a stale build/ served the OLD code silently (issue #486). Hashing every
+// src/**/*.ts closes that gap: any source edit changes the stamp.
 //
-// DETERMINISM: the hash is computed over the file bytes 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 normalization + sha256 over src/tool-specs.ts and compares to
-// the built REGISTRY_STAMP; the two must compute identically.
+// WHY hash the raw source text (not extracted structured data): the tool input
+// SCHEMAS live as `buildShape` functions which are NOT serializable, so we cannot
+// reduce them to structured data without a fragile zod shim. Hashing the STABLE,
+// deterministic source TEXT captures every field that must stay in sync with zero
+// probing fragility. Any edit to any source file changes the text → the stamp.
+//
+// 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
- * function (exported) so the algorithm has a single home; the loader duplicates
- * only the tiny normalize+sha256 steps because it lives in the CJS server build
- * and cannot import this ESM script. If you change the normalization here, mirror
- * it in apps/server/src/core/ai-chat/tools/docmost-client.loader.ts.
+ * Recursively enumerate every `*.ts` file under `dir`, EXCLUDING the codegen's
+ * own `*.generated.ts` output (a self-referential cycle otherwise). Returns
+ * absolute paths, unsorted (the caller sorts by relative path for determinism).
+ * Kept as a plain exported function so the algorithm has a single home; the
+ * 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) {
-  const normalized = toolSpecsSource.replace(/\r\n/g, '\n').replace(/\n$/, '');
-  return createHash('sha256').update(normalized, 'utf8').digest('hex');
+export function collectStampFiles(dir) {
+  const out = [];
+  for (const entry of readdirSync(dir)) {
+    const full = join(dir, entry);
+    if (statSync(full).isDirectory()) {
+      out.push(...collectStampFiles(full));
+    } else if (entry.endsWith('.ts') && !entry.endsWith('.generated.ts')) {
+      out.push(full);
+    }
+  }
+  return out;
+}
+
+/**
+ * Deterministic stamp of the whole src/ tree. Enumerate + sort by POSIX-relative
+ * path, then fold each file's relative path AND normalized content into one
+ * sha256. MUST stay byte-for-byte identical to the loader's recompute.
+ */
+export function computeRegistryStamp(srcDir) {
+  const files = collectStampFiles(srcDir)
+    .map((abs) => ({
+      rel: relative(srcDir, abs).split(sep).join('/'),
+      abs,
+    }))
+    .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0));
+  const hash = createHash('sha256');
+  for (const { rel, abs } of files) {
+    const normalized = readFileSync(abs, 'utf8')
+      .replace(/\r\n/g, '\n')
+      .replace(/\n$/, '');
+    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(source);
+  const stamp = computeRegistryStamp(SRC_DIR);
   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' +
-    '// tiers, catalog lines and input schemas). Regenerated on every build/pretest\n' +
-    '// so build/ always matches the compiled src. The in-app loader recomputes this\n' +
-    '// from src and refuses to run on a mismatch (issue #447). This file is\n' +
-    '// gitignored and produced by the build — see .gitignore.\n' +
+    '// A deterministic hash of the whole src/ tree (every src/**/*.ts except\n' +
+    '// *.generated.ts). Regenerated on every build/pretest so build/ always\n' +
+    '// matches the compiled src. The in-app loader recomputes this from src and\n' +
+    '// refuses to run on a mismatch (issue #447/#486). This file is gitignored\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/comments.ts

@@ -59,6 +59,39 @@ import {
   mergeFootnoteDefinitions,
 } from "../lib/transforms.js";
 
+// Max concurrent per-page comment fetches in checkNewComments (#490). The scan is
+// O(N) independent REST reads over the working set; running them one-at-a-time made
+// a large space linear in round-trips. A small cap parallelizes without hammering
+// the server (or exhausting sockets). 6 is a conservative middle of the 5–8 band.
+const COMMENT_SCAN_CONCURRENCY = 6;
+
+/**
+ * Map `items` through `fn` with at most `limit` in flight, preserving INPUT ORDER
+ * in the returned array. A tiny bounded pool (no p-limit dependency): `limit`
+ * workers pull the next index off a shared cursor until the list is drained.
+ */
+async function mapWithConcurrency<T, R>(
+  items: readonly T[],
+  limit: number,
+  fn: (item: T, index: number) => Promise<R>,
+): Promise<R[]> {
+  const results = new Array<R>(items.length);
+  let cursor = 0;
+  const worker = async (): Promise<void> => {
+    for (;;) {
+      const i = cursor++;
+      if (i >= items.length) return;
+      results[i] = await fn(items[i], i);
+    }
+  };
+  const workers = Array.from(
+    { length: Math.max(1, Math.min(limit, items.length)) },
+    () => worker(),
+  );
+  await Promise.all(workers);
+  return results;
+}
+
 // Public method surface of CommentsMixin (issue #450) — a NAMED type so the factory
 // return type is expressible in the emitted .d.ts (the anonymous mixin class
 // carries the base's protected shared state, which would otherwise trip TS4094).
@@ -655,27 +688,32 @@ export function CommentsMixin<TBase extends GConstructor<DocmostClientContext>>(
       parentPageId,
     );
 
-    // 2. Fetch comments for each page, keep ones created after since
-    const results: any[] = [];
-    for (const page of pagesInScope) {
-      try {
-        // Full feed (incl. resolved): a "new comments since" scan reports all
-        // recent activity; the active-only filter is scoped to listComments.
-        const comments = (await this.listComments(page.id, true)).items;
-        const newComments = comments.filter(
-          (c: any) => new Date(c.createdAt) > sinceDate,
-        );
-        if (newComments.length > 0) {
-          results.push({
-            pageId: page.id,
-            pageTitle: page.title,
-            comments: newComments,
-          });
+    // 2. Fetch comments for each page, keep ones created after since. Runs with
+    // bounded concurrency (#490) instead of one-at-a-time — the per-page reads are
+    // independent, so a large working set no longer costs O(N) serial round-trips.
+    // Order is preserved (mapWithConcurrency keeps input order), so the output is
+    // deterministic regardless of which fetch finishes first.
+    const perPage = await mapWithConcurrency(
+      pagesInScope,
+      COMMENT_SCAN_CONCURRENCY,
+      async (page: any) => {
+        try {
+          // Full feed (incl. resolved): a "new comments since" scan reports all
+          // recent activity; the active-only filter is scoped to listComments.
+          const comments = (await this.listComments(page.id, true)).items;
+          const newComments = comments.filter(
+            (c: any) => new Date(c.createdAt) > sinceDate,
+          );
+          return newComments.length > 0
+            ? { pageId: page.id, pageTitle: page.title, comments: newComments }
+            : null;
+        } catch (e: any) {
+          // Skip pages with errors (e.g. deleted between calls)
+          return null;
         }
-      } catch (e: any) {
-        // Skip pages with errors (e.g. deleted between calls)
-      }
-    }
+      },
+    );
+    const results: any[] = perPage.filter((r): r is any => r !== null);
 
     const totalNewComments = results.reduce(
       (sum, r) => sum + r.comments.length,

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";
@@ -167,6 +170,43 @@ export abstract class DocmostClientContext {
   // cached conversion can never leak across identities. See getpage-cache.ts.
   protected getPageCache = new GetPageConversionCache();
 
+  // #487: an OPTIONAL abort signal the in-app tool host sets before each tool
+  // call (a composite of the turn's Stop signal + a per-call wall-clock cap). It
+  // is checked at safe-points BETWEEN the sequential HTTP calls of a paginated
+  // read (paginateAll) and just before the atomic collab commit of a write (the
+  // mutatePage/replacePage/mutateLiveContentUnlocked seams), so a Stop / cap
+  // stops the NEXT network call from STARTING. An already-started single call may
+  // still land — a documented limitation (#487).
+  //
+  // SINGLE-WRITER by phase-1 assumption: exactly one DocmostClient is built per
+  // turn and shared by every tool call; the host sets this per call and does NOT
+  // restore the prior value on unwind (set-and-leave) — a fresh client per turn
+  // plus overwrite-by-the-next-call keeps it correct, and leaving a settled
+  // call's signal in place is what makes a discarded race-loser throw on its
+  // next safe-point. If the model emits PARALLEL in-app
+  // tool calls they share this one field, so the per-call CAP of one call is not
+  // guaranteed to bound another's in-flight pagination — but every composite the
+  // host sets carries the SAME turn Stop signal, so a Stop still aborts whichever
+  // signal is current. #487.
+  protected toolAbortSignal: AbortSignal | null = null;
+
+  /**
+   * #487: set (or clear with null) the in-app tool abort signal governing the
+   * NEXT client call's safe-points. The host wraps each in-app tool call: it sets
+   * the composite (Stop + per-call cap) here before invoking the tool and leaves
+   * it in place afterwards (set-and-leave, NOT restored) — the next call
+   * overwrites it, and a fresh client is built per turn. Public so the
+   * server-side tool wrapper can reach it; harmless (a no-op) when never set.
+   */
+  public setToolAbortSignal(signal: AbortSignal | null): void {
+    this.toolAbortSignal = signal;
+  }
+
+  /** #487: the abort signal currently governing this client's safe-points. */
+  public getToolAbortSignal(): AbortSignal | null {
+    return this.toolAbortSignal;
+  }
+
   // Two construction forms:
   //  - new DocmostClient(config)                  // discriminated union (current)
   //  - new DocmostClient(baseURL, email, password) // legacy positional creds
@@ -492,6 +532,37 @@ 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 —
@@ -526,19 +597,28 @@ 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, {
-      // Only the actual 25s collab connect timeout emits this — the connect-vs-
-      // unload signal; the other failure paths must NOT emit it.
-      onConnectTimeout: () =>
-        this.onMetricFn?.("collab_connect_timeouts_total", 1),
+    // 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: () =>
+          this.onMetricFn?.("collab_connect_timeouts_total", 1),
+      });
+      try {
+        // #487 PRE-COMMIT safe-point (reentrant twin of mutatePageContent): a
+        // Stop/cap after acquiring the session but before the atomic write skips
+        // this commit. Same limitation applies (stops the NEXT commit only).
+        this.toolAbortSignal?.throwIfAborted();
+        return await session.mutate(transform);
+      } catch (e) {
+        // Drop the session on any failure so the next call reconnects fresh.
+        session.destroy("mutate failed");
+        throw e;
+      }
     });
-    try {
-      return await session.mutate(transform);
-    } catch (e) {
-      // Drop the session on any failure so the next call reconnects fresh.
-      session.destroy("mutate failed");
-      throw e;
-    }
   }
 
   /**
@@ -563,6 +643,11 @@ export abstract class DocmostClientContext {
     let truncated = false;
 
     for (let page = 0; page < MAX_PAGES; page++) {
+      // #487 safe-point: a Stop (or the in-app tool per-call cap) that fires
+      // BETWEEN sequential page fetches must stop the NEXT request from starting
+      // — a read tool that would otherwise paginate for minutes is interrupted
+      // here. throwIfAborted() rejects with the signal's reason.
+      this.toolAbortSignal?.throwIfAborted();
       const payload: Record<string, any> = {
         ...basePayload,
         limit: clampedLimit,
@@ -667,7 +752,19 @@ 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) =>
+      // #487: thread the in-app tool signal to mutatePageContent's pre-commit
+      // safe-point so a Stop/cap during the connect/lock window skips the write.
+      mutatePageContent(
+        pageUuid,
+        token,
+        apiUrl,
+        transform,
+        this.toolAbortSignal ?? undefined,
+      ),
+    );
   }
 
   /**
@@ -687,7 +784,18 @@ 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) =>
+      // #487: same pre-commit safe-point as mutatePage, for full-document writes.
+      replacePageContent(
+        pageUuid,
+        doc,
+        token,
+        apiUrl,
+        this.toolAbortSignal ?? undefined,
+      ),
+    );
   }
 
   /**

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);
-      return buildPageTree(pages);
+      // #486: propagate `truncated` (same pattern as check_new_comments). The old
+      // 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";
@@ -31,6 +30,13 @@ export { destroyAllSessions } from "./lib/collab-session.js";
 // internals directly; it goes through loadDocmostMcp()).
 export { SHARED_TOOL_SPECS } from "./tool-specs.js";
 export type { SharedToolSpec } from "./tool-specs.js";
+// #489 — write-class registry consumed by the in-app external-MCP retry gate.
+export {
+  SHARED_TOOL_WRITE_CLASS,
+  isRetryableWriteClass,
+  assertEverySpecDeclaresWriteClass,
+} from "./tool-specs.js";
+export type { ToolWriteClass } from "./tool-specs.js";
 
 // Re-export the build-time REGISTRY_STAMP (issue #447): a deterministic hash of
 // the tool-specs registry content, generated into src/registry-stamp.generated.ts

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(
-            new Error("Authentication failed for collaboration connection"),
-            true,
-          );
+          // Tag the error so the client can tell a REJECTED collab token apart
+          // from a generic disconnect and invalidate + refresh it (#486).
+          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/collaboration.ts

@@ -254,6 +254,12 @@ export async function mutatePageContent(
   collabToken: string,
   baseUrl: string,
   transform: (liveDoc: any) => any | null,
+  // #487: optional abort signal carrying the turn's Stop + the in-app tool
+  // per-call cap. Checked as the PRE-COMMIT safe-point below (after the session
+  // is acquired, immediately before the atomic read->write), so a Stop that
+  // arrives during the connect/lock window stops THIS write from landing. See the
+  // limitation note at the check.
+  signal?: AbortSignal,
 ): Promise<MutationResult> {
   return withPageLock(pageId, async () => {
     if (process.env.DEBUG) {
@@ -266,6 +272,13 @@ export async function mutatePageContent(
 
     const session = await acquireCollabSession(pageId, collabToken, baseUrl);
     try {
+      // #487 PRE-COMMIT safe-point: if the turn was Stopped (or the in-app tool
+      // per-call cap fired) after we acquired the collab session but before the
+      // atomic write, throw NOW so this commit never runs. KNOWN LIMITATION
+      // (#487): this only stops THIS commit — a write tool that already committed
+      // an EARLIER call this turn leaves that op applied. Cancel guarantees "no
+      // NEW commit starts", NOT "the write didn't land".
+      signal?.throwIfAborted();
       return await session.mutate(transform);
     } catch (e) {
       // Drop the session on any failure so the next call reconnects fresh (this
@@ -291,6 +304,8 @@ export async function replacePageContent(
   prosemirrorDoc: any,
   collabToken: string,
   baseUrl: string,
+  // #487: threaded straight to mutatePageContent's pre-commit safe-point.
+  signal?: AbortSignal,
 ): Promise<MutationResult> {
   // Fail fast on a bad document instead of deferring the failure into the
   // collaboration write (where TiptapTransformer.toYdoc(undefined) used to
@@ -307,6 +322,7 @@ export async function replacePageContent(
     collabToken,
     baseUrl,
     () => prosemirrorDoc,
+    signal,
   );
 }
 

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
-// (layout:"elk" in drawioCreate/drawioUpdate) and elkjs runs synchronously on
-// the MCP server's event loop, so an unbounded graph would block it for
-// seconds-to-minutes. 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
-// wall-clock timeout; on either bound we fall back to the ORIGINAL model, the
-// same best-effort contract the catch already honours.
+// DoS bounds for the ELK layout. The mxGraph XML is LLM-supplied (layout:"elk"
+// in drawioCreate/drawioUpdate). elkjs' layout() returns a Promise but runs the
+// crossing-minimisation SYNCHRONOUSLY — it blocks whatever thread it runs on for
+// the whole pass. A ~1MB XML (well under the stage-1 16MB cap) can carry
+// thousands of nodes. We (a) cap the graph size before ever calling ELK and
+// (b) run the layout in a WORKER THREAD so the main event loop stays free, with
+// 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
-//     pathological input can never wedge the server.
+//   - The timeout is a HARD kill of the worker thread — the only way to interrupt
+//     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
- * (normalized) model is returned unchanged — layout is best-effort polish, never
- * a reason to fail the write.
+ * the layout runs on a worker thread. On any layout failure (including a
+ * terminate-on-timeout) the ORIGINAL (normalized) model is returned unchanged —
+ * 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
-    // module systems; resolve the real constructor at runtime, then cast (the
-    // runtime call is verified — see the layout unit test).
-    const Ctor: any = (ELK as any).default ?? ELK;
-    const elk = new Ctor();
-    // 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;
+    // Run the (synchronous-under-the-hood) ELK pass on a worker thread so the
+    // main event loop is never blocked, and enforce the wall-clock ceiling by
+    // terminating that worker on timeout. A graph under the node/edge caps but
+    // still pathologically slow is hard-killed instead of wedging anything.
+    laid = await layoutInWorker(graph, resolveElkTimeoutMs());
   } 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

@@ -127,6 +127,18 @@ export interface SharedToolSpec {
   mcpName: string;
   /** camelCase key in the ai-SDK tools object (the in-app layer). */
   inAppKey: string;
+  /**
+   * Write-class of the tool (#489), declared on EVERY spec (a registration-time
+   * assert enforces completeness; `satisfies Record<string, SharedToolSpec>`
+   * makes it a compile error to omit). 'readOnly' = a pure read that mutates
+   * NOTHING durable, so it is safe to auto-retry once after a transport break.
+   * 'write' = anything that mutates a page/comment/share/diagram/etc — a
+   * transport error is INDETERMINATE (the server may have applied it before the
+   * connection reset), so it is NEVER blind-retried (a retry would double-apply,
+   * the #435 incident class). Consumed by the external-MCP retry path
+   * (mcp-clients.service.ts) to gate its single auto-retry.
+   */
+  writeClass: 'readOnly' | 'write';
   /** Single canonical model-facing description used by both layers. */
   description: string;
   /**
@@ -240,6 +252,7 @@ export const SHARED_TOOL_SPECS = {
   getWorkspace: {
     mcpName: 'getWorkspace',
     inAppKey: 'getWorkspace',
+    writeClass: 'readOnly',
     description: 'Fetch metadata about the current workspace (name, settings).',
     tier: 'core',
     catalogLine: 'getWorkspace — fetch current workspace metadata (name, settings).',
@@ -249,6 +262,7 @@ export const SHARED_TOOL_SPECS = {
   listSpaces: {
     mcpName: 'listSpaces',
     inAppKey: 'listSpaces',
+    writeClass: 'readOnly',
     description:
       'List the spaces the current user can access. Returns the array of ' +
       'spaces (id, name, slug, ...).',
@@ -260,6 +274,7 @@ export const SHARED_TOOL_SPECS = {
   listShares: {
     mcpName: 'listShares',
     inAppKey: 'listShares',
+    writeClass: 'readOnly',
     description:
       'List all public shares in the workspace with page titles and public URLs.',
     tier: 'deferred',
@@ -272,6 +287,7 @@ export const SHARED_TOOL_SPECS = {
   getPageJson: {
     mcpName: 'getPageJson',
     inAppKey: 'getPageJson',
+    writeClass: 'readOnly',
     description:
       'Get page details with the raw ProseMirror JSON content (lossless: ' +
       'includes block ids, callouts, tables, link/image attributes) plus the ' +
@@ -289,6 +305,7 @@ export const SHARED_TOOL_SPECS = {
   getOutline: {
     mcpName: 'getOutline',
     inAppKey: 'getOutline',
+    writeClass: 'readOnly',
     description:
       "Return a COMPACT outline of a page's top-level blocks ({index, type, " +
       'id, level, firstText}; tables add rows/cols/header; lists add item ' +
@@ -309,6 +326,7 @@ export const SHARED_TOOL_SPECS = {
   getNode: {
     mcpName: 'getNode',
     inAppKey: 'getNode',
+    writeClass: 'readOnly',
     description:
       "Fetch a single block for editing. `nodeId` is a block id from the page " +
       'outline or page-JSON view (works for headings/paragraphs/callouts/images), OR ' +
@@ -350,6 +368,7 @@ export const SHARED_TOOL_SPECS = {
   searchInPage: {
     mcpName: 'searchInPage',
     inAppKey: 'searchInPage',
+    writeClass: 'readOnly',
     description:
       'Find every occurrence of a string (or regex) INSIDE one page and get ' +
       'WHERE each is — instead of pulling blocks one-by-one with getNode. ' +
@@ -413,6 +432,7 @@ export const SHARED_TOOL_SPECS = {
   deleteNode: {
     mcpName: 'deleteNode',
     inAppKey: 'deleteNode',
+    writeClass: 'write',
     description:
       'Remove a single block by its attrs.id (from the page outline or ' +
       'page-JSON view) WITHOUT resending the whole document.',
@@ -438,6 +458,7 @@ export const SHARED_TOOL_SPECS = {
   patchNode: {
     mcpName: 'patchNode',
     inAppKey: 'patchNode',
+    writeClass: 'write',
     description:
       'Replace a single content block identified by its attrs.id, WITHOUT ' +
       'resending the whole document; the replacement keeps the same block id. ' +
@@ -505,6 +526,7 @@ export const SHARED_TOOL_SPECS = {
   insertNode: {
     mcpName: 'insertNode',
     inAppKey: 'insertNode',
+    writeClass: 'write',
     description:
       'Insert content before/after another block (by attrs.id or anchor text) ' +
       'or append it at the end (top level). For before/after you MUST provide ' +
@@ -597,6 +619,7 @@ export const SHARED_TOOL_SPECS = {
   sharePage: {
     mcpName: 'sharePage',
     inAppKey: 'sharePage',
+    writeClass: 'write',
     // CANONICAL: merges the MCP copy's URL-format + idempotency detail with the
     // in-app copy's reversibility note; keeps the security framing both had.
     description:
@@ -626,6 +649,7 @@ export const SHARED_TOOL_SPECS = {
   unsharePage: {
     mcpName: 'unsharePage',
     inAppKey: 'unsharePage',
+    writeClass: 'write',
     description: 'Remove the public share of a page (revokes the public URL).',
     tier: 'deferred',
     catalogLine: "unsharePage — revoke a page's public share (removes the public URL).",
@@ -640,6 +664,7 @@ export const SHARED_TOOL_SPECS = {
   diffPageVersions: {
     mcpName: 'diffPageVersions',
     inAppKey: 'diffPageVersions',
+    writeClass: 'readOnly',
     description:
       'Diff two versions of a page and return a Docmost-equivalent change set ' +
       '(inserted/deleted text, integrity counts for images/links/tables/' +
@@ -672,6 +697,7 @@ export const SHARED_TOOL_SPECS = {
   listPageHistory: {
     mcpName: 'listPageHistory',
     inAppKey: 'listPageHistory',
+    writeClass: 'readOnly',
     description:
       "List a page's saved versions (Docmost auto-snapshots on every save), " +
       'newest first, cursor-paginated. Returns { items, nextCursor }; each ' +
@@ -693,6 +719,7 @@ export const SHARED_TOOL_SPECS = {
   restorePageVersion: {
     mcpName: 'restorePageVersion',
     inAppKey: 'restorePageVersion',
+    writeClass: 'write',
     description:
       'Restore a page to a saved version: writes that version\'s content back ' +
       'as the page\'s current content (Docmost has no restore endpoint, so ' +
@@ -713,6 +740,7 @@ export const SHARED_TOOL_SPECS = {
   importPageMarkdown: {
     mcpName: 'importPageMarkdown',
     inAppKey: 'importPageMarkdown',
+    writeClass: 'write',
     // IN-APP ONLY (issue #411): the external /mcp surface no longer exposes
     // importPageMarkdown — the registry loop in index.ts skips inAppOnly specs,
     // so this stays available to the in-app agent (round-tripping an EXPORTED
@@ -742,6 +770,7 @@ export const SHARED_TOOL_SPECS = {
   copyPageContent: {
     mcpName: 'copyPageContent',
     inAppKey: 'copyPageContent',
+    writeClass: 'write',
     description:
       "Replace targetPageId's content with a copy of sourcePageId's content, " +
       'entirely server-side — the document is NOT sent through the model. The ' +
@@ -770,6 +799,7 @@ export const SHARED_TOOL_SPECS = {
   editPageText: {
     mcpName: 'editPageText',
     inAppKey: 'editPageText',
+    writeClass: 'write',
     description:
       "Surgical find/replace inside a page's text, preserving all block " +
       'ids and marks. A find MAY cross bold/italic/link boundaries; the ' +
@@ -819,6 +849,7 @@ export const SHARED_TOOL_SPECS = {
   stashPage: {
     mcpName: 'stashPage',
     inAppKey: 'stashPage',
+    writeClass: 'readOnly',
     description:
       'Serialize a whole page (the full ProseMirror JSON, as getPageJson ' +
       'returns) into an ephemeral in-memory blob and return ONLY a short ' +
@@ -880,6 +911,7 @@ export const SHARED_TOOL_SPECS = {
   getPage: {
     mcpName: 'getPage',
     inAppKey: 'getPage',
+    writeClass: 'readOnly',
     description:
       'Fetch a single page as Markdown by its id. Returns the page title and ' +
       'its Markdown content. The converter is canonical (round-trips text and ' +
@@ -919,13 +951,15 @@ export const SHARED_TOOL_SPECS = {
   listPages: {
     mcpName: 'listPages',
     inAppKey: 'listPages',
+    writeClass: 'readOnly',
     description:
       'List the most recent pages (ordered by updatedAt, descending), ' +
       'optionally scoped to a single space. Returns a bounded list (default ' +
       '50, max 100) — use search for lookups in large spaces. tree:true (with ' +
-      "spaceId) returns the space's full page hierarchy as a nested tree, but " +
-      'is DEPRECATED — use getTree instead (leaner nodes, plus rootPageId / ' +
-      'maxDepth).',
+      "spaceId) returns { tree, truncated } — the space's full page hierarchy " +
+      'as a nested tree, plus a `truncated` flag that is true when the tree was ' +
+      '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).",
@@ -964,6 +998,7 @@ export const SHARED_TOOL_SPECS = {
   getTree: {
     mcpName: 'getTree',
     inAppKey: 'getTree',
+    writeClass: 'readOnly',
     description:
       "Get a space's page hierarchy (or one subtree) as a nested tree in a " +
       'SINGLE request — completely and without loss. Each node is ' +
@@ -1008,6 +1043,7 @@ export const SHARED_TOOL_SPECS = {
   getPageContext: {
     mcpName: 'getPageContext',
     inAppKey: 'getPageContext',
+    writeClass: 'readOnly',
     description:
       'Given a pageId, get its LOCATION and immediate surroundings (metadata ' +
       'only, no page content) in one call — answers "where am I / what is ' +
@@ -1037,6 +1073,7 @@ export const SHARED_TOOL_SPECS = {
   createPage: {
     mcpName: 'createPage',
     inAppKey: 'createPage',
+    writeClass: 'write',
     description:
       'Create a new page with a Markdown body in a space, optionally under a ' +
       'parent page (omit parentPageId to create at the space root). Returns ' +
@@ -1087,6 +1124,7 @@ export const SHARED_TOOL_SPECS = {
   movePage: {
     mcpName: 'movePage',
     inAppKey: 'movePage',
+    writeClass: 'write',
     description:
       'Move a page under a new parent page, or to the space root when no ' +
       'parent is given. Reversible: move it back at any time.',
@@ -1180,6 +1218,7 @@ export const SHARED_TOOL_SPECS = {
   renamePage: {
     mcpName: 'renamePage',
     inAppKey: 'renamePage',
+    writeClass: 'write',
     description:
       'Rename a page (change its title only; the body is untouched, never ' +
       'resent). Reversible: rename back at any time.',
@@ -1201,6 +1240,7 @@ export const SHARED_TOOL_SPECS = {
   deletePage: {
     mcpName: 'deletePage',
     inAppKey: 'deletePage',
+    writeClass: 'write',
     description:
       'Move a page to the trash — SOFT delete only: the page can be restored ' +
       'from trash and nothing is ever permanently deleted.',
@@ -1234,6 +1274,7 @@ export const SHARED_TOOL_SPECS = {
   updatePageJson: {
     mcpName: 'updatePageJson',
     inAppKey: 'updatePageJson',
+    writeClass: 'write',
     description:
       "Replace a page's content with a raw ProseMirror JSON document (lossless " +
       'write: preserves the block ids, callouts, tables and attributes you pass ' +
@@ -1290,6 +1331,7 @@ export const SHARED_TOOL_SPECS = {
   updatePageMarkdown: {
     mcpName: 'updatePageMarkdown',
     inAppKey: 'updatePageMarkdown',
+    writeClass: 'write',
     description:
       "Replace a page's body with new Markdown content (and optionally its " +
       'title). The whole body is re-imported from the markdown (block ids ' +
@@ -1324,6 +1366,7 @@ export const SHARED_TOOL_SPECS = {
   exportPageMarkdown: {
     mcpName: 'exportPageMarkdown',
     inAppKey: 'exportPageMarkdown',
+    writeClass: 'readOnly',
     // CANONICAL: the MCP copy (a strict superset of the terse in-app wording).
     description:
       'Export a page to a single self-contained Docmost-flavoured Markdown ' +
@@ -1372,6 +1415,7 @@ export const SHARED_TOOL_SPECS = {
   createComment: {
     mcpName: 'createComment',
     inAppKey: 'createComment',
+    writeClass: 'write',
     // CANONICAL: the in-app copy (the more-maintained one). It keeps the same
     // rules as the MCP copy — inline-only, top-level requires a `selection`, no
     // page-level comments, replies inherit the anchor, suggestedText must be
@@ -1504,6 +1548,7 @@ export const SHARED_TOOL_SPECS = {
   listComments: {
     mcpName: 'listComments',
     inAppKey: 'listComments',
+    writeClass: 'readOnly',
     // CANONICAL: the two copies are near-identical; the MCP copy is the
     // superset (it keeps the "(pagination is handled internally)" note the
     // in-app copy dropped), so it is used verbatim.
@@ -1531,6 +1576,7 @@ export const SHARED_TOOL_SPECS = {
   resolveComment: {
     mcpName: 'resolveComment',
     inAppKey: 'resolveComment',
+    writeClass: 'write',
     // CANONICAL: the MCP copy's richer wording, minus its reference
     // to `deleteComment` (a sibling tool that does NOT exist in the in-app
     // layer) — rephrased transport-neutrally per the registry convention.
@@ -1571,6 +1617,7 @@ export const SHARED_TOOL_SPECS = {
   checkNewComments: {
     mcpName: 'checkNewComments',
     inAppKey: 'checkNewComments',
+    writeClass: 'readOnly',
     // CANONICAL: the MCP copy (the more detailed of the two). The MCP layer's
     // execute-side guard that rejects an unparseable `since` timestamp stays in
     // its execute body (per-layer logic), not in the shared schema.
@@ -1650,6 +1697,7 @@ export const SHARED_TOOL_SPECS = {
   tableInsertRow: {
     mcpName: 'tableInsertRow',
     inAppKey: 'tableInsertRow',
+    writeClass: 'write',
     description:
       'Insert a row of plain-text cells into a table. `table` is `#<index>` ' +
       'from the page outline, or a block id inside it. `cells` is the text per ' +
@@ -1683,6 +1731,7 @@ export const SHARED_TOOL_SPECS = {
   tableDeleteRow: {
     mcpName: 'tableDeleteRow',
     inAppKey: 'tableDeleteRow',
+    writeClass: 'write',
     description:
       'Delete the row at 0-based `index` from a table (`table` is `#<index>` ' +
       'from the page outline, or a block id inside it). Refuses to delete the ' +
@@ -1706,6 +1755,7 @@ export const SHARED_TOOL_SPECS = {
   tableUpdateCell: {
     mcpName: 'tableUpdateCell',
     inAppKey: 'tableUpdateCell',
+    writeClass: 'write',
     description:
       'Set the plain-text content of cell [row, col] (0-based) in a table ' +
       '(`table` is `#<index>` from the page outline, or a block id inside it). ' +
@@ -1746,6 +1796,7 @@ export const SHARED_TOOL_SPECS = {
   insertFootnote: {
     mcpName: 'insertFootnote',
     inAppKey: 'insertFootnote',
+    writeClass: 'write',
     description:
       'Insert an AUTHOR-INLINE footnote: you specify only WHERE (anchorText) ' +
       'and WHAT (text). The footnote marker is placed right after anchorText in ' +
@@ -1783,6 +1834,7 @@ export const SHARED_TOOL_SPECS = {
   insertImage: {
     mcpName: 'insertImage',
     inAppKey: 'insertImage',
+    writeClass: 'write',
     description:
       'Download an image from a web (http/https) URL and insert it into ' +
       'a page in one step. By default ' +
@@ -1827,6 +1879,7 @@ export const SHARED_TOOL_SPECS = {
   replaceImage: {
     mcpName: 'replaceImage',
     inAppKey: 'replaceImage',
+    writeClass: 'write',
     description:
       'Replace an existing image on a page with a new image fetched from a web ' +
       '(http/https) URL: uploads the new file as a NEW ' +
@@ -1865,6 +1918,7 @@ export const SHARED_TOOL_SPECS = {
   drawioGet: {
     mcpName: 'drawioGet',
     inAppKey: 'drawioGet',
+    writeClass: 'readOnly',
     description:
       'Read a draw.io diagram on a page as mxGraph XML (default) or as its raw ' +
       '`.drawio.svg`. `node` is the drawio node\'s attrs.id (from getOutline / ' +
@@ -1898,6 +1952,7 @@ export const SHARED_TOOL_SPECS = {
   drawioCreate: {
     mcpName: 'drawioCreate',
     inAppKey: 'drawioCreate',
+    writeClass: 'write',
     description:
       'Create a draw.io diagram from mxGraph XML and insert it as a diagram ' +
       'block. `xml` is a bare `<mxGraphModel>` OR a list of `<mxCell>` elements ' +
@@ -1968,6 +2023,7 @@ export const SHARED_TOOL_SPECS = {
   drawioUpdate: {
     mcpName: 'drawioUpdate',
     inAppKey: 'drawioUpdate',
+    writeClass: 'write',
     description:
       'Replace a draw.io diagram\'s content with new mxGraph XML (same lint ' +
       'pipeline as drawioCreate). `baseHash` is MANDATORY: pass the hash from ' +
@@ -2019,6 +2075,7 @@ export const SHARED_TOOL_SPECS = {
   drawioEditCells: {
     mcpName: 'drawioEditCells',
     inAppKey: 'drawioEditCells',
+    writeClass: 'write',
     description:
       'Make TARGETED, id-based edits to an existing draw.io diagram instead of ' +
       'resending the whole XML (a full-XML diff is fragile — draw.io reorders ' +
@@ -2077,6 +2134,7 @@ export const SHARED_TOOL_SPECS = {
   drawioFromGraph: {
     mcpName: 'drawioFromGraph',
     inAppKey: 'drawioFromGraph',
+    writeClass: 'write',
     description:
       'Build a draw.io diagram from a SEMANTIC graph — you describe nodes, groups ' +
       'and edges by MEANING and the server picks every coordinate, color and icon ' +
@@ -2201,6 +2259,7 @@ export const SHARED_TOOL_SPECS = {
   drawioFromMermaid: {
     mcpName: 'drawioFromMermaid',
     inAppKey: 'drawioFromMermaid',
+    writeClass: 'write',
     description:
       'Convert Mermaid `flowchart` text into an EDITABLE draw.io diagram (LLMs ' +
       'write Mermaid reliably). Best for STANDARD flowcharts/decision trees: ' +
@@ -2247,6 +2306,7 @@ export const SHARED_TOOL_SPECS = {
   drawioShapes: {
     mcpName: 'drawioShapes',
     inAppKey: 'drawioShapes',
+    writeClass: 'readOnly',
     description:
       'Look up VERIFIED draw.io stencil style-strings so you never guess a ' +
       '`shape=mxgraph.*` name (a wrong name renders as an EMPTY BOX). Searches a ' +
@@ -2293,6 +2353,7 @@ export const SHARED_TOOL_SPECS = {
   drawioGuide: {
     mcpName: 'drawioGuide',
     inAppKey: 'drawioGuide',
+    writeClass: 'readOnly',
     description:
       'Progressive-disclosure draw.io authoring reference. Call with a `section` ' +
       'to pull one focused, <=4KB chapter instead of bloating context: ' +
@@ -2322,3 +2383,53 @@ export const SHARED_TOOL_SPECS = {
     inlineBothHosts: true,
   },
 } satisfies Record<string, SharedToolSpec>;
+
+// --- write-class registry (#489) ------------------------------------------
+
+/** A tool's retry-safety class. 'readOnly' may be auto-retried once after a
+ *  transport break; 'write' is indeterminate and must never be blind-retried. */
+export type ToolWriteClass = 'readOnly' | 'write';
+
+/**
+ * Name → write-class map for the shared registry, keyed by mcpName (=== inAppKey).
+ * The external-MCP retry path (mcp-clients.service.ts) looks a tool up here by its
+ * RAW (un-namespaced) name to decide whether a transport failure may be retried.
+ * A tool NOT in this map (a third-party external MCP tool) is treated as 'write'
+ * by the consumer — the safe default (never blind-retry an unknown tool).
+ */
+export const SHARED_TOOL_WRITE_CLASS: Record<string, ToolWriteClass> =
+  Object.fromEntries(
+    Object.values(SHARED_TOOL_SPECS).map((spec) => [spec.mcpName, spec.writeClass]),
+  );
+
+/** Whether a write-class permits a single automatic retry after a transport
+ *  break. Only a pure read is retry-safe; everything mutating is indeterminate. */
+export function isRetryableWriteClass(
+  writeClass: ToolWriteClass | undefined,
+): boolean {
+  return writeClass === 'readOnly';
+}
+
+/**
+ * Registration-time assert (#489): EVERY spec must declare a valid write-class.
+ * `satisfies Record<string, SharedToolSpec>` already makes an omission a compile
+ * error, but this guards a raw/cast construction path and documents the invariant
+ * at the point of use. Runs once on import — both hosts import this module, so
+ * both get the check. Throws (fails startup) rather than silently mis-gating a
+ * retry in production.
+ */
+export function assertEverySpecDeclaresWriteClass(): void {
+  for (const [key, spec] of Object.entries(SHARED_TOOL_SPECS)) {
+    const wc = (spec as SharedToolSpec).writeClass;
+    if (wc !== 'readOnly' && wc !== 'write') {
+      throw new Error(
+        `tool-specs: spec "${key}" must declare writeClass ('readOnly' | 'write'), got ${JSON.stringify(
+          wc,
+        )}`,
+      );
+    }
+  }
+}
+
+// Enforce at module load (registration time) on both hosts.
+assertEverySpecDeclaresWriteClass();

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/paginate-abort-safepoint.test.mjs

@@ -0,0 +1,143 @@
+// #487 commit 1 — the in-app tool cancellation safe-point inside paginateAll.
+//
+// The in-app tool host sets a composite abort signal on the client
+// (setToolAbortSignal) before each tool call; paginateAll checks it at a
+// safe-point BEFORE every sequential page fetch, so a Stop that lands mid-read
+// stops the NEXT HTTP request from STARTING (a read tool can no longer paginate
+// for minutes past a Stop). This pins the HONEST observable property against the
+// REAL client + a real HTTP server: "after Stop, no NEW request starts".
+import { test, after } from "node:test";
+import assert from "node:assert/strict";
+import http from "node:http";
+import { DocmostClient } from "../../build/client.js";
+
+function readBody(req) {
+  return new Promise((resolve) => {
+    let raw = "";
+    req.on("data", (c) => (raw += c));
+    req.on("end", () => resolve(raw));
+  });
+}
+function sendJson(res, status, obj, extra = {}) {
+  res.writeHead(status, { "Content-Type": "application/json", ...extra });
+  res.end(JSON.stringify(obj));
+}
+const openServers = [];
+async function spawn(handler) {
+  const server = await new Promise((resolve) => {
+    const s = http.createServer(handler);
+    s.listen(0, "127.0.0.1", () => resolve(s));
+  });
+  openServers.push(server);
+  const { port } = server.address();
+  return { baseURL: `http://127.0.0.1:${port}/api` };
+}
+after(async () => {
+  await Promise.all(openServers.map((s) => new Promise((r) => s.close(r))));
+});
+function handleLogin(req, res) {
+  if (req.url === "/api/auth/login") {
+    sendJson(res, 200, { success: true }, {
+      "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+    });
+    return true;
+  }
+  return false;
+}
+
+// A Stop that lands DURING pagination: the server aborts the client signal as it
+// serves page 1 (more pages remain). The loop's next safe-point must throw before
+// the page-2 request is sent.
+test("paginateAll stops the NEXT request when the signal aborts mid-pagination", async () => {
+  let requests = 0;
+  const ac = new AbortController();
+  const { baseURL } = await spawn(async (req, res) => {
+    await readBody(req);
+    if (handleLogin(req, res)) return;
+    if (req.url === "/api/spaces") {
+      requests++;
+      // Simulate a user Stop that lands while page 1 is in flight.
+      if (requests === 1) ac.abort(new Error("user stop"));
+      sendJson(res, 200, {
+        success: true,
+        data: {
+          items: [{ id: `p${requests}` }],
+          meta: { hasNextPage: true, nextCursor: `c${requests}` },
+        },
+      });
+      return;
+    }
+    sendJson(res, 404, {});
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  client.setToolAbortSignal(ac.signal);
+
+  await assert.rejects(
+    () => client.paginateAll("/spaces", {}),
+    /user stop/,
+    "the aborted safe-point rejects with the signal's reason",
+  );
+  assert.equal(requests, 1, "page 2 never started after the Stop");
+});
+
+// A Stop that is already in effect before the read starts: zero requests fire.
+test("paginateAll starts no request when the signal is already aborted", async () => {
+  let requests = 0;
+  const { baseURL } = await spawn(async (req, res) => {
+    await readBody(req);
+    if (handleLogin(req, res)) return;
+    if (req.url === "/api/spaces") {
+      requests++;
+      sendJson(res, 200, {
+        success: true,
+        data: { items: [], meta: { hasNextPage: false, nextCursor: null } },
+      });
+      return;
+    }
+    sendJson(res, 404, {});
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  // Warm the auth so ensureAuthenticated does not itself POST after the abort.
+  await client.ensureAuthenticated();
+  const ac = new AbortController();
+  ac.abort(new Error("already stopped"));
+  client.setToolAbortSignal(ac.signal);
+
+  await assert.rejects(() => client.paginateAll("/spaces", {}), /already stopped/);
+  assert.equal(requests, 0, "no /spaces request started once already aborted");
+});
+
+// Without a tool signal (default), pagination is unaffected — the safe-point is a
+// pure no-op, so pre-#487 behaviour is byte-identical.
+test("paginateAll is unaffected when no tool signal is set", async () => {
+  let requests = 0;
+  const PAGES = {
+    "": { items: [{ id: "a" }], nextCursor: "c1" },
+    c1: { items: [{ id: "b" }], nextCursor: null },
+  };
+  const { baseURL } = await spawn(async (req, res) => {
+    const raw = await readBody(req);
+    if (handleLogin(req, res)) return;
+    if (req.url === "/api/spaces") {
+      requests++;
+      const body = JSON.parse(raw || "{}");
+      const page = PAGES[body.cursor ?? ""] ?? { items: [], nextCursor: null };
+      sendJson(res, 200, {
+        success: true,
+        data: {
+          items: page.items,
+          meta: { hasNextPage: page.nextCursor != null, nextCursor: page.nextCursor },
+        },
+      });
+      return;
+    }
+    sendJson(res, 404, {});
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  const all = await client.paginateAll("/spaces", {});
+  assert.equal(requests, 2, "both pages fetched with no signal set");
+  assert.deepEqual(all.map((p) => p.id), ["a", "b"]);
+});

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.
-  const tree = await client.listPages("space-1", 50, true);
+  // listPages tree:true -> enumerateSpacePages(spaceId) -> { tree, truncated }.
+  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");
 });
@@ -438,3 +442,71 @@ test("checkNewComments subtree includes the root without a separate getPageRaw",
   assert.equal(result.checkedPages, 2, "root + one descendant scanned");
   assert.equal(result.totalNewComments, 1, "the root's fresh comment found");
 });
+
+// -----------------------------------------------------------------------------
+// 6) checkNewComments parallelism (#490): the per-page comment fetches run with
+//    bounded concurrency (not one-at-a-time), and the results still preserve the
+//    page order deterministically regardless of which fetch finishes first.
+// -----------------------------------------------------------------------------
+test("checkNewComments fetches pages concurrently (bounded) and preserves order", async () => {
+  // A subtree with 12 descendants so the scan has plenty to parallelize.
+  const NODES = [{ id: "parent", title: "Parent", parentPageId: null, hasChildren: true }];
+  for (let i = 0; i < 12; i++) {
+    NODES.push({ id: `k${i}`, title: `Kid ${i}`, parentPageId: "parent", hasChildren: false });
+  }
+
+  let inFlight = 0;
+  let maxInFlight = 0;
+
+  const { baseURL } = await spawn(async (req, res) => {
+    const raw = await readBody(req);
+    if (handleLogin(req, res)) return;
+    if (req.url === "/api/pages/tree") {
+      sendJson(res, 200, { success: true, data: { items: NODES } });
+      return;
+    }
+    if (req.url === "/api/comments") {
+      const body = JSON.parse(raw || "{}");
+      inFlight++;
+      maxInFlight = Math.max(maxInFlight, inFlight);
+      // Hold the response briefly so concurrent fetches actually overlap.
+      setTimeout(() => {
+        inFlight--;
+        // Every page carries one fresh comment so ordering is observable.
+        sendJson(res, 200, {
+          success: true,
+          data: {
+            items: [
+              { id: `c-${body.pageId}`, createdAt: "2030-01-01T00:00:00.000Z", content: null },
+            ],
+            meta: { nextCursor: null },
+          },
+        });
+      }, 25);
+      return;
+    }
+    sendJson(res, 404, {});
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  const result = await client.checkNewComments(
+    "space-1",
+    "2020-01-01T00:00:00.000Z",
+    "parent",
+  );
+
+  // 13 pages (parent + 12 kids) were scanned; each had a fresh comment.
+  assert.equal(result.checkedPages, 13, "all pages scanned");
+  assert.equal(result.totalNewComments, 13, "one fresh comment per page");
+  // Parallelism: more than one request was in flight at once, but never above the
+  // cap (6). A serial implementation would show maxInFlight === 1.
+  assert.ok(maxInFlight > 1, `expected concurrent fetches, saw max ${maxInFlight}`);
+  assert.ok(maxInFlight <= 6, `concurrency must be bounded, saw ${maxInFlight}`);
+  // Deterministic order: results follow the page-enumeration order (parent first).
+  assert.equal(result.comments[0].pageId, "parent", "results preserve page order");
+  assert.deepEqual(
+    result.comments.map((r) => r.pageId),
+    ["parent", ...Array.from({ length: 12 }, (_, i) => `k${i}`)],
+    "result order matches the enumeration order regardless of finish order",
+  );
+});

packages/mcp/test/mock/write-abort-safepoint.test.mjs

@@ -0,0 +1,164 @@
+// #487 F4 — the WRITE-side cancellation safe-point.
+//
+// Every content-mutating collab write (collaboration.mutatePageContent and the
+// reentrant twin client.mutateLiveContentUnlocked used by replaceImage) checks
+// the in-app tool abort signal at a PRE-COMMIT safe-point — after the collab
+// session is acquired but immediately BEFORE the atomic read->write
+// (session.mutate). So a Stop (or the per-call cap) that lands during the
+// connect/lock window stops THIS write from landing: no new commit starts once
+// aborted. paginate-abort-safepoint.test.mjs pins the READ half; this pins the
+// integrity-critical WRITE half — remove the `throwIfAborted()` and the transform
+// would run and the doc would be mutated past a Stop.
+//
+// There is no collab server in the unit env, so we swap the provider factory
+// (__setCollabProviderFactory) for a fake that reports an immediate successful
+// sync. That makes acquireCollabSession SUCCEED and hand back a live, ready
+// session, so the ONLY thing standing between the call and session.mutate is the
+// safe-point under test. The transform is instrumented to prove it never runs.
+import { test, afterEach } from "node:test";
+import assert from "node:assert/strict";
+import { mutatePageContent } from "../../build/lib/collaboration.js";
+import {
+  __setCollabProviderFactory,
+  destroyAllSessions,
+} from "../../build/lib/collab-session.js";
+import { DocmostClient } from "../../build/client.js";
+
+const BASE_URL = "http://127.0.0.1:1/api";
+// mutatePageContent locks via withPageLock, which demands a canonical page UUID
+// (resolve-then-lock invariant, #260/#449); the unlocked twin does not.
+const PAGE_UUID = "11111111-1111-4111-8111-111111111111";
+
+// A fake HocuspocusProvider that immediately reports a successful initial sync so
+// CollabSession.open() resolves to a ready session, and stays "synced" with zero
+// unsynced changes so a reached session.mutate() would resolve at once. It speaks
+// only the tiny CollabProviderLike surface the session depends on.
+function syncedProviderFactory(config) {
+  // Fire the initial-sync callback so open() settles as ready.
+  config.onSynced();
+  return {
+    synced: true,
+    unsyncedChanges: 0,
+    destroy() {},
+    on() {},
+    off() {},
+  };
+}
+
+// Disable the session cache so every acquire opens (and the failure path destroys)
+// its own ephemeral session — no cross-test session reuse.
+process.env.MCP_COLLAB_SESSION_IDLE_MS = "0";
+
+afterEach(() => {
+  __setCollabProviderFactory(null); // restore the real factory
+  destroyAllSessions();
+});
+
+// --- collaboration.mutatePageContent (the page-locked write path) -------------
+
+test("mutatePageContent rejects at the pre-commit safe-point BEFORE session.mutate when the signal is already aborted", async () => {
+  __setCollabProviderFactory(syncedProviderFactory);
+  let transformCalls = 0;
+  const ac = new AbortController();
+  ac.abort(new Error("user stop"));
+
+  await assert.rejects(
+    () =>
+      mutatePageContent(
+        PAGE_UUID,
+        "collab-jwt",
+        BASE_URL,
+        (liveDoc) => {
+          transformCalls++;
+          return liveDoc;
+        },
+        ac.signal,
+      ),
+    /user stop/,
+    "the aborted safe-point rejects with the signal's reason before committing",
+  );
+  assert.equal(
+    transformCalls,
+    0,
+    "the transform (and therefore session.mutate) must NEVER run once aborted",
+  );
+});
+
+test("mutatePageContent (control) DOES reach session.mutate and invoke the transform when the signal is live", async () => {
+  __setCollabProviderFactory(syncedProviderFactory);
+  let transformCalls = 0;
+  const ac = new AbortController(); // never aborted
+
+  const result = await mutatePageContent(
+    PAGE_UUID,
+    "collab-jwt",
+    BASE_URL,
+    (liveDoc) => {
+      transformCalls++;
+      return null; // null -> no-op write; still proves the transform was invoked
+    },
+    ac.signal,
+  );
+  assert.equal(
+    transformCalls,
+    1,
+    "with a live signal the safe-point is a no-op and session.mutate runs the transform",
+  );
+  assert.ok(result && result.verify, "a MutationResult is returned");
+});
+
+// --- client.mutateLiveContentUnlocked (the reentrant twin, replaceImage) ------
+
+test("mutateLiveContentUnlocked rejects at the pre-commit safe-point BEFORE session.mutate when the tool signal is already aborted", async () => {
+  __setCollabProviderFactory(syncedProviderFactory);
+  const client = new DocmostClient({
+    apiUrl: BASE_URL,
+    getToken: async () => "access",
+    getCollabToken: async () => "collab-jwt",
+  });
+  let transformCalls = 0;
+  const ac = new AbortController();
+  ac.abort(new Error("cap fired"));
+  client.setToolAbortSignal(ac.signal);
+
+  await assert.rejects(
+    () =>
+      client.mutateLiveContentUnlocked("page-1", "collab-jwt", (liveDoc) => {
+        transformCalls++;
+        return liveDoc;
+      }),
+    /cap fired/,
+    "the aborted safe-point rejects with the signal's reason before committing",
+  );
+  assert.equal(
+    transformCalls,
+    0,
+    "the transform (and therefore session.mutate) must NEVER run once aborted",
+  );
+});
+
+test("mutateLiveContentUnlocked (control) DOES reach session.mutate and invoke the transform when the tool signal is live", async () => {
+  __setCollabProviderFactory(syncedProviderFactory);
+  const client = new DocmostClient({
+    apiUrl: BASE_URL,
+    getToken: async () => "access",
+    getCollabToken: async () => "collab-jwt",
+  });
+  let transformCalls = 0;
+  client.setToolAbortSignal(new AbortController().signal); // live
+
+  const result = await client.mutateLiveContentUnlocked(
+    "page-1",
+    "collab-jwt",
+    (liveDoc) => {
+      transformCalls++;
+      return null;
+    },
+  );
+  assert.equal(
+    transformCalls,
+    1,
+    "with a live signal the safe-point is a no-op and session.mutate runs the transform",
+  );
+  assert.ok(result && result.verify, "a MutationResult is returned");
+});

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 { readFileSync } from "node:fs";
+import {
+  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
-// exports `computeRegistryStamp(sourceText)` — a sha256 over normalized source
-// text (CRLF->LF, single trailing newline stripped). The in-app loader
-// (apps/server/.../docmost-client.loader.ts) DUPLICATES that normalize+sha256 to
-// recompute the stamp from src and refuse a stale build. These tests pin the
-// algorithm's behaviour AND assert the built stamp matches the current src, so a
-// stale generated file OR a normalize divergence reddens.
+// Guard tests for the build/src-skew stamp (issues #447/#486). The codegen script
+// exports `computeRegistryStamp(srcDir)` — a sha256 over the WHOLE src/ tree
+// (every src/**/*.ts EXCEPT *.generated.ts), each file folded in as its
+// POSIX-relative path + its normalized content (CRLF->LF, single trailing newline
+// stripped). Hashing the whole tree (not just tool-specs.ts) is #486: an edit to
+// client.ts / a client/* module without a rebuild must ALSO redden. The in-app
+// 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", () => {
-  const input = "export const X = 1;\nexport const Y = 2;\n";
-  assert.equal(computeRegistryStamp(input), computeRegistryStamp(input));
+// Build a throwaway src/ tree from a { relPath: content } map and return its dir.
+function makeSrcTree(files) {
+  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");
-  assert.match(stamp, /^[0-9a-f]{64}$/);
+  const t = makeSrcTree({ "tool-specs.ts": "anything\n" });
+  try {
+    assert.match(computeRegistryStamp(t.src), /^[0-9a-f]{64}$/);
+  } finally {
+    t.cleanup();
+  }
 });
 
-test("normalizes CRLF vs LF: the same content hashes equal", () => {
-  const lf = "line1\nline2\nline3";
-  const crlf = "line1\r\nline2\r\nline3";
-  assert.equal(computeRegistryStamp(crlf), computeRegistryStamp(lf));
+// #486 CORE: an edit to a NON-tool-specs source file (client.ts) must change the
+// stamp. Under the old single-file (tool-specs.ts only) hash this edit was
+// invisible and a stale build/ served the old client.ts silently.
+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({
+    "tool-specs.ts": "export const SPECS = 1;\n",
+    "client.ts": "export const impl = 'v2';\n",
+  });
+  try {
+    assert.notEqual(
+      computeRegistryStamp(before.src),
+      computeRegistryStamp(after.src),
+      "a client.ts edit with an unchanged tool-specs.ts must move the stamp",
+    );
+  } finally {
+    before.cleanup();
+    after.cleanup();
+  }
 });
 
-test("normalizes a trailing newline: with/without a final \\n hashes equal", () => {
-  const noTrailing = "alpha\nbeta";
-  const trailing = "alpha\nbeta\n";
-  assert.equal(computeRegistryStamp(trailing), computeRegistryStamp(noTrailing));
+test("editing a nested client/* module changes the stamp", () => {
+  const before = makeSrcTree({
+    "tool-specs.ts": "x\n",
+    "client/read.ts": "export const READ = 1;\n",
+  });
+  const after = makeSrcTree({
+    "tool-specs.ts": "x\n",
+    "client/read.ts": "export const READ = 2;\n",
+  });
+  try {
+    assert.notEqual(
+      computeRegistryStamp(before.src),
+      computeRegistryStamp(after.src),
+    );
+  } finally {
+    before.cleanup();
+    after.cleanup();
+  }
 });
 
-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),
-  );
+// *.generated.ts is EXCLUDED (else the codegen's own output is a fixed-point
+// cycle): adding/removing/changing it must not move the stamp.
+test("*.generated.ts is excluded from the stamp", () => {
+  const without = makeSrcTree({ "tool-specs.ts": "x\n" });
+  const withGen = makeSrcTree({
+    "tool-specs.ts": "x\n",
+    "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();
+  }
 });
 
-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));
+test("a CRLF checkout WITH trailing CRLF hashes equal to bare LF", () => {
+  const bare = makeSrcTree({ "tool-specs.ts": "alpha\nbeta" });
+  const crlfTrailing = makeSrcTree({ "tool-specs.ts": "alpha\r\nbeta\r\n" });
+  try {
+    assert.equal(
+      computeRegistryStamp(crlfTrailing.src),
+      computeRegistryStamp(bare.src),
+    );
+  } finally {
+    bare.cleanup();
+    crlfTrailing.cleanup();
+  }
 });
 
-// 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.
+// Only a SINGLE trailing newline is stripped — a second blank line is content.
 test("only ONE trailing newline is stripped (two differ from one)", () => {
-  assert.notEqual(
-    computeRegistryStamp("x\n"),
-    computeRegistryStamp("x\n\n"),
-  );
+  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 input. The SAME literal input
-// and expected hash are asserted in the server-side jest test
-// (docmost-client.loader.spec.ts). If either side's normalize+sha256 ever
-// diverges, one of the two tests reddens. Input exercises BOTH normalize steps.
-test("fixed-input hash matches the documented cross-impl value", () => {
-  const FIXED_INPUT = "line1\r\nline2\n";
-  const EXPECTED =
-    "683376e290829b482c2655745caffa7a1dccfa10afaa62dac2b42dd6c68d0f83";
-  assert.equal(computeRegistryStamp(FIXED_INPUT), EXPECTED);
+// 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();
+  }
 });
 
-// DESYNC GUARD (covers reviewer suggestion 2). Recompute the stamp from the
-// actual src/tool-specs.ts 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 normalize ever diverges from what produced the built stamp.
-test("built REGISTRY_STAMP equals the stamp recomputed from src/tool-specs.ts", () => {
-  const source = readFileSync(TOOL_SPECS_PATH, "utf8");
-  assert.equal(computeRegistryStamp(source), REGISTRY_STAMP);
+// 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();
+  }
 });
 
-// Sanity: the fixed-input helper computes the SAME way the codegen does, proving
-// the EXPECTED constant above is not an arbitrary magic value but the documented
-// normalize+sha256 of FIXED_INPUT. Belt-and-braces so a bad EXPECTED can't hide a
-// real regression.
-test("the documented EXPECTED constant is the normalize+sha256 of FIXED_INPUT", () => {
-  const FIXED_INPUT = "line1\r\nline2\n";
-  const normalized = FIXED_INPUT.replace(/\r\n/g, "\n").replace(/\n$/, "");
-  const expected = createHash("sha256")
-    .update(normalized, "utf8")
-    .digest("hex");
-  assert.equal(computeRegistryStamp(FIXED_INPUT), expected);
+// 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);
 });

packages/mcp/test/unit/tool-specs.test.mjs

@@ -2,7 +2,12 @@ import { test } from "node:test";
 import assert from "node:assert/strict";
 import { z } from "zod";
 
-import { SHARED_TOOL_SPECS } from "../../build/tool-specs.js";
+import {
+  SHARED_TOOL_SPECS,
+  SHARED_TOOL_WRITE_CLASS,
+  isRetryableWriteClass,
+  assertEverySpecDeclaresWriteClass,
+} from "../../build/tool-specs.js";
 
 // The shared registry is consumed by BOTH the zod-v3 MCP server and the zod-v4
 // in-app AI-SDK service, so every spec must carry the cross-layer wiring
@@ -43,6 +48,41 @@ test("mcpName and inAppKey are each unique across the registry", () => {
   }
 });
 
+// #489 — every spec must declare its write-class so the external-MCP retry path
+// can gate a single auto-retry ONLY on a pure read (a blind retry of a write =
+// double-apply). The declaration is enforced at registration time.
+test("#489: every spec declares a valid writeClass ('readOnly' | 'write')", () => {
+  for (const [key, spec] of Object.entries(SHARED_TOOL_SPECS)) {
+    assert.ok(
+      spec.writeClass === "readOnly" || spec.writeClass === "write",
+      `${key}: missing/invalid writeClass: ${JSON.stringify(spec.writeClass)}`,
+    );
+  }
+  // The registration-time assert must not throw for the shipped registry.
+  assert.doesNotThrow(() => assertEverySpecDeclaresWriteClass());
+});
+
+test("#489: SHARED_TOOL_WRITE_CLASS maps every mcpName to its class; helper gates on readOnly", () => {
+  const specs = Object.values(SHARED_TOOL_SPECS);
+  assert.equal(Object.keys(SHARED_TOOL_WRITE_CLASS).length, specs.length);
+  for (const spec of specs) {
+    assert.equal(SHARED_TOOL_WRITE_CLASS[spec.mcpName], spec.writeClass);
+  }
+  // Only a readOnly tool is retry-eligible; a write tool and an unknown tool are not.
+  assert.equal(isRetryableWriteClass("readOnly"), true);
+  assert.equal(isRetryableWriteClass("write"), false);
+  assert.equal(isRetryableWriteClass(undefined), false);
+});
+
+test("#489: representative reads are readOnly and representative writes are write", () => {
+  for (const name of ["getPage", "getTree", "searchInPage", "listComments"]) {
+    assert.equal(SHARED_TOOL_SPECS[name].writeClass, "readOnly", `${name} should be readOnly`);
+  }
+  for (const name of ["patchNode", "createPage", "deletePage", "createComment", "drawioCreate"]) {
+    assert.equal(SHARED_TOOL_SPECS[name].writeClass, "write", `${name} should be write`);
+  }
+});
+
 test("buildShape (when present) returns a usable ZodRawShape with a real zod", () => {
   for (const [key, spec] of Object.entries(SHARED_TOOL_SPECS)) {
     if (!spec.buildShape) continue;

packages/token-estimate/package.json

@@ -0,0 +1,19 @@
+{
+  "name": "@docmost/token-estimate",
+  "version": "0.1.0",
+  "description": "Shared, provider-agnostic token estimator (chars/2.5) used by the AI-chat client counter and the server history-replay budgeter, so the two never diverge.",
+  "private": true,
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "scripts": {
+    "build": "tsc",
+    "watch": "tsc --watch",
+    "test": "vitest run",
+    "test:watch": "vitest"
+  },
+  "license": "MIT",
+  "devDependencies": {
+    "typescript": "^5.0.0",
+    "vitest": "4.1.6"
+  }
+}

packages/token-estimate/src/index.test.ts

@@ -0,0 +1,31 @@
+import { describe, it, expect } from 'vitest';
+import { estimateTokens, CHARS_PER_TOKEN } from './index';
+
+describe('estimateTokens (shared chars/2.5)', () => {
+  it('returns 0 for empty / nullish input', () => {
+    expect(estimateTokens('')).toBe(0);
+    expect(estimateTokens(null)).toBe(0);
+    expect(estimateTokens(undefined)).toBe(0);
+  });
+
+  it('uses the chars/2.5 ratio, ceiled', () => {
+    expect(CHARS_PER_TOKEN).toBe(2.5);
+    // 5 chars / 2.5 = 2
+    expect(estimateTokens('abcde')).toBe(2);
+    // any non-empty string is at least 1 token (ceil)
+    expect(estimateTokens('a')).toBe(1);
+    // 100 chars / 2.5 = 40
+    expect(estimateTokens('x'.repeat(100))).toBe(40);
+  });
+
+  it('counts Cyrillic ~2x higher than the old chars/4 rule (no undercount)', () => {
+    const cyr = 'привет мир как дела'; // 19 chars
+    expect(estimateTokens(cyr)).toBe(Math.ceil(19 / 2.5)); // 8
+    expect(estimateTokens(cyr)).toBeGreaterThan(Math.ceil(19 / 4)); // > 5
+  });
+
+  it('is deterministic / byte-stable (same input => same output)', () => {
+    const s = 'the quick brown fox';
+    expect(estimateTokens(s)).toBe(estimateTokens(s));
+  });
+});

packages/token-estimate/src/index.ts

@@ -0,0 +1,35 @@
+/**
+ * Shared, provider-agnostic token estimator (#490).
+ *
+ * No provider exposes an exact tokenizer we can afford to run on the hot path (a
+ * real BPE pass is O(n²)-ish, bloats the client bundle, and is wrong for
+ * Gemini/Ollama anyway), so both the client's in-body counter AND the server's
+ * history-replay budgeter use this ONE cheap chars-based heuristic. Keeping it in
+ * a single shared module is deliberate: two independent estimators drift, and then
+ * "the badge shows 60%" while "the budgeter already trimmed" — the exact confusion
+ * this package prevents.
+ *
+ * Ratio: **chars / 2.5**. Most content here is Cyrillic, where a token is ~2.5
+ * characters; the common English `chars/4` rule of thumb UNDER-counts Cyrillic by
+ * ~2×, which for a budget check is the dangerous direction (it lets the context
+ * overflow). 2.5 slightly over-estimates pure English/code, which is the SAFE
+ * direction for a budget. This is an estimate, never an exact count — the
+ * authoritative figure is always the provider's reported usage; the estimate is
+ * for UI affordances, the delta of not-yet-sent messages, and deciding what to
+ * trim.
+ */
+
+/** Characters per token for the shared estimate. See the module comment. */
+export const CHARS_PER_TOKEN = 2.5;
+
+/**
+ * Rough token estimate for a piece of text (chars / {@link CHARS_PER_TOKEN}).
+ * Returns 0 for empty/nullish input, and ceils so any non-empty text counts as at
+ * least one token. Pure and deterministic (byte-stable), so the same text always
+ * yields the same estimate — which the server budgeter relies on to keep replay
+ * trimming stable turn to turn (provider prompt-cache friendliness).
+ */
+export function estimateTokens(text: string | null | undefined): number {
+  if (!text) return 0;
+  return Math.ceil(text.length / CHARS_PER_TOKEN);
+}

packages/token-estimate/tsconfig.json

@@ -0,0 +1,16 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "CommonJS",
+    "moduleResolution": "Node",
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "declaration": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["src/**/*.test.ts"]
+}

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
@@ -287,6 +287,9 @@ importers:
       '@docmost/prosemirror-markdown':
         specifier: workspace:*
         version: link:../../packages/prosemirror-markdown
+      '@docmost/token-estimate':
+        specifier: workspace:*
+        version: link:../../packages/token-estimate
       '@excalidraw/excalidraw':
         specifier: 0.18.0-3a5ef40
         version: 0.18.0-3a5ef40(@types/react-dom@18.3.1)(@types/react@18.3.12)(react-dom@18.3.1(react@18.3.1))(react@18.3.1)
@@ -561,6 +564,9 @@ importers:
       '@docmost/prosemirror-markdown':
         specifier: workspace:*
         version: link:../../packages/prosemirror-markdown
+      '@docmost/token-estimate':
+        specifier: workspace:*
+        version: link:../../packages/token-estimate
       '@fastify/compress':
         specifier: ^9.0.0
         version: 9.0.0
@@ -644,10 +650,10 @@ 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
@@ -1148,6 +1154,15 @@ importers:
         specifier: 4.1.6
         version: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@20.19.43)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@25.0.0)(vite@8.0.5(@types/node@20.19.43)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
 
+  packages/token-estimate:
+    devDependencies:
+      typescript:
+        specifier: ^5.0.0
+        version: 5.9.3
+      vitest:
+        specifier: 4.1.6
+        version: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@25.5.0)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@27.4.0(@noble/hashes@2.0.1))(vite@8.0.5(@types/node@25.5.0)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
+
 packages:
 
   '@aashutoshrathi/word-wrap@1.2.6':
@@ -16455,17 +16470,17 @@ 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