fix(ai-chat): ре-ревью #518 — CI-фиделити delta, attach-дубль, чистки (#491)

Ребейз на обновлённый feat/490 (PR #510, Option-A): стек #491 перенесён с
6e872f2d на 6e42752a; run-fsm.ts/chat-thread.tsx приехали из develop-версии
(W1/S4: RUN_ALREADY_ACTIVE несёт activeRunId + supersede-CAS адопция) — мои
#491-правки легли сверху без конфликтов, W1 НЕ откачен (run-fsm.ts == develop
байт-в-байт; RUN_ALREADY_ACTIVE диспатчит activeRunId).

1. [CI-fidelity] Дельта-курсор спек падал в CI unit-лейне: `.spec.ts` дефолтил
   на НЕмигрированный docmost → 5/6 ERROR relation does not exist (skip-гард ловит
   только connection-fail). Переименован в `*.int-spec.ts` (исключается из unit-
   regex `.spec.ts$`, гоняется в test:int, чей global-setup мигрирует docmost_test)
   + DSN по умолчанию → docmost_test. Теперь 6/6 реально исполняются в CI-верном
   окружении; overlap-мутация роняет RACE-тесты (не вакуумен).

2. [regression #137/#161] attach: отсутствие `n` схлопывалось в frontier 0 →
   finished-неротированный ран (coverageFloor 0) отдавал ВЕСЬ tail вместо 204;
   парамслесс/легаси-вкладка допишет полный replay → дубль. Различаем ОТСУТСТВИЕ
   `n` (null — не tail-aware) от `n=0` (tail-aware): контроллер шлёт null при
   missing/invalid; registry.attach(n: number|null) 204-ит finished-ран при
   n===null (старый `finished && !expectLive` гейт), n=0 по-прежнему отдаёт хвост.
   Тесты (registry + controller) + mutation-verify: нейтрализация гейта роняет их.

4. [conventions] Ring-кап env-var переименован RUN_STREAM_MAX_BUFFER_BYTES →
   AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES (префикс как у сиблингов) + запись в
   .env.example (дефолт 4MB, 0/invalid→дефолт, subscriber-cap=2×).

5. [docs] run-fsm.spec.md: item4 переписан («реализовано в #491 — дельта несёт
   run:{id,status}|null; клиент run-поле ещё не потребляет») + добавлена строка
   перехода POLL_IDLE_CAP stopping→idle (Review #4, редьюсер это делает).

6. [simplification] Удалена мёртвая цепочка reconstructRunParts /
   reconstructPartsFromRow (ноль прод-вызовов) + опц. messageRepo-инъекция в
   AiChatRunService + спек-блоки; вернётся с первым реальным вызывателем. Маркер
   metadata.stepsPersisted (реально используемый) сохранён.

DROP-пункты ревьюера (осиротевший import, DELTA_POLL_MAX_ROWS) не трогаю.
Прогон: server tsc 0, ai-chat unit 202, delta-int 6/6 (int-lane), int attach 6/6,
client vitest 403, tsc client 0 ai-chat, mcp 834/0. FSM-инварианты #488 сохранены.

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.
@@ -287,6 +302,39 @@ MCP_DOCMOST_PASSWORD=
 # enabled for a workspace, and the same single-instance constraint applies (the
 # registry is process-local).
 # AI_CHAT_RESUMABLE_STREAM=false
+#
+# Per-run replay ring cap (#491), in BYTES, for the resumable-stream registry
+# above. The registry buffers the run's recent SSE tail so a reopened tab can
+# attach and continue from the step it already persisted; the ring is bounded and
+# rotates on every confirmed step-persist. This caps the un-persisted tail between
+# rotations — an overflow evicts the oldest frames and a late attach falls back to
+# 204 -> degraded poll, so correctness never depends on the size. Default 4194304
+# (4MB); a 0/invalid value falls back to the default. The per-subscriber backpressure
+# cap is derived as 2x this value. Only meaningful with AI_CHAT_RESUMABLE_STREAM on.
+# AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES=4194304
+
+# --- 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).

.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:

CHANGELOG.md

@@ -202,6 +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
@@ -282,6 +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
@@ -314,6 +336,44 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- **A long AI chat no longer bricks on the model's context window, and each turn
+  stops re-persisting the whole tool-output history.** Tool outputs are now
+  stored ONCE, in `metadata.parts`; the `tool_calls` trace keeps only per-step
+  outcome flags (a v2 trace shape), ending the O(N²) write amplification that
+  re-wrote every prior output on every step (measured on a live Postgres via the
+  `pg_current_wal_lsn()` delta: the trace column shrank ~3200×, the full
+  assistant row ~51%). The persisted record is unchanged in content — the full
+  history still lives in `metadata.parts`. At REPLAY time only, the history sent
+  to the provider is now bounded by a deterministic, prompt-cache-friendly token
+  budget: `floor(0.7 × chatContextWindow)` when a window is configured (no cap —
+  anti-brick protection, not a cost limiter), a flat 100k fallback for installs
+  with no window set (exactly the ones that hit terminal overflow), or off when
+  the window is explicitly `0`. Trimming truncates old tool outputs first, then
+  mechanically collapses the oldest turns, always keeping the recent turns full
+  and the tool-call/result pairing balanced. A provider context-overflow 400 is
+  now classified and used as a reactive signal: the row is stamped so the NEXT
+  turn re-trims aggressively (0.5×), which un-bricks a chat that just 400'd. The
+  client token badge and the server budgeter now share one estimator (new
+  `@docmost/token-estimate` package) so they can never diverge. Deferred-tool
+  activation is also cached in the chat metadata to avoid re-resolving it each
+  turn. (#490)
+- **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

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/ai-chat-window.tsx

@@ -58,8 +58,11 @@ import ConversationList from "@/features/ai-chat/components/conversation-list.ts
 import ChatThread from "@/features/ai-chat/components/chat-thread.tsx";
 import {
   exportAiChat,
+  getAiChatMessagesDelta,
   stopRun,
 } from "@/features/ai-chat/services/ai-chat-service.ts";
+import { mergeDeltaRowsIntoPages } from "@/features/ai-chat/utils/resume-helpers.ts";
+import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.ts";
 import { useChatSession } from "@/features/ai-chat/hooks/use-chat-session.ts";
 import {
   shouldCollapseOnOutsidePointer,
@@ -86,19 +89,11 @@ const MIN_HEIGHT = 400;
 // Margin kept between the window and the viewport edges while dragging.
 const EDGE_MARGIN = 8;
 
-// #184 phase 1.5 / #430: backstop for the degraded-poll fallback. The poll is
-// armed when a resume attempt could not attach to the live run and disarmed by the
-// thread on settle / local stream; this cap is the ONLY backstop against an endless
-// tick (a stuck 'streaming' row before the boot-sweep, or a user-tail 204 with no
-// run).
-//
-// #430: measured from RUN ACTIVITY, not from arm-time. A real autonomous run takes
-// 11-25 min — longer than a fixed 10-min-from-start cap, which used to cut the poll
-// off mid-run. Instead we cap on INACTIVITY: keep polling as long as the run is
-// still making progress (its persisted rows keep changing), and only give up after
-// this long with NO new activity. A genuinely stuck run produces no row changes, so
-// the idle cap still bounds it; a long-but-progressing run polls to completion.
-const DEGRADED_POLL_IDLE_MAX_MS = 10 * 60_000;
+// #184 phase 1.5 / #430 / #488: the degraded-poll fallback. The window owns only
+// a DUMB 2.5s timer, gated by an armed flag; the THREAD's run-lifecycle FSM owns
+// arm/disarm AND the inactivity cap that turns a stuck run into a `stalled` banner
+// (#488 commit 4a — the cap moved into the thread so polling->stalled is a single
+// FSM transition; the window no longer silently stops polling at the cap).
 
 /** Compact token formatter: 1.2M / 3.4k / 950. */
 function formatTokens(n: number): string {
@@ -259,17 +254,13 @@ export default function AiChatWindow() {
     [roles],
   );
 
-  // #184 phase 1.5: degraded-poll fallback (replaces the F4/F5/F7 latches). When
-  // ChatThread could not attach to a still-running run it arms this via
-  // onResumeFallback(true); the thread disarms it on settle / local stream. The
-  // window only OWNS the timer (armedAtRef stamps when it was armed for the cap).
+  // #184 phase 1.5 / #488: degraded-poll fallback. ChatThread's FSM arms this via
+  // onResumeFallback(true) when it enters a poll-bearing recovery (attach 204 /
+  // starved finish / stop) and disarms it on settle / local stream / stalled. The
+  // window owns ONLY the dumb 2.5s timer; the THREAD owns arm/disarm AND the
+  // inactivity cap (a stuck run -> the thread's `stalled` banner disarms this).
   const [degradedPoll, setDegradedPoll] = useState(false);
-  // #430: timestamp of the LAST run activity while the poll is armed — stamped on
-  // arm and re-stamped whenever the polled rows change (see the effect below). The
-  // idle cap is measured from this, so a long-but-progressing run keeps polling.
-  const lastActivityAtRef = useRef(0);
   const onResumeFallback = useCallback((active: boolean): void => {
-    if (active) lastActivityAtRef.current = Date.now();
     setDegradedPoll(active);
   }, []);
   // Reset the degraded poll whenever the open chat changes: it is scoped to the
@@ -281,32 +272,63 @@ export default function AiChatWindow() {
   const { data: messageRows, isLoading: messagesLoading } =
     useAiChatMessagesQuery(
       activeChatId ?? undefined,
-      // DELIBERATELY DUMB (invariant 8 / task 2.4): poll every 2.5s while armed
-      // and while the run is still active (#430: under the INACTIVITY cap, not a
-      // fixed-from-start cap); otherwise off. NO error checks (TanStack v5 resets
-      // fetchFailureCount each fetch, so consecutive errors are not expressible —
-      // and the poll must survive a server restart) and NO tail checks (the
-      // settled/local-stream semantics live in ChatThread, which disarms via
-      // onResumeFallback(false)). The idle cap is the only backstop.
-      () =>
-        degradedPoll === true &&
-        Date.now() - lastActivityAtRef.current < DEGRADED_POLL_IDLE_MAX_MS
-          ? 2500
-          : false,
-      // #344: gate on windowOpen too — no message history is fetched (and no
-      // degraded poll runs) while the window is closed; it loads when the window
-      // opens with an active chat.
+      // #491: the full infinite-query no longer POLLS. It seeds the thread ONCE; the
+      // degraded fallback now runs a DELTA poller (below) that augments THIS cache
+      // idempotently, instead of refetching every page (with full parts) every 2.5s.
+      false,
+      // #344: gate on windowOpen too — no message history is fetched while the window
+      // is closed; it loads when the window opens with an active chat.
       windowOpen,
     );
 
-  // #430: re-stamp the activity clock whenever the polled rows change while the
-  // poll is armed. TanStack keeps the same `messageRows` reference across refetches
-  // that return deep-equal data (structural sharing), so a new reference means the
-  // run genuinely progressed — which extends the inactivity cap above. A stuck run
-  // yields no reference change, so the cap eventually fires and stops the poll.
+  // #491 degraded DELTA poll. While armed (degradedPoll) and the window is open on a
+  // chat, poll POST /ai-chat/messages/delta every 2.5s: it returns only the rows
+  // CHANGED since the previous cursor (+ the run fact) in ONE round-trip. We merge
+  // those rows into the SAME infinite-query cache the thread reads (idempotently by
+  // id — the delta's overlap window re-delivers rows), so the thread's reconcile
+  // effect follows the detached run to its terminal row from a fraction of the wire
+  // cost. The run-fact settle stays the thread FSM's job (row-status reconcile), so
+  // we do NOT double-poll /run here. Cursor resets when the chat changes / disarms.
+  const deltaCursorRef = useRef<string | undefined>(undefined);
   useEffect(() => {
-    if (degradedPoll) lastActivityAtRef.current = Date.now();
-  }, [degradedPoll, messageRows]);
+    deltaCursorRef.current = undefined;
+  }, [activeChatId, degradedPoll]);
+  useEffect(() => {
+    if (!degradedPoll || !windowOpen || !activeChatId) return;
+    const chatId = activeChatId;
+    let cancelled = false;
+    const tick = async (): Promise<void> => {
+      try {
+        const res = await getAiChatMessagesDelta(chatId, deltaCursorRef.current);
+        if (cancelled) return;
+        deltaCursorRef.current = res.cursor;
+        if (res.rows.length > 0) {
+          queryClient.setQueryData(
+            AI_CHAT_MESSAGES_RQ_KEY(chatId),
+            (
+              old:
+                | {
+                    pages: { items: IAiChatMessageRow[]; meta: unknown }[];
+                    pageParams: unknown[];
+                  }
+                | undefined,
+            ) =>
+              old
+                ? { ...old, pages: mergeDeltaRowsIntoPages(old.pages, res.rows) }
+                : old,
+          );
+        }
+      } catch {
+        // Transient failure (e.g. a server restart mid-run): swallow and retry on
+        // the next tick — the poll must survive a bounce, like the old dumb refetch.
+      }
+    };
+    const id = setInterval(() => void tick(), 2500);
+    return () => {
+      cancelled = true;
+      clearInterval(id);
+    };
+  }, [degradedPoll, windowOpen, activeChatId, queryClient]);
 
   // #184 reconnect-and-live-follow. Whether detached agent runs are enabled for
   // this workspace. When the feature is off no runs are ever created, so the

apps/client/src/features/ai-chat/services/ai-chat-service.ts

@@ -57,6 +57,50 @@ export async function stopRun(
   return req.data;
 }
 
+/**
+ * Delta poll (#491): the chat's message rows changed since `cursor` (a DB-clock
+ * timestamp echoed from the previous poll) plus the current run fact, in ONE
+ * round-trip — the degraded-poll fallback's payload, replacing the old "refetch
+ * ALL infinite-query pages every 2.5s with full parts" poll. Omit `cursor` on the
+ * first poll (returns just a fresh cursor, no rows, to start the chain). The
+ * overlap window guarantees occasional REPEATS, so the caller MUST merge rows
+ * idempotently by id (mergeById). Owner-gated server-side.
+ */
+export async function getAiChatMessagesDelta(
+  chatId: string,
+  cursor?: string,
+): Promise<{
+  rows: IAiChatMessageRow[];
+  cursor: string;
+  run: { id: string; status: string } | null;
+}> {
+  const req = await api.post<{
+    rows: IAiChatMessageRow[];
+    cursor: string;
+    run: { id: string; status: string } | null;
+  }>("/ai-chat/messages/delta", { chatId, cursor });
+  return req.data;
+}
+
+/**
+ * #488: the run-fact — "is a run active on this chat?" — first-class from the
+ * server (POST /ai-chat/run). Called on mount to seed the client FSM's run-fact
+ * and to VERIFY after a supersede mismatch (an observer following a superseded
+ * run asks for the latest run and follows it). Returns the latest run row (with
+ * its `id` and `status`) and its projected assistant message, or `run: null` when
+ * the chat has never had a run. Owner-gated server-side.
+ */
+export async function getRun(chatId: string): Promise<{
+  run: { id: string; status: string } | null;
+  message: IAiChatMessageRow | null;
+}> {
+  const req = await api.post<{
+    run: { id: string; status: string } | null;
+    message: IAiChatMessageRow | null;
+  }>("/ai-chat/run", { chatId });
+  return req.data;
+}
+
 /**
  * Resolve the chat bound to a document (the current user's most-recent chat
  * created on that page), or null when there is none. Drives auto-open-on-page.

apps/client/src/features/ai-chat/state/run-fsm.spec.md

@@ -0,0 +1,190 @@
+# AI-chat run-lifecycle FSM — design spec (#488)
+
+This is the written design that `run-fsm.ts` implements. It ships in the PR (issue
+#488 commit 1: "the spec is written FIRST and enters the PR"). It has four parts:
+(1) the event × state transition table, (2) the map of every `chat-thread.tsx` ref
+to {FSM state | FSM context | stays data}, (3) the run-fact protocol, (4) the
+invariants.
+
+The reducer is a **pure function** `reduce(machine, event) → machine`. The returned
+machine carries the **command effects** for that transition; a thin runtime in
+`chat-thread.tsx` dispatches events and executes effects. Because it is pure, the
+whole machine is enumerable and unit-tested directly (event × state → next state is
+the observable property) — see `run-fsm.test.ts`.
+
+---
+
+## 1. Event × state transition table
+
+Phases: `idle | sending | streaming | attaching | reconnecting(attempt,failed) |
+polling(reason) | stalled | stopping | superseding | error(kind)`.
+Context (orthogonal): `epoch`, `ownership: local|observer`, `runFact: {runId}|null`,
+`liveFollow` (are we following a live run we locally streamed — the reconnect
+ladder — vs a one-shot mount-attach resume? both are `observer`, but a live-follow
+drop RE-ENTERS the ladder (#488 commit 3) while a mount-resume drop polls).
+
+Legend: **†** = command-transition (bumps `epoch`, I1). Effects in `[…]`.
+
+| Event (source) | From phase(s) | → To phase | Effects / ctx |
+|---|---|---|---|
+| `SEND_LOCAL` (user send) | idle, error, polling, stalled, reconnecting | sending **†** | `[cancelReconnect, disarmPoll]`, ownership=local |
+| `STREAM_START{runId}` (SDK `start` metadata) | sending, attaching, reconnecting, superseding | streaming | `[cancelReconnect, disarmPoll]`, runFact←runId |
+| `FINISH_CLEAN` (onFinish clean) | streaming, … | idle | `[disarmPoll, cancelReconnect]`, runFact←null |
+| `FINISH_ABORT` (onFinish isAbort) | streaming, stopping | idle | `[disarmPoll, cancelReconnect]`, runFact←null (I4 exits stopping by this DATA) |
+| `FINISH_DISCONNECT` (observer, NOT liveFollow) | streaming(observer) | polling(disconnect-visible) | `[armPoll]` (a mount-resume drop polls) |
+| `FINISH_DISCONNECT{hasVisibleContent}` (local drop OR liveFollow) | streaming | reconnecting(1) **†** *iff runFact\|liveFollow* | `[scheduleReconnect(1)]` (+`armPoll` if visible), ownership=observer, liveFollow=true (commit 3: repeatable) |
+| `FINISH_DISCONNECT` (no runFact, not liveFollow) | streaming | idle | runFact←null (plain terminal "connection lost") |
+| `STREAM_INCOMPLETE{reason}` (observer starved/torn clean finish) | streaming(observer) | polling(reason) | `[armPoll(reason)]` |
+| `FINISH_ERROR{kind}` (onFinish isError) | any | error(kind) | `[disarmPoll, cancelReconnect]`, runFact←null |
+| `STREAM_START{runId}` (first assistant frame of a local turn) | sending | streaming | runFact←runId, `[cancelReconnect, disarmPoll]` |
+| `ATTACH_START{runId}` (mount resume) | **idle only** (F2) | attaching **†** | `[resumeStream]`, ownership=observer, runFact←runId; ignored from any non-idle phase |
+| `ATTACH_LIVE` (attach GET 2xx) | attaching | streaming | — |
+| `ATTACH_NONE` (attach GET 204/err/throw) | attaching | polling(attach-none) | `[armPoll(attach-none)]` |
+| `RECONNECT_ATTEMPT{n}` (backoff timer) | reconnecting | reconnecting(n) **†** | `[resumeStream]` |
+| `RECONNECT_ATTACHED` (reconnect GET 2xx) | reconnecting | streaming | `[cancelReconnect, disarmPoll]` — **counter reset** (commit 3) |
+| `RECONNECT_NONE` (reconnect GET 204/err), attempt<MAX | reconnecting | reconnecting(n+1) **†** | `[armPoll(attach-none), scheduleReconnect(n+1)]` |
+| `RECONNECT_NONE`, attempt=MAX | reconnecting | reconnecting(MAX, failed) | `[armPoll(reconnect-exhausted)]` |
+| `RETRY` (manual, failed banner) | reconnecting(failed) | reconnecting(1) **†** | `[resumeStream]` |
+| `RETRY` (manual, stalled banner) | stalled | polling(attach-none) **†** | `[armPoll]` |
+| `POLL_TERMINAL` (settled tail merged) | polling, reconnecting, stopping | idle | `[disarmPoll, cancelReconnect]`, runFact←null (I4) |
+| `POLL_IDLE_CAP` (inactivity cap) | polling, reconnecting | stalled | `[disarmPoll, cancelReconnect]` (commit 4a — no more silent) |
+| `POLL_IDLE_CAP` (inactivity cap) | stopping | idle | `[disarmPoll, cancelReconnect]`, runFact←null (Review #4: a Stop-armed poll with no SDK/terminal backstop gets a bounded exit — NOT `stalled`, Stop was already pressed so nothing to retry) |
+| `RUN_FACT{null}` (POST /run → null/terminal, 204) | reconnecting/attaching/polling/stopping | idle | `[cancelReconnect, disarmPoll]`, runFact←null (I3 fresh-negative gate) |
+| `RUN_FACT{runId}` | any | (same) | runFact←runId (pessimism toward an attempt) |
+| `STOP_REQUESTED` (user Stop) | streaming, reconnecting, polling | stopping **†** | `[stopRun, abortAttach, cancelReconnect, armPoll]` (poll drives the terminal — I4 exit by data) |
+| `SUPERSEDE_REQUESTED{targetRunId}` (interrupt+send) | streaming, reconnecting, polling, error | superseding **†** | `[supersede(target), cancelReconnect, disarmPoll]` |
+| `SUPERSEDE_READY{runId}` (CAS ok) | superseding | streaming | ownership=local, runFact←runId |
+| `SUPERSEDE_MISMATCH{currentRunId}` (409 SUPERSEDE_TARGET_MISMATCH) | superseding | error(supersede-mismatch) | `[postRun(verify)]`, runFact←currentRunId |
+| `SUPERSEDE_TIMEOUT` (409 SUPERSEDE_TIMEOUT) | superseding | error(supersede-timeout) | — (composer keeps text; no auto-retry) |
+| `SUPERSEDE_INVALID` (409 SUPERSEDE_INVALID) | superseding | error(supersede-invalid) | — |
+| `RUN_ALREADY_ACTIVE{activeRunId}` (409 A_RUN_ALREADY_ACTIVE, plain POST) | sending | error(run-already-active) | runFact←activeRunId (composer offers supersede; NO auto-retry) |
+| `DISPOSE` (unmount) | any | idle **†** | `[abortAttach, cancelReconnect, disarmPoll]` (I1/I5 — epoch++ kills late callbacks) |
+
+**`stopping` honors any finish (re-review MEDIUM):** BEFORE the epoch filter, a
+stream finish (`FINISH_*`/`STREAM_INCOMPLETE`) arriving in phase `stopping` exits
+`stopping -> idle` regardless of generation. A plain Stop has no successor stream,
+so the aborted stream's finish IS the expected end (I4 exit by data) — and it
+carries the PRE-stop generation (STOP_REQUESTED bumped the epoch), so the filter
+would otherwise strand the machine in `stopping` (no idle-cap covers it). The filter
+stays in force for `superseding` (that is the F1 supersede drop).
+
+**Epoch filter (I1):** the reducer then drops any event carrying an `epoch` that
+does not equal the current `ctx.epoch`. Outcome events (`STREAM_START`, `ATTACH_*`,
+`RECONNECT_*`, `SUPERSEDE_*`, **`FINISH_*`/`STREAM_INCOMPLETE`**, `RUN_FACT`) are
+stamped with the generation the corresponding STREAM started under (the runtime
+holds a per-owned-stream `turnEpoch`); trigger events (user actions, fresh
+disconnects) carry no epoch. **F1:** this is what makes a SUPERSEDED stream's late
+`onFinish` (a dead stream A closing after the CAS started stream B) get dropped, so
+A cannot drive the live new run into a false reconnect or reset its run-fact. The
+supersede path additionally ABORTS A and starts B only from A's onFinish (a
+microtask), because ai@6 `AbstractChat.makeRequest` corrupts overlapping streams
+(A's `finally` reads then nulls the shared `activeResponse`).
+
+**Removed events (scope-cut, internal review):** `RUN_SUPERSEDED` (a ghost feature —
+never dispatched; the observer-superseded case is handled by the degraded poll,
+which follows the latest rows regardless of runId), `RECONNECT_BEGIN` (reconnect is
+entered by `FINISH_DISCONNECT`), and `POLL_ACTIVITY` (the window's activity clock was
+removed when the idle-cap moved into the thread). The reducer and this table now
+share exactly the dispatched event set.
+
+### 409-code → event map (the real #487 contract consumed here)
+
+| Server response | Event dispatched | error kind → banner |
+|---|---|---|
+| 409 `A_RUN_ALREADY_ACTIVE` (+ body.activeRunId) | `RUN_ALREADY_ACTIVE{activeRunId}` | run-already-active → "already answering / interrupt & send" |
+| 409 `SUPERSEDE_TARGET_MISMATCH` (+ body.activeRunId) | `SUPERSEDE_MISMATCH{currentRunId}` | supersede-mismatch → verify via /run |
+| 409 `SUPERSEDE_TIMEOUT` | `SUPERSEDE_TIMEOUT` | supersede-timeout → "couldn't interrupt in time, resend" |
+| 409 `SUPERSEDE_INVALID` | `SUPERSEDE_INVALID` | supersede-invalid → "couldn't interrupt this run" |
+| 503 `A_RUN_BEGIN_FAILED` | `FINISH_ERROR{begin-failed}` | begin-failed → "could not start, temporary" |
+
+---
+
+## 2. Ref-map — every `chat-thread.tsx` ref → its new home  (MIGRATION RESOLVED)
+
+The migration is COMPLETE: the 13 run-lifecycle FLAGS below are GONE from
+`chat-thread.tsx` (collapsed into FSM phase/ctx/effects, or deleted). What remains
+are identity/data mirrors, effect-owned controllers/timers, and ONE React-liveness
+bit — none of which is a run-lifecycle flag, so the post-merge "no new flags" rule
+holds. **Pending column: empty.**
+
+| # | Old ref | Resolved to | Where now |
+|---|---|---|---|
+| 1 | `reconcileTailRef` | **FSM phase** | reconcile-merge gated on `phase ∈ {polling, reconnecting, stopping}` |
+| 2 | `noStreamHandledRef` | **FSM epoch (I1)** | the attach outcome's epoch guard drops the stale/second outcome |
+| 3 | `onNoActiveStreamRef` | **FSM event** | transport → `handleAttachOutcome` dispatches `ATTACH_NONE`/`RECONNECT_NONE` |
+| 4 | `onReconnectAttachedRef` | **FSM event** | transport dispatches `ATTACH_LIVE` / `RECONNECT_ATTACHED` |
+| 5 | `resumedTurnRef` + `resumedTurn` state | **FSM ctx `ownership`** | `ownership==='observer'` ⇒ never flush; hides "Send now" |
+| 6 | `reconnectStateRef` + `reconnectState` state | **FSM phase** | `reconnecting(attempt,failed)` renders the banner |
+| 7 | `reconnectTimerRef` | **effect-owned timer** | owned by `scheduleReconnect`/`cancelReconnect` effects (not a flag) |
+| 8 | `flushOnAbortRef` | **DELETED** | the stop→flush dance is replaced by the CAS supersede (commit 5) |
+| 9 | `interruptNextSendRef` | **DELETED** | the server injects the interrupt note from the supersede itself |
+| 10 | `supersedeRetryRef` | **DELETED** (commit 5) | the client 409 retry ladder is gone; CAS supersede replaces it |
+| 11 | `stopPendingRef` | **FSM phase `stopping`** | the deferred stop fires from the chat-id adoption effect while `stopping` |
+| 12 | `mountedRef` | **retained (React liveness)** | orthogonal to run-lifecycle; gates imperative onFinish side-effects post-unmount. Epoch (I1) handles stale COMMAND-outcomes; DISPOSE bumps it |
+| 13 | `attemptResumeRef` | **FSM `ATTACH_START` + run-fact** | mount arms attach ONLY on a confirmed active run (commit 4b: streaming-tail status, or POST /run for a user tail) |
+| 14–15 | `anchorRef {id, stepsPersisted}` | **data** (attachStrategy) | #491 tail-only: replaced `stripRef`/`strippedRowRef`. The PERSISTED assistant row that pins the run (server invariant 6) + its step frontier N; feeds `?anchor=<id>&n=<stepsPersisted>`. No strip — the seed keeps every row; entering reconnecting re-seeds from persist |
+| 16 | `attachAbortRef` | **effect-owned controller** | aborted by the `abortAttach` effect in cleanup (I5) |
+| 17–25 | `chatIdRef`, `openPageRef`, `getEditorSelectionRef`, `roleIdRef`, `stableIdRef`, `queuedRef`, `sendMessageRef`, `statusRef`, `lastForwardedChatIdRef` | **data** (identity/send mirrors) | unchanged — not lifecycle flags |
+| NEW | `pendingSupersedeRef` | **data** (send-plumbing) | the runId injected into the next `POST /stream {supersede}`; the single replacement for the 3 DELETED one-shots (#8/#9/#10) — net −2 refs |
+| NEW | `idleCapTimerRef` | **effect-owned timer** | the stalled inactivity cap → `POLL_IDLE_CAP` (commit 4a); not a flag |
+
+Net: the 13 lifecycle flags (#1–#13) are eliminated: **8** → FSM phase/ctx/epoch/event
+(#1–#6, #11, #13), **3** deleted (#8/#9/#10), **`reconnectTimerRef` (#7)** becomes an
+effect-owned controller, and **`mountedRef` (#12)** is retained as React liveness
+(8 + 3 + 1 + 1 = 13). (`attachAbortRef` (#16) is outside the #1–#13 set — it was
+already an effect-owned controller.) Two effect-owned timers + one send-plumbing data
+ref are added — none is a boolean lifecycle latch.
+
+---
+
+## 3. Run-fact protocol (`runFact: {runId} | null`) — I3
+
+"A run is active" is first-class from the SERVER, not inferred from an assistant
+message. Sources, in the order they update `ctx.runFact`:
+
+1. **Init (mount):** `POST /ai-chat/run { chatId }` → `{ run, message }`. A `run`
+   with a non-terminal `status` seeds `runFact = { runId: run.id }`; a null/terminal
+   run seeds `null`. This is what arms the resume attempt (`ATTACH_START`) — the
+   attempt is armed ONLY on a positive fact (commit 4b: a user-tail with no active
+   run no longer arms a pointless poll on every open).
+2. **Live update:** the `start` stream metadata carries `runId` → `STREAM_START{runId}`.
+3. **Attach outcomes:** `ATTACH_LIVE` (2xx) confirms active; a 204 on a non-stripped
+   path is an authoritative NEGATIVE fact → the runtime dispatches `RUN_FACT{null}`,
+   which cancels recovery (I3 fresh-negative gate).
+4. **Poll (#491, implemented):** the degraded poll now hits the delta endpoint
+   (`POST /ai-chat/messages/delta`), which ALREADY carries the run fact
+   (`run: {id, status} | null`) alongside the changed rows. The client does NOT yet
+   consume that run field — it still drives to a terminal ROW (merged by id),
+   dispatched as `POLL_TERMINAL` — so the run field rides the wire for a future
+   client that settles straight off it.
+
+Pessimism rule: a stale-but-positive fact PERMITS entering recovery (attach); the
+204 then cuts it. A fresh negative fact gates recovery OUT immediately.
+
+---
+
+## 4. Invariants
+
+- **I1 — Epoch (generation counter).** Every command-emitting transition bumps
+  `ctx.epoch`; every async outcome event carries its issuing epoch; the reducer
+  drops stale-epoch outcomes. Replaces the one-shot-ref zoo (`noStreamHandledRef`,
+  the flush/interrupt/supersede one-shots, the `mountedRef` late-callback gate).
+- **I2 — Ownership is context, not state.** `local | observer` is orthogonal to the
+  transport phase. The queue flushes ONLY under local ownership; an observer
+  following a detached run never flushes (was `resumedTurnRef`).
+- **I3 — Run-fact is first-class from the server.** Reconnect is entered by the
+  run-fact, not by an assistant message (commit 2). A fresh negative fact cancels
+  recovery.
+- **I4 — Exit `stopping` by DATA.** A terminal row / negative run-fact / terminal
+  finish exits `stopping`, never the stopRun HTTP response (which returns after the
+  abort but before finalization — keying off it would unlock the composer on a 409).
+- **I5 — Dispose protocol.** Command controllers (attach GET, POST /stream, POST
+  /run) are effect-owned and aborted in cleanup (`abortAttach` on `DISPOSE`), not
+  render-phase refs. A client abort of an already-sent POST does not cancel the
+  server action, so disarming on unmount is safe.
+- **attachStrategy** is behind the `resumeStream` effect; #491 swapped it to
+  tail-only (`?anchor=&n=`, `anchorRef` data) WITHOUT touching the FSM. Entering
+  reconnecting always re-seeds from persist; on a getRun failure the live partial
+  is dropped + replay-from-start so it is never the tail-apply base (no #137/#161
+  duplication).
+- **Queue** stays a data structure; flush/interrupt decisions are transitions.

apps/client/src/features/ai-chat/state/run-fsm.test.ts

@@ -0,0 +1,482 @@
+import { describe, it, expect } from "vitest";
+import {
+  reduce,
+  initialMachine,
+  reconnectDelayMs,
+  RECONNECT_MAX_ATTEMPTS,
+  type Machine,
+  type Effect,
+  type Event,
+} from "./run-fsm";
+
+// Drive a sequence of events through the reducer, returning the final machine.
+function run(m: Machine, ...events: Event[]): Machine {
+  return events.reduce(reduce, m);
+}
+function withRunFact(runId = "run-1"): Machine {
+  return {
+    ...initialMachine(),
+    ctx: { epoch: 0, ownership: "local", runFact: { runId }, liveFollow: false },
+  };
+}
+function effectTypes(m: Machine): string[] {
+  return m.effects.map((e) => e.type);
+}
+function hasEffect(m: Machine, type: Effect["type"]): boolean {
+  return m.effects.some((e) => e.type === type);
+}
+
+describe("run-fsm — epoch invariant (I1)", () => {
+  it("drops an outcome carrying a stale epoch", () => {
+    // A command bumps the epoch; an outcome stamped with the OLD epoch is dropped.
+    const m0 = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" }); // epoch 0->1, attaching
+    expect(m0.ctx.epoch).toBe(1);
+    expect(m0.phase.name).toBe("attaching");
+    // A late ATTACH_LIVE from a SUPERSEDED attempt (epoch 0) must NOT drive us.
+    const stale = reduce(m0, { type: "ATTACH_LIVE", epoch: 0 });
+    expect(stale.phase.name).toBe("attaching");
+    expect(stale.effects).toEqual([]);
+  });
+
+  it("applies an outcome carrying the current epoch", () => {
+    const m0 = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    const live = reduce(m0, { type: "ATTACH_LIVE", epoch: m0.ctx.epoch });
+    expect(live.phase.name).toBe("streaming");
+  });
+
+  it("an outcome with no epoch is never dropped (trigger events)", () => {
+    const m0 = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    const disposed = reduce(m0, { type: "DISPOSE" });
+    expect(disposed.phase.name).toBe("idle");
+    expect(hasEffect(disposed, "abortAttach")).toBe(true);
+  });
+
+  it("every command-transition increments the epoch exactly once", () => {
+    let m = initialMachine();
+    const before = m.ctx.epoch;
+    m = reduce(m, { type: "SEND_LOCAL" });
+    expect(m.ctx.epoch).toBe(before + 1);
+    m = reduce(m, { type: "STOP_REQUESTED" });
+    expect(m.ctx.epoch).toBe(before + 2);
+  });
+});
+
+describe("run-fsm — local turn", () => {
+  it("SEND_LOCAL → sending, local ownership, cancels recovery", () => {
+    const m = reduce(withRunFact(), { type: "SEND_LOCAL" });
+    expect(m.phase.name).toBe("sending");
+    expect(m.ctx.ownership).toBe("local");
+    expect(effectTypes(m)).toEqual(
+      expect.arrayContaining(["cancelReconnect", "disarmPoll"]),
+    );
+  });
+
+  it("STREAM_START adopts the runId into the run-fact and goes streaming", () => {
+    const m = run(initialMachine(), { type: "SEND_LOCAL" });
+    const s = reduce(m, { type: "STREAM_START", runId: "run-9", epoch: m.ctx.epoch });
+    expect(s.phase.name).toBe("streaming");
+    expect(s.ctx.runFact).toEqual({ runId: "run-9" });
+  });
+
+  it("FINISH_CLEAN → idle, run-fact cleared, poll/reconnect disarmed", () => {
+    const streaming = run(initialMachine(), { type: "SEND_LOCAL" }, { type: "STREAM_START", runId: "r" });
+    const done = reduce(streaming, { type: "FINISH_CLEAN" });
+    expect(done.phase.name).toBe("idle");
+    expect(done.ctx.runFact).toBeNull();
+  });
+});
+
+// #488 commit 2 — SSE break BEFORE the first assistant frame must still recover.
+describe("run-fsm — commit 2: reconnect by run-fact, not by assistant message", () => {
+  it("FINISH_DISCONNECT with an active run-fact → reconnecting (even with no visible content)", () => {
+    // Setup-phase break: no assistant frame yet, but a run-fact exists.
+    const streaming = withRunFact("run-2");
+    const m = reduce(streaming, {
+      type: "FINISH_DISCONNECT",
+      hasVisibleContent: false,
+      epoch: streaming.ctx.epoch,
+    });
+    expect(m.phase.name).toBe("reconnecting");
+    if (m.phase.name === "reconnecting") expect(m.phase.attempt).toBe(1);
+    expect(m.ctx.ownership).toBe("observer");
+    expect(hasEffect(m, "scheduleReconnect")).toBe(true);
+    // No visible content -> no poll arm yet (the reconnect ladder rebuilds it).
+    expect(hasEffect(m, "armPoll")).toBe(false);
+  });
+
+  it("FINISH_DISCONNECT WITH visible content also arms the poll", () => {
+    const m = reduce(withRunFact("run-2"), {
+      type: "FINISH_DISCONNECT",
+      hasVisibleContent: true,
+      epoch: 0,
+    });
+    expect(m.phase.name).toBe("reconnecting");
+    expect(hasEffect(m, "armPoll")).toBe(true);
+  });
+
+  it("FINISH_DISCONNECT with NO run-fact → idle (plain connection-lost)", () => {
+    const m = reduce(initialMachine(), {
+      type: "FINISH_DISCONNECT",
+      hasVisibleContent: true,
+      epoch: 0,
+    });
+    expect(m.phase.name).toBe("idle");
+  });
+});
+
+// #488 commit 3 — a SECOND break after a successful re-attach starts a NEW ladder.
+describe("run-fsm — commit 3: repeated reconnect cycles", () => {
+  it("two breaks in a row produce two reconnect cycles (counter resets on attach)", () => {
+    let m = withRunFact("run-3");
+    // First break -> reconnecting(1).
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: false, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("reconnecting");
+    // Attempt fires, re-attaches live.
+    m = reduce(m, { type: "RECONNECT_ATTEMPT", attempt: 1, epoch: m.ctx.epoch });
+    m = reduce(m, { type: "RECONNECT_ATTACHED", epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("streaming");
+    // SECOND break: the counter was reset, so a fresh ladder starts at attempt 1
+    // (the old one-shot !wasResumed gate would have sent this to silent poll).
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: false, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("reconnecting");
+    if (m.phase.name === "reconnecting") expect(m.phase.attempt).toBe(1);
+    expect(hasEffect(m, "scheduleReconnect")).toBe(true);
+  });
+
+  it("a MOUNT-attach observer drop falls to POLL, not the reconnect ladder", () => {
+    // Distinguishes commit 3 from a one-shot resume: an observer that never
+    // live-followed (liveFollow false) polls on a drop.
+    let m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_LIVE", epoch: m.ctx.epoch });
+    expect(m.ctx.ownership).toBe("observer");
+    expect(m.ctx.liveFollow).toBe(false);
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: true, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("polling");
+    expect(hasEffect(m, "armPoll")).toBe(true);
+  });
+
+  it("STREAM_INCOMPLETE (observer starved/torn finish) → polling", () => {
+    let m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_LIVE", epoch: m.ctx.epoch });
+    m = reduce(m, { type: "STREAM_INCOMPLETE", reason: "starved", epoch: m.ctx.epoch });
+    expect(m.phase).toEqual({ name: "polling", reason: "starved" });
+    expect(hasEffect(m, "armPoll")).toBe(true);
+  });
+
+  it("liveFollow is set on the first local drop and kept across a re-attach", () => {
+    let m = withRunFact("run-3");
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: false, epoch: m.ctx.epoch });
+    expect(m.ctx.liveFollow).toBe(true);
+    m = reduce(m, { type: "RECONNECT_ATTEMPT", attempt: 1, epoch: m.ctx.epoch });
+    m = reduce(m, { type: "RECONNECT_ATTACHED", epoch: m.ctx.epoch });
+    expect(m.ctx.liveFollow).toBe(true); // kept — so a second drop reconnects
+    // A clean finish clears it.
+    m = reduce(m, { type: "FINISH_CLEAN", epoch: m.ctx.epoch });
+    expect(m.ctx.liveFollow).toBe(false);
+  });
+
+  it("RECONNECT_NONE backs off through the ladder, then fails at the cap", () => {
+    let m = withRunFact("run-3");
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: false, epoch: m.ctx.epoch });
+    for (let n = 1; n < RECONNECT_MAX_ATTEMPTS; n++) {
+      m = reduce(m, { type: "RECONNECT_ATTEMPT", attempt: n, epoch: m.ctx.epoch });
+      m = reduce(m, { type: "RECONNECT_NONE", epoch: m.ctx.epoch });
+      expect(m.phase.name).toBe("reconnecting");
+      if (m.phase.name === "reconnecting") {
+        expect(m.phase.attempt).toBe(n + 1);
+        expect(m.phase.failed).toBe(false);
+      }
+      // The belt-and-suspenders poll is armed each failed attempt.
+      expect(hasEffect(m, "armPoll")).toBe(true);
+    }
+    // Final attempt fails -> failed banner (Retry), poll armed.
+    m = reduce(m, { type: "RECONNECT_ATTEMPT", attempt: RECONNECT_MAX_ATTEMPTS, epoch: m.ctx.epoch });
+    m = reduce(m, { type: "RECONNECT_NONE", epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("reconnecting");
+    if (m.phase.name === "reconnecting") expect(m.phase.failed).toBe(true);
+    // RETRY restarts at attempt 1.
+    m = reduce(m, { type: "RETRY" });
+    expect(m.phase.name).toBe("reconnecting");
+    if (m.phase.name === "reconnecting") {
+      expect(m.phase.attempt).toBe(1);
+      expect(m.phase.failed).toBe(false);
+    }
+    expect(hasEffect(m, "resumeStream")).toBe(true);
+  });
+
+  it("reconnectDelayMs is the exponential backoff 1s,2s,4s,8s,16s", () => {
+    expect([1, 2, 3, 4, 5].map(reconnectDelayMs)).toEqual([1000, 2000, 4000, 8000, 16000]);
+  });
+});
+
+// #488 commit 4 — polling stalled-state + user-tail gating.
+describe("run-fsm — commit 4: stalled + run-fact gating", () => {
+  it("POLL_IDLE_CAP: polling → stalled with a banner (poll disarmed), not silent", () => {
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_NONE", epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("polling");
+    m = reduce(m, { type: "POLL_IDLE_CAP" });
+    expect(m.phase.name).toBe("stalled");
+    expect(hasEffect(m, "disarmPoll")).toBe(true);
+  });
+
+  it("RETRY from stalled re-arms the poll", () => {
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_NONE", epoch: m.ctx.epoch });
+    m = reduce(m, { type: "POLL_IDLE_CAP" });
+    m = reduce(m, { type: "RETRY" });
+    expect(m.phase.name).toBe("polling");
+    expect(hasEffect(m, "armPoll")).toBe(true);
+  });
+
+  it("a fresh NEGATIVE run-fact while attaching cancels recovery (user-tail, no active run)", () => {
+    // The mount POST /run returns no active run: attaching → idle, no poll armed.
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "RUN_FACT", runFact: null, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("idle");
+    expect(m.ctx.runFact).toBeNull();
+    expect(hasEffect(m, "disarmPoll")).toBe(true);
+  });
+
+  it("a negative run-fact while polling stops the poll", () => {
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_NONE", epoch: m.ctx.epoch });
+    m = reduce(m, { type: "RUN_FACT", runFact: null, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("idle");
+  });
+
+  it("POLL_TERMINAL settles polling → idle (I4 data-driven exit)", () => {
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_NONE", epoch: m.ctx.epoch });
+    m = reduce(m, { type: "POLL_TERMINAL" });
+    expect(m.phase.name).toBe("idle");
+    expect(m.ctx.runFact).toBeNull();
+  });
+});
+
+// #488 commit 5 — error classification + supersede CAS transitions.
+describe("run-fsm — commit 5: supersede CAS + error classification", () => {
+  it("SUPERSEDE_REQUESTED → superseding, fires the CAS effect, bumps epoch", () => {
+    const streaming = withRunFact("run-old");
+    const m = reduce(streaming, { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    expect(m.phase.name).toBe("superseding");
+    expect(m.ctx.epoch).toBe(streaming.ctx.epoch + 1);
+    const sup = m.effects.find((e) => e.type === "supersede");
+    expect(sup).toEqual({ type: "supersede", targetRunId: "run-old" });
+  });
+
+  it("SUPERSEDE_READY → streaming as the new local owner", () => {
+    let m = reduce(withRunFact("run-old"), { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    m = reduce(m, { type: "SUPERSEDE_READY", runId: "run-new", epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("streaming");
+    expect(m.ctx.ownership).toBe("local");
+    expect(m.ctx.runFact).toEqual({ runId: "run-new" });
+  });
+
+  it("SUPERSEDE_MISMATCH → error(supersede-mismatch) + verify via /run (no blind banner)", () => {
+    let m = reduce(withRunFact("run-old"), { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    m = reduce(m, { type: "SUPERSEDE_MISMATCH", currentRunId: "run-x", epoch: m.ctx.epoch });
+    expect(m.phase).toEqual({ name: "error", kind: "supersede-mismatch" });
+    expect(hasEffect(m, "postRun")).toBe(true);
+    expect(m.ctx.runFact).toEqual({ runId: "run-x" });
+  });
+
+  it("SUPERSEDE_TIMEOUT → error(supersede-timeout), no auto-retry effect", () => {
+    let m = reduce(withRunFact("run-old"), { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    m = reduce(m, { type: "SUPERSEDE_TIMEOUT", epoch: m.ctx.epoch });
+    expect(m.phase).toEqual({ name: "error", kind: "supersede-timeout" });
+    expect(m.effects).toEqual([]);
+  });
+
+  it("SUPERSEDE_INVALID → error(supersede-invalid)", () => {
+    let m = reduce(withRunFact("run-old"), { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    m = reduce(m, { type: "SUPERSEDE_INVALID", epoch: m.ctx.epoch });
+    expect(m.phase).toEqual({ name: "error", kind: "supersede-invalid" });
+  });
+
+  it("a stale SUPERSEDE outcome from a superseded epoch is dropped", () => {
+    let m = reduce(withRunFact("run-old"), { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    const supersedingEpoch = m.ctx.epoch;
+    // The user retriggers, bumping the epoch again.
+    m = reduce(m, { type: "SUPERSEDE_REQUESTED", targetRunId: "run-old" });
+    // The first CAS's late TIMEOUT (old epoch) must NOT knock us out of superseding.
+    const late = reduce(m, { type: "SUPERSEDE_TIMEOUT", epoch: supersedingEpoch });
+    expect(late.phase.name).toBe("superseding");
+  });
+
+  it("RUN_ALREADY_ACTIVE (plain POST gate) → error(run-already-active), no retry effect", () => {
+    const m = reduce(run(initialMachine(), { type: "SEND_LOCAL" }), { type: "RUN_ALREADY_ACTIVE" });
+    expect(m.phase).toEqual({ name: "error", kind: "run-already-active" });
+    expect(m.effects).toEqual([]);
+  });
+
+  it("#497/S4: RUN_ALREADY_ACTIVE{activeRunId} ADOPTS the server's active run as the run-fact", () => {
+    // The server sends `activeRunId` so a later supersede can TARGET that run
+    // instead of a blind promote+abort. Absorb it into runFact.
+    const m = reduce(run(initialMachine(), { type: "SEND_LOCAL" }), {
+      type: "RUN_ALREADY_ACTIVE",
+      activeRunId: "run-foreign",
+    });
+    expect(m.phase).toEqual({ name: "error", kind: "run-already-active" });
+    expect(m.ctx.runFact).toEqual({ runId: "run-foreign" });
+    expect(m.effects).toEqual([]);
+  });
+
+  it("#497/S4: RUN_ALREADY_ACTIVE without an activeRunId keeps the prior run-fact", () => {
+    const seeded = reduce(run(initialMachine(), { type: "SEND_LOCAL" }), {
+      type: "RUN_FACT",
+      runFact: { runId: "run-prior" },
+    });
+    const m = reduce(seeded, { type: "RUN_ALREADY_ACTIVE" });
+    expect(m.ctx.runFact).toEqual({ runId: "run-prior" });
+  });
+});
+
+// #488 F2 — a late mount `getRun → ATTACH_START` must not hijack a local turn.
+describe("run-fsm — F2: ATTACH_START only from idle", () => {
+  it("ATTACH_START from a local `sending` turn is ignored (no observer hijack)", () => {
+    const sending = reduce(initialMachine(), { type: "SEND_LOCAL" }); // idle -> sending, local
+    const m = reduce(sending, { type: "ATTACH_START", runId: "r" });
+    expect(m.phase.name).toBe("sending");
+    expect(m.ctx.ownership).toBe("local"); // NOT flipped to observer
+    expect(m.effects).toEqual([]); // no resumeStream
+  });
+
+  it("ATTACH_START from idle attaches as normal", () => {
+    const m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    expect(m.phase.name).toBe("attaching");
+    expect(m.ctx.ownership).toBe("observer");
+    expect(hasEffect(m, "resumeStream")).toBe(true);
+  });
+});
+
+describe("run-fsm — stop (I4: exit by data)", () => {
+  it("STOP_REQUESTED → stopping, fires stopRun + abortAttach, no data-independent exit", () => {
+    const m = reduce(withRunFact(), { type: "STOP_REQUESTED" });
+    expect(m.phase.name).toBe("stopping");
+    expect(effectTypes(m)).toEqual(expect.arrayContaining(["stopRun", "abortAttach"]));
+  });
+
+  it("stopping exits on the aborted stream's finish carrying the PRE-STOP epoch", () => {
+    // MEDIUM (#488 re-review): STOP_REQUESTED is a command that BUMPS the epoch, but
+    // the runtime stamps the aborted stream's onFinish with the stream's START (pre-
+    // stop) generation — exactly what the component sends. `stopping` must HONOR
+    // that finish regardless of generation (no idle-cap covers `stopping`).
+    // MUTATION-VERIFY: remove the honor-in-`stopping` branch and this hangs in
+    // `stopping` (the epoch filter drops the pre-stop finish) -> red.
+    const preStopEpoch = withRunFact().ctx.epoch; // E1 (the stream's start epoch)
+    let m = reduce(withRunFact(), { type: "STOP_REQUESTED" }); // E1 -> E2, stopping
+    expect(m.ctx.epoch).toBe(preStopEpoch + 1);
+    m = reduce(m, { type: "FINISH_ABORT", epoch: preStopEpoch }); // NOT the current epoch
+    expect(m.phase.name).toBe("idle");
+    expect(m.ctx.runFact).toBeNull();
+  });
+
+  it("stopping exits on a clean finish carrying the pre-stop epoch too", () => {
+    const preStopEpoch = withRunFact().ctx.epoch;
+    let m = reduce(withRunFact(), { type: "STOP_REQUESTED" });
+    m = reduce(m, { type: "FINISH_CLEAN", epoch: preStopEpoch });
+    expect(m.phase.name).toBe("idle");
+  });
+
+  it("stopping exits on a negative run-fact (data)", () => {
+    let m = reduce(withRunFact(), { type: "STOP_REQUESTED" });
+    m = reduce(m, { type: "RUN_FACT", runFact: null, epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("idle");
+  });
+
+  // Review #4: `stopping` arms the poll but had no inactivity backstop.
+  it("review-4: POLL_IDLE_CAP in `stopping` exits to idle (bounded), NOT stalled", () => {
+    let m = reduce(withRunFact(), { type: "STOP_REQUESTED" });
+    expect(m.phase.name).toBe("stopping");
+    expect(hasEffect(m, "armPoll")).toBe(true);
+    // MUTATION-VERIFY: drop the `stopping` branch in POLL_IDLE_CAP and this hangs
+    // in `stopping` (poll forever) -> red.
+    m = reduce(m, { type: "POLL_IDLE_CAP" });
+    expect(m.phase.name).toBe("idle");
+    expect(hasEffect(m, "disarmPoll")).toBe(true);
+    expect(m.ctx.ownership).toBe("local");
+  });
+});
+
+// Review #1: positive attach outcomes must be guarded by the SOURCE phase — the
+// epoch filter alone is insufficient because POLL_TERMINAL uses to() (no epoch
+// bump) and does not abort the in-flight GET.
+describe("run-fsm — review-1: attach outcomes guarded by source phase", () => {
+  it("a late RECONNECT_ATTACHED after POLL_TERMINAL stays idle (no phantom streaming)", () => {
+    let m = withRunFact("run-1");
+    m = reduce(m, { type: "FINISH_DISCONNECT", hasVisibleContent: true, epoch: m.ctx.epoch });
+    m = reduce(m, { type: "RECONNECT_ATTEMPT", attempt: 1, epoch: m.ctx.epoch }); // attach GET
+    const epoch = m.ctx.epoch;
+    // The armed degraded poll reaches the terminal row FIRST (epoch unchanged).
+    m = reduce(m, { type: "POLL_TERMINAL" });
+    expect(m.phase.name).toBe("idle");
+    expect(m.ctx.epoch).toBe(epoch); // POLL_TERMINAL did NOT bump the epoch
+    // The slow GET returns live 2xx under the SAME epoch — must NOT resurrect.
+    m = reduce(m, { type: "RECONNECT_ATTACHED", epoch });
+    expect(m.phase.name).toBe("idle");
+  });
+
+  it("a late ATTACH_LIVE / ATTACH_NONE after leaving `attaching` is ignored", () => {
+    let m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    const epoch = m.ctx.epoch;
+    m = reduce(m, { type: "ATTACH_NONE", epoch }); // attaching -> polling
+    m = reduce(m, { type: "POLL_TERMINAL" }); // -> idle (epoch unchanged)
+    expect(m.phase.name).toBe("idle");
+    m = reduce(m, { type: "ATTACH_LIVE", epoch }); // late 2xx, same epoch
+    expect(m.phase.name).toBe("idle");
+    // And a late ATTACH_NONE (not `attaching`) is a no-op too.
+    m = reduce(m, { type: "ATTACH_NONE", epoch });
+    expect(m.phase.name).toBe("idle");
+  });
+});
+
+// Review #2: every terminal transition resets ownership to local.
+describe("run-fsm — review-2: terminal transitions reset ownership to local", () => {
+  const observer = (): Machine => {
+    let m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    m = reduce(m, { type: "ATTACH_LIVE", epoch: m.ctx.epoch });
+    expect(m.ctx.ownership).toBe("observer");
+    return m;
+  };
+  it("FINISH_CLEAN resets ownership", () => {
+    const m = reduce(observer(), { type: "FINISH_CLEAN", epoch: observer().ctx.epoch });
+    expect(m.ctx.ownership).toBe("local");
+  });
+  it("FINISH_ERROR / POLL_TERMINAL / RUN_FACT(null) reset ownership", () => {
+    let o = observer();
+    expect(reduce(o, { type: "FINISH_ERROR", kind: "stream", epoch: o.ctx.epoch }).ctx.ownership).toBe("local");
+    // POLL_TERMINAL from an observer polling phase
+    let p = reduce(observer(), { type: "STREAM_INCOMPLETE", reason: "starved", epoch: observer().ctx.epoch });
+    expect(reduce(p, { type: "POLL_TERMINAL" }).ctx.ownership).toBe("local");
+    // RUN_FACT(null) from an observer attaching phase
+    let a = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    expect(reduce(a, { type: "RUN_FACT", runFact: null, epoch: a.ctx.epoch }).ctx.ownership).toBe("local");
+  });
+});
+
+describe("run-fsm — ownership (I2) is context, orthogonal to phase", () => {
+  it("attach/reconnect set observer; send/supersede-ready set local", () => {
+    let m = reduce(initialMachine(), { type: "ATTACH_START", runId: "r" });
+    expect(m.ctx.ownership).toBe("observer");
+    m = reduce(m, { type: "ATTACH_LIVE", epoch: m.ctx.epoch });
+    expect(m.phase.name).toBe("streaming");
+    expect(m.ctx.ownership).toBe("observer"); // still observing a detached run
+    // A local send flips ownership back to local.
+    m = reduce(m, { type: "SEND_LOCAL" });
+    expect(m.ctx.ownership).toBe("local");
+  });
+});
+
+describe("run-fsm — dispose (I5)", () => {
+  it("DISPOSE from any phase aborts controllers and bumps epoch", () => {
+    let m = reduce(withRunFact(), { type: "ATTACH_START", runId: "r" });
+    const before = m.ctx.epoch;
+    m = reduce(m, { type: "DISPOSE" });
+    expect(m.phase.name).toBe("idle");
+    expect(m.ctx.epoch).toBe(before + 1);
+    expect(effectTypes(m)).toEqual(
+      expect.arrayContaining(["abortAttach", "cancelReconnect", "disarmPoll"]),
+    );
+  });
+});

apps/client/src/features/ai-chat/state/run-fsm.ts

@@ -0,0 +1,600 @@
+/**
+ * Run-lifecycle finite state machine for a single AI-chat thread (#488).
+ *
+ * ============================================================================
+ * WHY THIS EXISTS
+ * ----------------------------------------------------------------------------
+ * The resume/reconnect/poll/stop/supersede lifecycle used to be spread across
+ * ~26 `useRef` one-shot flags in `chat-thread.tsx`, each disarmed "on every
+ * path". Ownerless flag combinations produced silent UI freezes, and every fix
+ * added another ref (the #381 -> #432 -> #456 spiral). This module replaces that
+ * ref-zoo with ONE pure reducer whose transitions are enumerable and unit-
+ * testable in isolation (event x state -> next state is the observable property).
+ *
+ * The reducer is PURE: it owns no timers, no fetches, no React state. It maps
+ * `(machine, event) -> machine`, where the returned machine carries the list of
+ * COMMAND EFFECTS to run for that transition. A thin runtime in `chat-thread.tsx`
+ * dispatches events (from SDK callbacks / HTTP outcomes) and executes the
+ * effects (attach GET, POST /stream, POST /run, POST /stop, backoff timers,
+ * poll arm/disarm). The runtime lives in a THREAD, not the window, so a late SDK
+ * callback dies with the owner (kills the "event from a dead view" class, #161).
+ *
+ * ============================================================================
+ * INVARIANTS (see run-fsm.spec.md for the full spec + tables)
+ * ----------------------------------------------------------------------------
+ *  I1  EPOCH (generation counter). Commands (`resumeStream`, `postRun`, `stop`,
+ *      `supersede`, `scheduleReconnect`) are async; their outcomes arrive on the
+ *      SAME SDK/HTTP callbacks. Every command-emitting transition increments
+ *      `ctx.epoch`; every OUTCOME event carries the epoch it was issued under;
+ *      the reducer DROPS an outcome whose epoch != the current epoch. This is
+ *      what the one-shot-ref zoo used to approximate by hand.
+ *  I2  OWNERSHIP is a CONTEXT FIELD (`'local' | 'observer'`), not a state —
+ *      orthogonal to the transport phase. The queue is flushed ONLY by a local
+ *      owner (an observer following a detached run never flushes).
+ *  I3  RUN-FACT ("a run is active") is first-class from the server: `runFact`
+ *      holds the server-confirmed active run id (POST /run on mount, the `start`
+ *      metadata runId, attach outcomes). Reconnect is entered by the RUN-FACT,
+ *      not by the presence of an assistant message (#488 commit 2). A fresh
+ *      negative fact (null) cancels reconnect immediately.
+ *  I4  Exit `stopping` by DATA (a terminal row / negative run-fact), NEVER by the
+ *      stopRun HTTP response (which returns after abort, before finalization).
+ *  I5  Command controllers are effect-owned (abort in cleanup), NOT render-phase
+ *      refs — expressed here as the `abortAttach` effect on disposing transitions.
+ * ============================================================================
+ */
+
+// ---------------------------------------------------------------------------
+// Phases (the transport lifecycle). Ownership / runFact are CONTEXT, not here.
+// ---------------------------------------------------------------------------
+
+/** Why the degraded poll is the active recovery. */
+export type PollReason =
+  | "attach-none" // mount attach returned 204 / error — nothing live to attach
+  | "starved" // a resumed finish carried no visible content
+  | "disconnect-visible" // a live disconnect WITH on-screen content — poll to terminal
+  | "reconnect-exhausted"; // the live re-attach ladder gave up
+
+/** The classified error kind (drives the banner text + composer behavior). */
+export type ErrorKind =
+  | "stream" // a generic provider/network stream error (useChat error)
+  | "run-already-active" // 409 A_RUN_ALREADY_ACTIVE (a plain POST hit the gate)
+  | "supersede-mismatch" // 409 SUPERSEDE_TARGET_MISMATCH (CAS target moved)
+  | "supersede-timeout" // 409 SUPERSEDE_TIMEOUT (old run did not settle in W)
+  | "supersede-invalid" // 409 SUPERSEDE_INVALID (bad supersede target)
+  | "begin-failed"; // 503 A_RUN_BEGIN_FAILED (could not start the run)
+
+export type Phase =
+  | { name: "idle" }
+  | { name: "sending" } // local POST in flight, before the first frame
+  | { name: "streaming" } // receiving frames
+  | { name: "attaching" } // mount-time attach GET in flight
+  | { name: "reconnecting"; attempt: number; failed: boolean }
+  | { name: "polling"; reason: PollReason }
+  | { name: "stalled" } // poll hit the inactivity cap — banner + Retry
+  | { name: "stopping" }
+  | { name: "superseding" }
+  | { name: "error"; kind: ErrorKind };
+
+export type Ownership = "local" | "observer";
+
+/** The server-confirmed active run, or null when no run is active. */
+export type RunFact = { runId: string } | null;
+
+export interface Ctx {
+  /** I1: generation counter — every command-transition increments it. */
+  epoch: number;
+  /** I2: does THIS client own the turn's writes (local streamer) or observe? */
+  ownership: Ownership;
+  /** I3: the server-confirmed active run. */
+  runFact: RunFact;
+  /**
+   * Are we FOLLOWING a live run we were locally streaming (the reconnect ladder),
+   * as opposed to a one-shot mount-attach resume? Both are `ownership: 'observer'`,
+   * but they recover DIFFERENTLY on a drop: a live-follow drop RE-ENTERS the
+   * reconnect ladder (#488 commit 3 — the second break after a successful re-attach
+   * must reconnect again, not fall to silent poll), while a mount-resume drop falls
+   * to the degraded poll. This is the ctx bit that separates the two WITHOUT a new
+   * component ref (it is why commit 3 needs the FSM, not a surgical patch).
+   */
+  liveFollow: boolean;
+}
+
+export interface Machine {
+  phase: Phase;
+  ctx: Ctx;
+  /** Command effects to run for the transition that produced THIS machine.
+   *  The runtime executes them and does not read them again. */
+  effects: Effect[];
+}
+
+// ---------------------------------------------------------------------------
+// Command effects (the reducer's only side-channel — executed by the runtime).
+// ---------------------------------------------------------------------------
+
+export type Effect =
+  /** POST /run to (re)establish or verify the run-fact. `reason` is diagnostic. */
+  | { type: "postRun"; reason: "mount" | "verify" }
+  /** Trigger the SDK `resumeStream()` (attach GET via prepareReconnectToStream). */
+  | { type: "resumeStream" }
+  /** Schedule a reconnect attempt after a backoff, then dispatch RECONNECT_ATTEMPT. */
+  | { type: "scheduleReconnect"; attempt: number; delayMs: number }
+  /** Cancel any pending reconnect backoff timer. */
+  | { type: "cancelReconnect" }
+  /** Arm the degraded poll (the window's dumb timer follows the run in the DB). */
+  | { type: "armPoll"; reason: PollReason }
+  /** Disarm the degraded poll. */
+  | { type: "disarmPoll" }
+  /** POST /stop the chat's active run (authoritative detached-run stop). */
+  | { type: "stopRun" }
+  /** POST /stream { supersede: { runId } } — the CAS "interrupt and send now". */
+  | { type: "supersede"; targetRunId: string }
+  /** Abort the in-flight attach/reconnect GET controller (dispose / observer stop). */
+  | { type: "abortAttach" };
+
+// ---------------------------------------------------------------------------
+// Events. An OUTCOME event MAY carry `epoch`; if it does and it does not equal
+// the current epoch, the reducer drops it (I1). Trigger events (user actions,
+// fresh disconnects) carry no epoch and are never dropped.
+// ---------------------------------------------------------------------------
+
+export type Event =
+  // -- local turn --
+  | { type: "SEND_LOCAL" }
+  | { type: "STREAM_START"; runId?: string; epoch?: number }
+  /** An OBSERVER's attached stream ended WITHOUT reaching terminal (a starved
+   *  clean replay, or a torn resume) — fall to the degraded poll to drive the row
+   *  to its real terminal state. (A live-follow drop uses FINISH_DISCONNECT.) */
+  | { type: "STREAM_INCOMPLETE"; reason: PollReason; epoch?: number }
+  | { type: "FINISH_CLEAN"; epoch?: number }
+  | { type: "FINISH_ABORT"; epoch?: number }
+  | { type: "FINISH_DISCONNECT"; hasVisibleContent: boolean; epoch?: number }
+  | { type: "FINISH_ERROR"; kind: ErrorKind; epoch?: number }
+  // -- mount attach (resume) --
+  | { type: "ATTACH_START"; runId?: string }
+  | { type: "ATTACH_LIVE"; epoch?: number }
+  | { type: "ATTACH_NONE"; epoch?: number }
+  // -- reconnect after a live disconnect (entered by FINISH_DISCONNECT, #488 c2) --
+  | { type: "RECONNECT_ATTEMPT"; attempt: number; epoch?: number }
+  | { type: "RECONNECT_ATTACHED"; epoch?: number }
+  | { type: "RECONNECT_NONE"; epoch?: number }
+  | { type: "RETRY" }
+  // -- degraded poll --
+  | { type: "POLL_TERMINAL" }
+  | { type: "POLL_IDLE_CAP" }
+  // -- run-fact (server-confirmed active run) --
+  | { type: "RUN_FACT"; runFact: RunFact; epoch?: number }
+  // -- stop --
+  | { type: "STOP_REQUESTED" }
+  // -- supersede (CAS) --
+  | { type: "SUPERSEDE_REQUESTED"; targetRunId: string }
+  | { type: "SUPERSEDE_READY"; runId?: string; epoch?: number }
+  | { type: "SUPERSEDE_MISMATCH"; currentRunId?: string; epoch?: number }
+  | { type: "SUPERSEDE_TIMEOUT"; epoch?: number }
+  | { type: "SUPERSEDE_INVALID"; epoch?: number }
+  | { type: "RUN_ALREADY_ACTIVE"; activeRunId?: string }
+  // -- lifecycle --
+  | { type: "DISPOSE" };
+
+export const RECONNECT_MAX_ATTEMPTS = 5;
+export const RECONNECT_BASE_DELAY_MS = 1000;
+/** Backoff before attempt N (1-based): 1s, 2s, 4s, 8s, 16s. */
+export function reconnectDelayMs(attempt: number): number {
+  return RECONNECT_BASE_DELAY_MS * 2 ** (attempt - 1);
+}
+
+// ---------------------------------------------------------------------------
+// Constructors / helpers.
+// ---------------------------------------------------------------------------
+
+export function initialMachine(overrides?: Partial<Ctx>): Machine {
+  return {
+    phase: { name: "idle" },
+    ctx: { epoch: 0, ownership: "local", runFact: null, liveFollow: false, ...overrides },
+    effects: [],
+  };
+}
+
+/** Build a machine result: a phase, optional ctx patch, and effects. Empty
+ *  effects by default. Never mutates the input. */
+function to(
+  m: Machine,
+  phase: Phase,
+  opts?: { ctx?: Partial<Ctx>; effects?: Effect[] },
+): Machine {
+  return {
+    phase,
+    ctx: { ...m.ctx, ...(opts?.ctx ?? {}) },
+    effects: opts?.effects ?? [],
+  };
+}
+
+/** No transition: keep the phase, clear effects (so a re-run does not re-fire). */
+function stay(m: Machine): Machine {
+  return { phase: m.phase, ctx: m.ctx, effects: [] };
+}
+
+/** A command-transition: same as `to` but bumps the epoch (I1). Any outcome
+ *  event issued under the old epoch is dropped once this lands. */
+function command(
+  m: Machine,
+  phase: Phase,
+  effects: Effect[],
+  ctx?: Partial<Ctx>,
+): Machine {
+  return {
+    phase,
+    ctx: { ...m.ctx, ...(ctx ?? {}), epoch: m.ctx.epoch + 1 },
+    effects,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// The pure reducer.
+// ---------------------------------------------------------------------------
+
+/** The terminal stream-finish events (one turn's stream ended). */
+function isFinishEvent(event: Event): boolean {
+  return (
+    event.type === "FINISH_ABORT" ||
+    event.type === "FINISH_CLEAN" ||
+    event.type === "FINISH_DISCONNECT" ||
+    event.type === "FINISH_ERROR" ||
+    event.type === "STREAM_INCOMPLETE"
+  );
+}
+
+export function reduce(m: Machine, event: Event): Machine {
+  // MEDIUM (#488 re-review): honor ANY stream finish in `stopping` regardless of
+  // generation. A plain user Stop has NO successor stream — the aborted stream's
+  // finish IS the expected end of the stop, so exit `stopping -> idle` by that DATA
+  // (I4). The epoch filter below must NOT drop it: STOP_REQUESTED bumped the epoch,
+  // but the finish carries the PRE-stop generation (the runtime stamps it with the
+  // stream's start epoch), so I1 would otherwise strand the machine in `stopping`
+  // forever (no idle-cap covers `stopping`). The epoch filter stays in force for
+  // `superseding` (a successor B owns) — that is the F1 supersede drop.
+  if (m.phase.name === "stopping" && isFinishEvent(event)) {
+    return to(m, { name: "idle" }, {
+      // Reset ownership to local on this terminal transition (review #2): otherwise
+      // an observer-stop leaves ownership 'observer' and hides "Send now" forever.
+      ctx: { runFact: null, liveFollow: false, ownership: "local" },
+      effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+    });
+  }
+
+  // I1: drop a stale outcome (an event issued under a superseded epoch).
+  if ("epoch" in event && event.epoch !== undefined && event.epoch !== m.ctx.epoch) {
+    return stay(m);
+  }
+
+  switch (event.type) {
+    // ---- local turn ----------------------------------------------------
+    case "SEND_LOCAL":
+      // A local send owns the view: leave any recovery, become the local
+      // streamer, disarm poll/reconnect. epoch++ so a late recovery outcome
+      // from the previous phase is dropped.
+      return command(
+        m,
+        { name: "sending" },
+        [{ type: "cancelReconnect" }, { type: "disarmPoll" }],
+        { ownership: "local", liveFollow: false },
+      );
+
+    case "STREAM_INCOMPLETE":
+      // An OBSERVER's attached stream ended incomplete (starved / torn) — follow
+      // the run to terminal via the degraded poll.
+      return to(m, { name: "polling", reason: event.reason }, {
+        effects: [{ type: "armPoll", reason: event.reason }],
+      });
+
+    case "STREAM_START": {
+      // First frame arrived. Adopt the run-fact runId if present. sending ->
+      // streaming; a reconnect/attach that just went live also lands here.
+      const runFact = event.runId ? { runId: event.runId } : m.ctx.runFact;
+      return to(m, { name: "streaming" }, {
+        ctx: { runFact },
+        effects: [{ type: "cancelReconnect" }, { type: "disarmPoll" }],
+      });
+    }
+
+    case "FINISH_CLEAN":
+      // A clean terminal outcome. The run is done — clear the run-fact and go
+      // idle. (The queue flush is a component concern gated by ownership; the
+      // FSM only models the phase.) Review #2: reset ownership to local so a
+      // just-finished observer-attach turn re-exposes "Send now" for the queue.
+      return to(m, { name: "idle" }, {
+        ctx: { runFact: null, liveFollow: false, ownership: "local" },
+        effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+      });
+
+    case "FINISH_ABORT":
+      // A user Stop / intentional abort finished. If we were stopping, the
+      // terminal data has now arrived (I4) — go idle. The run-fact is cleared.
+      return to(m, { name: "idle" }, {
+        ctx: { runFact: null, liveFollow: false, ownership: "local" },
+        effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+      });
+
+    case "FINISH_DISCONNECT":
+      // A LIVE SSE drop. Recovery depends on WHO we are (I2 + liveFollow):
+      //  - a mount-attach OBSERVER (a one-shot resume, NOT live-follow) that drops
+      //    -> the degraded poll drives the row to terminal from the DB.
+      if (m.ctx.ownership === "observer" && !m.ctx.liveFollow) {
+        return to(m, { name: "polling", reason: "disconnect-visible" }, {
+          effects: [{ type: "armPoll", reason: "disconnect-visible" }],
+        });
+      }
+      //  - a LOCAL live turn (first drop) OR a live-follow re-attach (a SUBSEQUENT
+      //    drop) -> (re-)enter the reconnect ladder. #488 commit 3: allowed
+      //    REPEATEDLY — `liveFollow` is kept across a successful re-attach, so the
+      //    second break reconnects again instead of falling to silent poll.
+      // #488 commit 2: gated on the RUN-FACT (or an existing live-follow), NOT on
+      // the presence of an assistant message — a setup-phase break still recovers.
+      //  - visible content already on screen -> keep it, ALSO poll to terminal
+      //    (a full replay could clobber the fuller live tail);
+      //  - no visible content -> the reconnect ladder rebuilds it.
+      if (m.ctx.runFact || m.ctx.liveFollow) {
+        const effects: Effect[] = [
+          { type: "scheduleReconnect", attempt: 1, delayMs: reconnectDelayMs(1) },
+        ];
+        if (event.hasVisibleContent) effects.push({ type: "armPoll", reason: "disconnect-visible" });
+        return command(m, { name: "reconnecting", attempt: 1, failed: false }, effects, {
+          ownership: "observer",
+          liveFollow: true,
+        });
+      }
+      // No run to recover: a plain disconnect. Surface the terminal notice.
+      return to(m, { name: "idle" }, {
+        ctx: { runFact: null, liveFollow: false, ownership: "local" },
+      });
+
+    case "FINISH_ERROR":
+      return to(m, { name: "error", kind: event.kind }, {
+        ctx: { runFact: null, liveFollow: false, ownership: "local" },
+        effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+      });
+
+    // ---- mount attach (resume) ----------------------------------------
+    case "ATTACH_START":
+      // A reopened tab attaches to a still-running run: observer ownership.
+      // #488 F2: ONLY from idle. The mount `getRun` round-trip resolves async, and
+      // a local send may have started meanwhile (phase `sending`, ownership local);
+      // a late ATTACH_START must NOT hijack that local turn into an observer-attach
+      // (queue would stop flushing, "Send now" would hide). Guarding in the reducer
+      // covers every dispatch source.
+      if (m.phase.name !== "idle") return stay(m);
+      return command(m, { name: "attaching" }, [{ type: "resumeStream" }], {
+        ownership: "observer",
+        runFact: event.runId ? { runId: event.runId } : m.ctx.runFact,
+      });
+
+    case "ATTACH_LIVE":
+      // The attach GET returned a live 2xx stream — follow it as an observer.
+      // Review #1: guard by SOURCE phase. The epoch filter alone is not enough — a
+      // POLL_TERMINAL uses to() (no epoch bump) and does not abort the in-flight
+      // GET, so a slow 2xx landing after the machine already left `attaching` (e.g.
+      // the armed poll saw the terminal row -> idle) would resurrect a settled run
+      // into a phantom `streaming`. Only enter streaming FROM `attaching`.
+      if (m.phase.name !== "attaching") return stay(m);
+      return to(m, { name: "streaming" });
+
+    case "ATTACH_NONE":
+      // 204 / non-2xx / throw: nothing live to attach. Arm the degraded poll to
+      // follow the run to terminal from the DB. This is a soft-negative run-fact
+      // (204 on a non-stripped path is authoritative-negative; the runtime may
+      // pass a RUN_FACT null separately). Keep the run-fact as-is here.
+      // Review #1: guard by source phase for consistency (a late outcome after the
+      // machine already left `attaching` must not re-arm a poll).
+      if (m.phase.name !== "attaching") return stay(m);
+      return to(m, { name: "polling", reason: "attach-none" }, {
+        effects: [{ type: "armPoll", reason: "attach-none" }],
+      });
+
+    // ---- reconnect after a live disconnect ----------------------------
+    case "RECONNECT_ATTEMPT":
+      // A scheduled backoff fired — fire the attach GET. epoch++ so the previous
+      // attempt's late outcome cannot drive this one.
+      if (m.phase.name !== "reconnecting") return stay(m);
+      return command(
+        m,
+        { name: "reconnecting", attempt: event.attempt, failed: false },
+        [{ type: "resumeStream" }],
+      );
+
+    case "RECONNECT_ATTACHED":
+      // #488 commit 3: a live re-attach succeeded. Reset to streaming — the
+      // attempt counter is dropped, so a LATER disconnect can start a fresh
+      // ladder from attempt 1 (the old one-shot `!wasResumed` gate forbade a
+      // second cycle, sending the second break to silent poll).
+      // Review #1: guard by SOURCE phase. The armed degraded poll can reach the
+      // terminal row (POLL_TERMINAL -> idle, via to(), NO epoch bump, GET not
+      // aborted) BEFORE a slow reconnect GET returns 2xx; without this guard that
+      // late RECONNECT_ATTACHED (same epoch) would resurrect a settled run into a
+      // phantom `streaming`. Only re-enter streaming FROM `reconnecting`.
+      if (m.phase.name !== "reconnecting") return stay(m);
+      return to(m, { name: "streaming" }, {
+        effects: [{ type: "cancelReconnect" }, { type: "disarmPoll" }],
+      });
+
+    case "RECONNECT_NONE": {
+      // 204 / error during a reconnect attempt. Arm the degraded poll as the
+      // belt-and-suspenders fallback, then either back off to the next attempt
+      // or, at the cap, surface the manual Retry ("failed").
+      if (m.phase.name !== "reconnecting") return stay(m);
+      const attempt = m.phase.attempt;
+      if (attempt < RECONNECT_MAX_ATTEMPTS) {
+        return command(
+          m,
+          { name: "reconnecting", attempt: attempt + 1, failed: false },
+          [
+            { type: "armPoll", reason: "attach-none" },
+            { type: "scheduleReconnect", attempt: attempt + 1, delayMs: reconnectDelayMs(attempt + 1) },
+          ],
+        );
+      }
+      return to(m, { name: "reconnecting", attempt, failed: true }, {
+        effects: [{ type: "armPoll", reason: "reconnect-exhausted" }],
+      });
+    }
+
+    case "RETRY":
+      // Manual Retry from the "failed" reconnect banner OR the stalled banner.
+      if (m.phase.name === "reconnecting" && m.phase.failed) {
+        return command(
+          m,
+          { name: "reconnecting", attempt: 1, failed: false },
+          [{ type: "resumeStream" }],
+        );
+      }
+      if (m.phase.name === "stalled") {
+        // Re-arm the poll to try to catch the run up again.
+        return command(m, { name: "polling", reason: "attach-none" }, [
+          { type: "armPoll", reason: "attach-none" },
+        ]);
+      }
+      return stay(m);
+
+    // ---- degraded poll -------------------------------------------------
+    case "POLL_TERMINAL":
+      // The run reached a terminal row via the poll (or the reconcile merge). Go
+      // idle and disarm everything (I4: this is a DATA-driven exit, incl. exit
+      // from `stopping`). Review #2: reset ownership to local.
+      return to(m, { name: "idle" }, {
+        ctx: { runFact: null, liveFollow: false, ownership: "local" },
+        effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+      });
+
+    case "POLL_IDLE_CAP":
+      // Review #4: `stopping` also arms the poll (STOP_REQUESTED) but has NO other
+      // backstop — an observer-stop with no SDK stream to fire onFinish, whose
+      // server stop never drives the run terminal, would poll the DB forever. Give
+      // it a bounded exit: cap -> idle + disarm (NOT `stalled`; Stop was already
+      // pressed, so there is nothing for the user to retry).
+      if (m.phase.name === "stopping") {
+        return to(m, { name: "idle" }, {
+          ctx: { runFact: null, liveFollow: false, ownership: "local" },
+          effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+        });
+      }
+      // #488 commit 4a: the poll hit the inactivity cap. Instead of going SILENT
+      // (the old "forever half-done answer"), surface a stalled banner + Retry.
+      if (m.phase.name !== "polling" && m.phase.name !== "reconnecting") return stay(m);
+      return to(m, { name: "stalled" }, {
+        effects: [{ type: "disarmPoll" }, { type: "cancelReconnect" }],
+      });
+
+    // ---- run-fact ------------------------------------------------------
+    case "RUN_FACT": {
+      const runFact = event.runFact;
+      // A fresh NEGATIVE fact (no active run) cancels recovery immediately (I3):
+      // there is nothing to reconnect to / poll for.
+      if (!runFact) {
+        if (
+          m.phase.name === "reconnecting" ||
+          m.phase.name === "attaching" ||
+          m.phase.name === "polling" ||
+          m.phase.name === "stopping"
+        ) {
+          return to(m, { name: "idle" }, {
+            // Review #2: reset ownership to local on this terminal transition.
+            ctx: { runFact: null, liveFollow: false, ownership: "local" },
+            effects: [{ type: "cancelReconnect" }, { type: "disarmPoll" }],
+          });
+        }
+        return to(m, m.phase, { ctx: { runFact: null } });
+      }
+      // A positive fact just updates the context (pessimism toward an attempt: a
+      // stale-but-positive fact permits entering recovery; a 204 will cut it).
+      return to(m, m.phase, { ctx: { runFact } });
+    }
+
+    // ---- stop ----------------------------------------------------------
+    case "STOP_REQUESTED":
+      // Authoritative stop of a detached run. Enter `stopping` and fire stopRun +
+      // abort the local/attach reader. ALSO arm the poll so the terminal row is
+      // observed — the exit is by DATA (I4: a terminal row / negative run-fact),
+      // never by the stopRun HTTP response (which returns after abort, before
+      // finalization). For a local turn the aborted stream's onFinish (ANY finish)
+      // is HONORED in `stopping` at the top of reduce() — regardless of generation
+      // — and exits to idle; the armed poll is the fallback for an observer stop
+      // with no local onFinish.
+      return command(
+        m,
+        { name: "stopping" },
+        [
+          { type: "stopRun" },
+          { type: "abortAttach" },
+          { type: "cancelReconnect" },
+          { type: "armPoll", reason: "attach-none" },
+        ],
+      );
+
+    // ---- supersede (CAS) ----------------------------------------------
+    case "SUPERSEDE_REQUESTED":
+      // "Interrupt and send now": CAS POST /stream { supersede }. epoch++ so a
+      // late outcome of the interrupted run is dropped.
+      return command(
+        m,
+        { name: "superseding" },
+        [{ type: "supersede", targetRunId: event.targetRunId }, { type: "cancelReconnect" }, { type: "disarmPoll" }],
+      );
+
+    case "SUPERSEDE_READY": {
+      // CAS succeeded (old run stopped/settled, slot taken, new run begun). We
+      // are now the local streamer of the NEW run. Adopt its runId if provided.
+      const runFact = event.runId ? { runId: event.runId } : m.ctx.runFact;
+      return to(m, { name: "streaming" }, {
+        ctx: { ownership: "local", runFact, liveFollow: false },
+      });
+    }
+
+    case "SUPERSEDE_MISMATCH":
+      // The active run moved between the click and the CAS. Per the spec: verify
+      // via /run rather than blindly banner — the mismatch may be our own already-
+      // superseded run. Surface a classified error AND fire a run-fact verify.
+      return to(m, { name: "error", kind: "supersede-mismatch" }, {
+        ctx: { runFact: event.currentRunId ? { runId: event.currentRunId } : m.ctx.runFact },
+        effects: [{ type: "postRun", reason: "verify" }],
+      });
+
+    case "SUPERSEDE_TIMEOUT":
+      // The old run did not settle within W. Nothing persisted; the composer keeps
+      // its text. Classified error, NO auto-retry (the old client retry ladder is
+      // removed in #488 commit 5).
+      return to(m, { name: "error", kind: "supersede-timeout" });
+
+    case "SUPERSEDE_INVALID":
+      return to(m, { name: "error", kind: "supersede-invalid" });
+
+    case "RUN_ALREADY_ACTIVE":
+      // A plain POST hit the one-active-run gate. NO auto-retry — the composer
+      // offers "interrupt and send" (supersede) instead. #497/S4: adopt the
+      // server's activeRunId as the run-fact so that supersede can TARGET the
+      // (possibly foreign-tab) active run via the CAS, rather than a blind
+      // promote+abort that just 409s again. A stale/absent id keeps the prior fact.
+      return to(m, { name: "error", kind: "run-already-active" }, {
+        ctx: { runFact: event.activeRunId ? { runId: event.activeRunId } : m.ctx.runFact },
+      });
+
+    // ---- lifecycle -----------------------------------------------------
+    case "DISPOSE":
+      // Unmount: abort in-flight controllers, drop timers, and bump the epoch so
+      // NO late callback can drive this (now dead) machine (I5).
+      return command(
+        m,
+        { name: "idle" },
+        [
+          { type: "abortAttach" },
+          { type: "cancelReconnect" },
+          { type: "disarmPoll" },
+        ],
+        { liveFollow: false },
+      );
+
+    default: {
+      // Exhaustiveness guard.
+      const _never: never = event;
+      void _never;
+      return stay(m);
+    }
+  }
+}

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

@@ -181,6 +181,12 @@ export interface IAiChatMessageRow {
   toolCalls?: unknown;
   metadata?: {
     parts?: UIMessage["parts"];
+    // #491 step-alignment anchor: the count of FINISHED steps whose parts are in
+    // THIS row, written atomically with `parts` server-side (flushAssistant). The
+    // resume client reads it as its persisted step frontier N — the tail-only
+    // attach asks the run-stream registry for the frames of step N onward (the
+    // seed already carries steps 0..N-1). Absent on pre-#491 rows -> read as 0.
+    stepsPersisted?: number;
     // AI SDK v6 `totalUsage` persisted on assistant rows. Legacy cumulative
     // figure (sum of every step's usage for the turn); kept for back-compat and
     // as the fallback for older rows that have no `contextTokens`.

apps/client/src/features/ai-chat/utils/adopt-chat-id.test.ts

@@ -3,6 +3,7 @@ import {
   resolveAdoptedChatId,
   newlyAddedChatIds,
   extractServerChatId,
+  extractRunId,
 } from "./adopt-chat-id";
 
 describe("resolveAdoptedChatId", () => {
@@ -70,3 +71,17 @@ describe("extractServerChatId", () => {
     expect(extractServerChatId(undefined)).toBeUndefined();
   });
 });
+
+describe("extractRunId", () => {
+  it("reads a string runId from the start metadata", () => {
+    expect(extractRunId({ metadata: { runId: "run-1" } })).toBe("run-1");
+  });
+  it("returns undefined when runId is absent", () => {
+    expect(extractRunId({ metadata: { chatId: "c" } })).toBeUndefined();
+    expect(extractRunId({})).toBeUndefined();
+    expect(extractRunId(undefined)).toBeUndefined();
+  });
+  it("returns undefined for a non-string runId", () => {
+    expect(extractRunId({ metadata: { runId: 7 } })).toBeUndefined();
+  });
+});

apps/client/src/features/ai-chat/utils/adopt-chat-id.ts

@@ -56,6 +56,20 @@ export function extractServerChatId(
   return typeof m?.chatId === "string" ? m.chatId : undefined;
 }
 
+/**
+ * #488: read the authoritative RUN id off a streaming assistant message. The
+ * server attaches it as `message.metadata.runId` on the `start` part when a run
+ * wraps the turn (see server `chatStreamMetadata`, #184/#487). This is the live
+ * run-fact update the client FSM adopts (mirrors `extractServerChatId`). Returns
+ * it only when it is a string; undefined otherwise.
+ */
+export function extractRunId(
+  message: { metadata?: unknown } | undefined,
+): string | undefined {
+  const m = message?.metadata as { runId?: string } | undefined;
+  return typeof m?.runId === "string" ? m.runId : undefined;
+}
+
 /**
  * The deduped set of ids present in `afterIds` but not in `beforeIds`. A
  * paginated/flatMapped list can repeat the same id, so dedupe: one genuinely-new

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

@@ -42,6 +42,53 @@ describe("describeChatError", () => {
     );
   });
 
+  // #488 commit 5: the #487 concurrency-gate / supersede 409s. FULL real bodies:
+  // a ConflictException(object) whose response is serialized verbatim, carrying a
+  // `code` and statusCode 409. Each must classify to a human text, not raw JSON.
+  it("classifies A_RUN_ALREADY_ACTIVE (409) as already-answering, not raw JSON", () => {
+    const body =
+      '{"message":"A run is already active for this chat","code":"A_RUN_ALREADY_ACTIVE","statusCode":409}';
+    expect(describeChatError(body, t).title).toBe(
+      "The agent is already answering",
+    );
+    // Never leaks the raw code as the detail.
+    expect(describeChatError(body, t).detail).not.toContain("A_RUN_ALREADY_ACTIVE");
+  });
+
+  it("classifies SUPERSEDE_TARGET_MISMATCH (409) as run-changed", () => {
+    // Real server body shape: the current run id is `activeRunId` (NOT `runId`) —
+    // see ai-chat.controller.ts. describeChatError classifies off `code` only.
+    const body =
+      '{"message":"active run does not match the supersede target","code":"SUPERSEDE_TARGET_MISMATCH","activeRunId":"run-x","statusCode":409}';
+    expect(describeChatError(body, t).title).toBe(
+      "Couldn't interrupt — the run changed",
+    );
+  });
+
+  it("classifies SUPERSEDE_TIMEOUT (409) as couldn't-interrupt-in-time", () => {
+    const body =
+      '{"message":"the run did not settle within the supersede window","code":"SUPERSEDE_TIMEOUT","statusCode":409}';
+    expect(describeChatError(body, t).title).toBe("Couldn't interrupt in time");
+  });
+
+  it("classifies SUPERSEDE_INVALID (409) as couldn't-interrupt-that-run", () => {
+    const body =
+      '{"message":"supervise requires chatId","code":"SUPERSEDE_INVALID","statusCode":409}';
+    expect(describeChatError(body, t).title).toBe(
+      "Couldn't interrupt that run",
+    );
+  });
+
+  it("ORDER GUARD: A_RUN_ALREADY_ACTIVE wins over any generic status branch", () => {
+    // Even though the body could superficially look 4xx-ish, the code branch runs
+    // first, so it is never mislabeled by a generic status heading.
+    const body =
+      '{"message":"conflict","code":"A_RUN_ALREADY_ACTIVE","statusCode":409}';
+    const view = describeChatError(body, t);
+    expect(view.title).not.toBe("Something went wrong");
+    expect(view.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

@@ -39,6 +39,44 @@ export function describeChatError(
     };
   }
 
+  // #488 commit 5: the #487 concurrency-gate / supersede 409s. These arrive as a
+  // ConflictException(object) body carrying a `code` (and statusCode 409). They
+  // MUST be classified by `code` STRICTLY BEFORE any generic status branch, or the
+  // user sees the raw JSON `{"code":"A_RUN_ALREADY_ACTIVE",…}`. The code strings
+  // are the real #487 server contract (ai-chat.controller.ts) — do not invent.
+  if (/"code"\s*:\s*"A_RUN_ALREADY_ACTIVE"/.test(msg)) {
+    return {
+      title: t("The agent is already answering"),
+      detail: t(
+        "This chat already has a run in progress. Wait for it to finish, or interrupt it and send now.",
+      ),
+    };
+  }
+  if (/"code"\s*:\s*"SUPERSEDE_TARGET_MISMATCH"/.test(msg)) {
+    return {
+      title: t("Couldn't interrupt — the run changed"),
+      detail: t(
+        "The run you tried to interrupt is no longer the active one. Check the latest answer and try again.",
+      ),
+    };
+  }
+  if (/"code"\s*:\s*"SUPERSEDE_TIMEOUT"/.test(msg)) {
+    return {
+      title: t("Couldn't interrupt in time"),
+      detail: t(
+        "The previous run didn't stop in time. Nothing was sent — try sending again.",
+      ),
+    };
+  }
+  if (/"code"\s*:\s*"SUPERSEDE_INVALID"/.test(msg)) {
+    return {
+      title: t("Couldn't interrupt that run"),
+      detail: t(
+        "The run to interrupt doesn't belong to this chat. Reload and try again.",
+      ),
+    };
+  }
+
   if (/"statusCode"\s*:\s*403\b/.test(msg)) {
     return {
       title: t("AI chat is disabled"),

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

@@ -4,7 +4,8 @@ import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.t
 import {
   isStreamingTail,
   isSettledAssistantTail,
-  seedRows,
+  stepsPersistedOf,
+  mergeDeltaRowsIntoPages,
   mergeById,
 } from "./resume-helpers.ts";
 
@@ -12,8 +13,18 @@ function row(
   id: string,
   role: string,
   status?: string,
+  stepsPersisted?: number,
 ): IAiChatMessageRow {
-  return { id, role, content: "", status, createdAt: "2026-01-01T00:00:00Z" };
+  return {
+    id,
+    role,
+    content: "",
+    status,
+    createdAt: "2026-01-01T00:00:00Z",
+    ...(stepsPersisted !== undefined
+      ? { metadata: { stepsPersisted } }
+      : {}),
+  };
 }
 
 function makeMsg(id: string, text: string): UIMessage {
@@ -65,23 +76,92 @@ describe("isSettledAssistantTail", () => {
   });
 });
 
-describe("seedRows", () => {
-  const rows = [row("u1", "user"), row("a1", "assistant", "streaming")];
-
-  it("returns the rows unchanged when not stripping", () => {
-    expect(seedRows(rows, false)).toBe(rows);
+describe("stepsPersistedOf", () => {
+  it("reads metadata.stepsPersisted", () => {
+    expect(stepsPersistedOf(row("a1", "assistant", "streaming", 3))).toBe(3);
+    expect(stepsPersistedOf(row("a1", "assistant", "streaming", 0))).toBe(0);
   });
 
-  it("drops the last row when stripping", () => {
-    const seeded = seedRows(rows, true);
-    expect(seeded).toHaveLength(1);
-    expect(seeded[0].id).toBe("u1");
+  it("defaults to 0 for a pre-#491 row (absent), null/undefined, or a bad value", () => {
+    expect(stepsPersistedOf(row("a1", "assistant", "streaming"))).toBe(0);
+    expect(stepsPersistedOf(null)).toBe(0);
+    expect(stepsPersistedOf(undefined)).toBe(0);
+    expect(
+      stepsPersistedOf({
+        id: "a1",
+        role: "assistant",
+        content: "",
+        createdAt: "x",
+        metadata: { stepsPersisted: -2 },
+      }),
+    ).toBe(0);
   });
 
-  it("returns an empty list when stripping a single-row list", () => {
-    expect(seedRows([row("a1", "assistant", "streaming")], true)).toHaveLength(
-      0,
-    );
+  it("floors a non-integer count", () => {
+    expect(
+      stepsPersistedOf({
+        id: "a1",
+        role: "assistant",
+        content: "",
+        createdAt: "x",
+        metadata: { stepsPersisted: 2.9 },
+      }),
+    ).toBe(2);
+  });
+});
+
+describe("mergeDeltaRowsIntoPages", () => {
+  const pages = () => [
+    { items: [row("u1", "user"), row("a1", "assistant", "streaming", 1)], meta: {} },
+  ];
+
+  it("returns the pages unchanged for an empty delta", () => {
+    const p = pages();
+    expect(mergeDeltaRowsIntoPages(p, [])).toBe(p);
+  });
+
+  it("appends a genuinely new row to the last page in chronological order", () => {
+    const merged = mergeDeltaRowsIntoPages(pages(), [row("a2", "assistant", "streaming", 0)]);
+    expect(merged[0].items.map((i) => i.id)).toEqual(["u1", "a1", "a2"]);
+  });
+
+  it("replaces a grown row in place (per-step growth), never appends a duplicate", () => {
+    const merged = mergeDeltaRowsIntoPages(pages(), [
+      row("a1", "assistant", "streaming", 2),
+    ]);
+    expect(merged[0].items.map((i) => i.id)).toEqual(["u1", "a1"]);
+    // the in-place replacement carries the grown step frontier.
+    expect(stepsPersistedOf(merged[0].items[1])).toBe(2);
+  });
+
+  it("does not mutate the input pages", () => {
+    const input = pages();
+    const before = input[0].items.slice();
+    mergeDeltaRowsIntoPages(input, [row("a2", "assistant", "streaming", 0)]);
+    expect(input[0].items).toEqual(before); // untouched
+  });
+
+  // #491 CONTRACT: the delta overlap window re-delivers the same rows, so merging
+  // MUST be idempotent — applying a delta twice equals applying it once (no growth,
+  // no reorder). A regression re-introduces duplicate assistant bubbles per poll.
+  it("is idempotent: applying the same delta twice equals once", () => {
+    const delta = [
+      row("a1", "assistant", "streaming", 2), // grown existing row
+      row("a2", "assistant", "streaming", 0), // new row
+    ];
+    const once = mergeDeltaRowsIntoPages(pages(), delta);
+    const twice = mergeDeltaRowsIntoPages(once, delta);
+    const thrice = mergeDeltaRowsIntoPages(twice, delta);
+    expect(once[0].items.map((i) => i.id)).toEqual(["u1", "a1", "a2"]);
+    expect(twice[0].items.map((i) => i.id)).toEqual(["u1", "a1", "a2"]);
+    expect(twice).toEqual(once);
+    expect(thrice).toEqual(once);
+  });
+
+  it("seeds a first page when the cache is empty", () => {
+    const merged = mergeDeltaRowsIntoPages([], [row("u1", "user")]);
+    expect(merged).toHaveLength(1);
+    expect(merged[0].items.map((i) => i.id)).toEqual(["u1"]);
   });
 });
 
@@ -109,4 +189,37 @@ describe("mergeById", () => {
     expect(mergeById(prev, null)).toBe(prev);
     expect(mergeById(prev, undefined)).toBe(prev);
   });
+
+  // #491 CONTRACT: the delta poll's overlap window GUARANTEES the same row is
+  // re-delivered across close polls, so merging must be IDEMPOTENT by id — merging
+  // the same row (or an equal-length list of rows) twice must not duplicate or
+  // reorder. This is the property the whole delta-poll design leans on; a
+  // regression here would re-introduce duplicate assistant bubbles on every poll.
+  it("is idempotent by id: re-merging the same row does not duplicate or reorder", () => {
+    const seed = [makeMsg("u1", "hi"), makeMsg("a1", "step 1")];
+    const repeat = makeMsg("a1", "step 1"); // the SAME row the overlap re-delivers
+    const once = mergeById(seed, repeat);
+    const twice = mergeById(once, repeat);
+    const thrice = mergeById(twice, repeat);
+    // Length is stable (no growth), order is stable (user then assistant).
+    expect(once.map((m) => m.id)).toEqual(["u1", "a1"]);
+    expect(twice.map((m) => m.id)).toEqual(["u1", "a1"]);
+    expect(thrice.map((m) => m.id)).toEqual(["u1", "a1"]);
+    // The repeated merge converges: the row is replaced in place, never appended.
+    expect(twice[1]).toBe(repeat);
+  });
+
+  it("is idempotent across a batch of repeated + grown rows (delta re-delivery)", () => {
+    // A delta poll re-delivers a1 (unchanged) and a2 (grown one step). Applying the
+    // batch twice must equal applying it once — the poll can re-send either.
+    const start = [makeMsg("u1", "hi"), makeMsg("a1", "done")];
+    const batch = [makeMsg("a1", "done"), makeMsg("a2", "grown step 2")];
+    const apply = (list: typeof start) =>
+      batch.reduce((acc, row) => mergeById(acc, row), list);
+    const once = apply(start);
+    const twice = apply(once);
+    expect(once.map((m) => m.id)).toEqual(["u1", "a1", "a2"]);
+    expect(twice.map((m) => m.id)).toEqual(["u1", "a1", "a2"]);
+    expect(twice).toEqual(once);
+  });
 });

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

@@ -11,9 +11,10 @@ import type { IAiChatMessageRow } from "@/features/ai-chat/types/ai-chat.types.t
 
 /**
  * A STREAMING tail: the last persisted row is an assistant row still marked
- * `status === 'streaming'`. Such a tail is stripped from the seed and rebuilt by
- * the replay (`expect=live`), since the SDK's `text-start` always pushes a new
- * part and replaying over a seeded in-progress row would duplicate its text.
+ * `status === 'streaming'`. #491 (tail-only): such a tail is seeded UNCHANGED —
+ * it carries the persisted steps 0..N-1 — and the run-stream registry's tail
+ * (frames for steps >= N) is APPENDED to it by the SDK's `readUIMessageStream`
+ * continuation. Only the presence of this tail decides WHETHER to attach.
  */
 export function isStreamingTail(rows: IAiChatMessageRow[]): boolean {
   const tail = rows[rows.length - 1];
@@ -32,15 +33,61 @@ export function isSettledAssistantTail(rows: IAiChatMessageRow[]): boolean {
 }
 
 /**
- * Seed rows for `useChat`: return the rows unchanged, or without the last row when
- * `strip` is set (the streaming tail is stripped so the live replay rebuilds it
- * without duplicating parts).
+ * #491 tail-only anchor: the count of FINISHED steps whose parts are persisted in
+ * THIS assistant row (`metadata.stepsPersisted`), written atomically with `parts`
+ * server-side. The resume client reads it as its persisted step frontier N — the
+ * tail-only attach asks the run-stream registry for the frames of step N onward
+ * (the seed already carries steps 0..N-1). Absent on pre-#491 rows => 0.
  */
-export function seedRows(
+export function stepsPersistedOf(
+  row: IAiChatMessageRow | null | undefined,
+): number {
+  const n = row?.metadata?.stepsPersisted;
+  return typeof n === "number" && n >= 0 ? Math.floor(n) : 0;
+}
+
+/** One page of the messages infinite-query cache (`{ items, meta }`). */
+export interface IMessagePage {
+  items: IAiChatMessageRow[];
+  meta: unknown;
+}
+
+/**
+ * #491 delta-poll merge: upsert the delta poll's `rows` into the messages
+ * infinite-query page structure IDEMPOTENTLY by id. The delta endpoint's overlap
+ * window GUARANTEES occasional REPEATS, so this MUST converge: a row already
+ * present is REPLACED IN PLACE (per-step growth of an in-progress row), a new row
+ * is APPENDED to the last page in chronological order (the server returns delta
+ * rows oldest-first). Applying the same delta twice equals applying it once. Never
+ * mutates the input pages (returns fresh page objects with cloned item arrays).
+ */
+export function mergeDeltaRowsIntoPages(
+  pages: IMessagePage[],
   rows: IAiChatMessageRow[],
-  strip: boolean,
-): IAiChatMessageRow[] {
-  return strip ? rows.slice(0, -1) : rows;
+): IMessagePage[] {
+  if (rows.length === 0) return pages;
+  const next: IMessagePage[] = pages.map((p) => ({
+    ...p,
+    items: p.items.slice(),
+  }));
+  const locate = (id: string): [number, number] | null => {
+    for (let pi = 0; pi < next.length; pi++) {
+      const ii = next[pi].items.findIndex((it) => it.id === id);
+      if (ii !== -1) return [pi, ii];
+    }
+    return null;
+  };
+  for (const row of rows) {
+    const at = locate(row.id);
+    if (at) {
+      next[at[0]].items[at[1]] = row; // replace in place — idempotent by id
+    } else if (next.length > 0) {
+      next[next.length - 1].items.push(row); // append chronologically
+    } else {
+      next.push({ items: [row], meta: undefined });
+    }
+  }
+  return next;
 }
 
 /**

apps/client/src/features/ai-chat/utils/sdk-continuation-tripwire.test.ts

@@ -0,0 +1,83 @@
+import { describe, it, expect } from "vitest";
+import { readUIMessageStream, type UIMessage } from "ai";
+import pkg from "../../../../package.json";
+
+/**
+ * PIN-SPEC TRIP-WIRE (#491). The tail-only attach continuation relies on THREE
+ * behaviors of `ai@6.0.207`, verified line-by-line in the issue. Without this
+ * test, an `ai` bump could silently break attach (the client would append the
+ * live tail to the wrong message, or duplicate a step):
+ *
+ *   1. `readUIMessageStream({ message })` CONTINUES the passed message — it does
+ *      not start a fresh one — so the tail streamed after a re-seed is appended to
+ *      the seeded assistant row (the same DB id).
+ *   2. A `start` frame does NOT reset the existing message's parts (so the seeded
+ *      steps 0..N-1 survive; the synthetic `start` the registry prepends only
+ *      carries the run-fact metadata).
+ *   3. Text parts do NOT cross a `finish-step` boundary — a new `text-start` after
+ *      `finish-step` is a NEW part — so the reconstructed steps stay separated and
+ *      the step frontier stays meaningful.
+ *
+ * If an `ai` upgrade changes any of these, this test fails LOUD instead of the
+ * resume path silently corrupting.
+ */
+describe("ai SDK continuation trip-wire (#491, tail-only attach)", () => {
+  it("is pinned to the exact ai version the continuation was verified against", () => {
+    // A caret/range bump is exactly what would silently break attach — require an
+    // exact pin. Bumping ai MUST re-verify the behavior asserted below, then this.
+    expect((pkg as { dependencies: Record<string, string> }).dependencies.ai).toBe(
+      "6.0.207",
+    );
+  });
+
+  it("continues the seeded message: start does not reset parts, the tail appends as new parts", async () => {
+    // A seeded assistant row with ONE finished step already reconstructed.
+    const seeded: UIMessage = {
+      id: "assistant-1",
+      role: "assistant",
+      parts: [
+        { type: "step-start" },
+        { type: "text", text: "STEP0", state: "done" },
+      ],
+    } as UIMessage;
+
+    // The tail the registry delivers on re-attach: a synthetic start (run-fact),
+    // then step 1's frames, then finish. As UI-message chunks (what the SSE frames
+    // decode to).
+    const chunks = [
+      { type: "start", messageMetadata: { runId: "r1", chatId: "c1" } },
+      { type: "start-step" },
+      { type: "text-start", id: "t1" },
+      { type: "text-delta", id: "t1", delta: "STEP1" },
+      { type: "text-end", id: "t1" },
+      { type: "finish-step" },
+      { type: "finish" },
+    ];
+    const stream = new ReadableStream({
+      start(c) {
+        for (const ch of chunks) c.enqueue(ch);
+        c.close();
+      },
+    });
+
+    let last: UIMessage | undefined;
+    for await (const msg of readUIMessageStream({ message: seeded, stream })) {
+      last = msg;
+    }
+
+    expect(last).toBeDefined();
+    // Same message id (continuation, not a fresh message).
+    expect(last!.id).toBe("assistant-1");
+    // The seeded step-0 parts SURVIVED the `start` frame, and step 1 was appended
+    // as SEPARATE parts (text did not cross the finish-step boundary).
+    const shape = last!.parts.map((p) => `${p.type}:${(p as { text?: string }).text ?? ""}`);
+    expect(shape).toEqual([
+      "step-start:",
+      "text:STEP0",
+      "step-start:",
+      "text:STEP1",
+    ]);
+    // The run-fact metadata from the synthetic start frame is applied.
+    expect(last!.metadata).toMatchObject({ runId: "r1", chatId: "c1" });
+  });
+});

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-stream-registry.service.ts

@@ -1,42 +1,122 @@
 import { Injectable, Logger, OnModuleDestroy } from '@nestjs/common';
 
 /**
- * In-memory run-stream registry (#184 phase 1.5). A durable agent run tees its
- * SSE frames here (via `pipeUIMessageStreamToResponse({ consumeSseStream })`)
- * so a LATE tab — one that reloaded, or opened after the starter dropped — can
- * attach through `GET /ai-chat/runs/:chatId/stream`, replay the frames buffered
- * so far, and then follow the live tail as a normal streamer.
+ * In-memory run-stream registry (#184 phase 1.5, step-aligned retention #491). A
+ * durable agent run tees its SSE frames here (via
+ * `pipeUIMessageStreamToResponse({ consumeSseStream })`) so a LATE tab — one that
+ * reloaded, or opened after the starter dropped — can attach through
+ * `GET /ai-chat/runs/:chatId/stream`, be handed the TAIL past the step it already
+ * has persisted, and then follow the live tail as a normal streamer.
  *
  * This is deliberately single-process and best-effort: it holds nothing the DB
  * does not (the run + assistant row are the source of truth), so a process
  * restart simply drops in-flight entries and the client falls back to its
  * restore + degraded-poll path. The async `attach` return type is the seam for a
  * future phase-2 cross-process backend (Redis) — the interface does not change.
+ *
+ * ── #491 step-aligned retention (the OOM fix) ────────────────────────────────
+ * The old registry buffered up to 32MB of raw SSE frames PER active run (V8 ~2×
+ * in memory) and, on attach, blasted the WHOLE buffer to the socket synchronously
+ * with no drain — a handful of marathon runs on a 1GB container OOM'd. #491 caps
+ * the ring at a few MB (env-tunable, default 4MB) and keeps it there by ROTATING:
+ *
+ *  - Every buffered frame is STAMPED with a step number at tee (see ingestFrame).
+ *    Convention: the stamp of a frame is the number of `finish-step` parts seen
+ *    BEFORE it (starting at 0). The finish-step frame itself carries the current
+ *    value, THEN the counter increments. So a frame stamped `s` is the content of
+ *    the (s+1)-th step — 0-based step index `s` — and the stamp aligns EXACTLY
+ *    with `metadata.stepsPersisted`: a client whose persisted `stepsPersisted` is
+ *    N has steps 0..N-1 on disk (and in its seed) and needs the tail `stamp >= N`.
+ *
+ *  - The ring rotates ONLY on a CONFIRMED persist of step N
+ *    (`confirmPersistedStep`), dropping frames with `stamp < N` (those steps are
+ *    now on disk and a fresh client seed carries them). A NON-confirmed step is
+ *    never rotated away, so a persist FAILURE just makes the ring cover MORE
+ *    (auto-safe). This is the anti-inversion rule: a naive "rotate in .then()"
+ *    that rotated after an UNwritten step would drop a step nobody has → silent
+ *    hole. Rotation is gated on a real, successful persist.
+ *
+ *  - If the ring still exceeds its byte cap after rotation (a single fat step, or
+ *    a lagging persist), the OLDEST frames are evicted to stay bounded. Evicting a
+ *    not-yet-persisted frame opens a GAP: an attach whose N falls at or below an
+ *    evicted step answers 204 and the client degrades to restore+poll. The gap is
+ *    NOT sticky — the coverage floor is recomputed from the ring, so a later
+ *    persist that rotates past the holey steps clears it.
+ *
+ * ── attach numbering / coverage (the wire convention) ────────────────────────
+ * The step marker N comes ONLY FROM THE CLIENT (a query param). The server never
+ * reads the row to derive N — a server-side N from a stale seed would open a
+ * silent one-step hole. N is the client's persisted `stepsPersisted` (a COUNT):
+ *   - the tail it needs = frames with `stamp >= N`;
+ *   - coverage is OK ⟺ `coverageFloor(entry) <= N`, where coverageFloor is the
+ *     smallest step FULLY present in the ring (its smallest retained stamp, bumped
+ *     by one when that leading step was only partially evicted by overflow). If
+ *     `coverageFloor > N` the ring starts AFTER the client's frontier (a hole, or
+ *     the client's seed simply lagged behind a rotation) → 204 → the client
+ *     refetches (a larger N) and re-attaches.
+ * The N cutoff is applied in ALL branches, INCLUDING the finished-retained replay.
+ *
+ * ── same-tick invariants (unchanged, still load-bearing) ─────────────────────
+ * invariant 1: only the matching run may mutate/observe an entry (runId check).
+ * invariant 2: retention deletes ONLY its own entry (a replacement may own the key).
+ * invariant 3: open() over a live entry mirrors the done-path (subscribers released).
+ * invariant 4: the tail SLICE + subscriber registration happen in ONE synchronous
+ *              tick inside attach() — no await between them — so a concurrently
+ *              ingested frame is EITHER in the snapshot (buffered before the sync
+ *              block, and the just-added subscriber never sees it) OR fanned out to
+ *              the paused subscriber's `pending` (ingested after) — never both and
+ *              never neither: no loss, no duplication. NOTE (#491): the controller
+ *              now AWAITS the drain-respecting tail write BEFORE calling start(), so
+ *              frames ingested during that await accumulate in `pending`; this is
+ *              bounded by the subscriber cap (an overflow degrades start() to an
+ *              end(), a 204-equivalent). It is the SYNCHRONOUS snapshot+registration
+ *              — not a same-tick start() — that makes this correct.
+ * invariant 5: the controller wires close-cleanup BEFORE any write.
+ * invariant 6: no cross-run replay — the `anchor` (the client's assistant row id)
+ *              must match this run's assistant id, or a foreign run's transcript
+ *              would be appended to the client's message.
  */
 
 /** How long a finished entry is retained for late attach (replay + immediate end). */
 export const RUN_STREAM_RETAIN_FINISHED_MS = 30_000;
 
 /**
- * Per-run replay buffer cap. Past this the buffer is dropped (attach -> 204, and
- * the client falls back to its restore + degraded-poll path, #430).
- *
- * Raised from 4MB to 32MB (#430): marathon autonomous runs (11-25 min observed)
- * stream far more than 4MB of SSE frames, so a live disconnect mid-run would find
- * an already-overflowed buffer and could only degrade-poll instead of re-attaching
- * to the live tail. 32MB comfortably covers those runs while staying bounded.
- *
- * Memory cost: this is the WORST-CASE retained size PER ACTIVE run (the buffer is
- * freed on finish + retention, or dropped immediately on overflow). With the small
- * number of concurrent autonomous runs a single workspace realistically has, 32MB
- * each is an acceptable ceiling; the overflow->204->degraded-poll fallback remains
- * the backstop for anything larger, so correctness never depends on this bound.
+ * DEFAULT per-run replay ring cap (#491, down from 32MB). SSE frames carry
+ * UNcompacted tool outputs + framing overhead (×1.5–2 vs the persisted parts), so
+ * a "2–3 large reads + reasoning" step routinely blows past 2MB; 4MB comfortably
+ * holds a step or two of TAIL, which is all a resuming client needs (steps below
+ * its persisted frontier come from the seed, not the ring). The ring stays bounded
+ * because it rotates on every confirmed persist; this cap is only the ceiling for
+ * the un-persisted tail between rotations. Env-tunable via
+ * AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES (bytes); a 0/invalid value falls back to this.
  */
-export const RUN_STREAM_MAX_BUFFER_BYTES = 32 * 1024 * 1024;
+export const AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES = 4 * 1024 * 1024;
 
-// 2x the replay cap: a just-written full-replay burst alone can never trip the
-// per-subscriber cap (see controller); only a genuinely stalled socket can.
-export const SUBSCRIBER_MAX_BUFFERED_BYTES = 2 * RUN_STREAM_MAX_BUFFER_BYTES;
+// 2× the ring cap: a just-written full-tail burst alone can never trip the
+// per-subscriber cap (see controller); only a genuinely stalled socket can. This
+// derivative relationship is preserved even when the ring cap is env-overridden.
+export const SUBSCRIBER_MAX_BUFFERED_BYTES = 2 * AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES;
+
+/**
+ * A finish-step boundary frame is exactly `data: {"type":"finish-step"...}\n\n`
+ * (verified empirically against ai@6.0.207 — each UI-message-stream part is a
+ * single `data: {json}\n\n` event, never split across `data:` lines, and `type`
+ * is always the first key). A prefix match is cheaper than JSON.parse-per-frame
+ * and has no false positives: a literal `"type":"finish-step"` inside a text
+ * delta is JSON-escaped (`\"type\":...`), and the frame would start with
+ * `data: {"type":"text-delta"` anyway.
+ */
+const FINISH_STEP_FRAME_PREFIX = 'data: {"type":"finish-step"';
+
+/** Resolve the ring cap from the environment, falling back to the default. */
+function resolveMaxBufferBytes(): number {
+  const raw = process.env.AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES;
+  if (!raw) return AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES;
+  const parsed = Number(raw);
+  return Number.isFinite(parsed) && parsed > 0
+    ? Math.floor(parsed)
+    : AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES;
+}
 
 export interface RunStreamCallbacks {
   onFrame: (frame: string) => void;
@@ -44,6 +124,9 @@ export interface RunStreamCallbacks {
 }
 
 export interface RunStreamAttachment {
+  // The synthetic `start` frame (carrying { runId, chatId }) followed by the
+  // buffered TAIL filtered to `stamp >= N`. The controller writes these to the
+  // socket in chunks respecting drain, then calls start().
   replay: string[];
   finished: boolean;
   start(): void; // drain pending frames (order preserved) and go live
@@ -53,14 +136,19 @@ export interface RunStreamAttachment {
 interface Subscriber extends RunStreamCallbacks {
   started: boolean;
   pending: string[];
-  // Byte size of `pending`, capped at SUBSCRIBER_MAX_BUFFERED_BYTES. `start()` is
-  // called in the SAME tick as `attach()` today (see attach), so `pending` never
-  // holds more than one microtask of frames — but the async `attach` signature is
-  // a phase-2 seam: an await between attach and start would let a stalled paused
-  // subscriber buffer the WHOLE run here. The cap is the structural backstop.
+  // Byte size of `pending`, capped at the subscriber cap. `start()` is called in
+  // the SAME tick as `attach()` today, so `pending` never holds more than one
+  // microtask of frames — but the controller writes the (potentially large) tail
+  // respecting drain BEFORE start(), so a stalled socket can accumulate here; the
+  // cap is the structural backstop (an overflow degrades start() to an end()).
   pendingBytes: number;
   overflowed: boolean;
   pendingEnd: boolean;
+  // The client's step frontier N: this subscriber only receives frames with
+  // `stamp >= minStamp` (the tail past what it already persisted). Live frames
+  // always satisfy this (their stamp is the current, highest step), so it only
+  // filters the rare out-of-order below-frontier frame.
+  minStamp: number;
 }
 
 interface Entry {
@@ -68,8 +156,20 @@ interface Entry {
   // The persisted assistant row id of this run (set at bind; undefined if the
   // seed failed). Used by the attach anchor check (invariant 6).
   assistantMessageId?: string;
+  // Parallel arrays: frames[i] is the SSE string, stamps[i] its step number.
   frames: string[];
+  stamps: number[];
   bytes: number;
+  // The running step counter used to stamp the NEXT frame (number of finish-step
+  // frames seen so far).
+  currentStamp: number;
+  // The highest confirmed `stepsPersisted`: frames with stamp < persistedFloor are
+  // on disk (safe to drop, never re-buffered). Monotonic (confirmPersistedStep).
+  persistedFloor: number;
+  // The highest stamp EVICTED by an overflow (unsafe) drop, -1 if none. Used to
+  // detect a partially-evicted leading step when computing the coverage floor.
+  overflowThroughStamp: number;
+  // Sticky-for-logging only: at least one unsafe (overflow) eviction happened.
   overflowed: boolean;
   finished: boolean;
   subscribers: Set<Subscriber>;
@@ -80,6 +180,10 @@ interface Entry {
 export class AiChatStreamRegistryService implements OnModuleDestroy {
   private readonly logger = new Logger(AiChatStreamRegistryService.name);
   private readonly entries = new Map<string, Entry>(); // key: chatId
+  // Env-resolved caps (per instance) so a deployment can tune the ceiling without
+  // a code change. The subscriber cap keeps the documented 2× relationship.
+  readonly maxBufferBytes = resolveMaxBufferBytes();
+  readonly subscriberMaxBufferedBytes = 2 * this.maxBufferBytes;
 
   /**
    * Register a fresh entry at the START of a run (before any frame), so a tab
@@ -105,7 +209,11 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
     this.entries.set(chatId, {
       runId,
       frames: [],
+      stamps: [],
       bytes: 0,
+      currentStamp: 0,
+      persistedFloor: 0,
+      overflowThroughStamp: -1,
       overflowed: false,
       finished: false,
       subscribers: new Set<Subscriber>(),
@@ -150,6 +258,34 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
     void pump();
   }
 
+  /**
+   * Confirm that step `stepsPersisted` (a COUNT: steps 0..stepsPersisted-1) is on
+   * disk for this run, and ROTATE the ring: drop the buffered frames of those
+   * now-persisted steps (stamp < stepsPersisted). This is the ONLY thing that
+   * rotates the ring, and it is called ONLY after a genuinely SUCCESSFUL per-step
+   * persist (see ai-chat.service updateStreaming). A failed persist never calls
+   * it, so the ring covers more (auto-safe). Identity-checked (invariant 1) and
+   * monotonic (a stale lower count is ignored).
+   */
+  confirmPersistedStep(
+    chatId: string,
+    runId: string,
+    stepsPersisted: number,
+  ): void {
+    const entry = this.entries.get(chatId);
+    if (!entry || entry.runId !== runId) return;
+    if (!Number.isFinite(stepsPersisted) || stepsPersisted <= entry.persistedFloor)
+      return;
+    entry.persistedFloor = stepsPersisted;
+    // Clean rotation: drop the persisted steps from the head. These frames are on
+    // disk + carried by a fresh client seed, so this NEVER opens a gap.
+    while (entry.frames.length > 0 && entry.stamps[0] < stepsPersisted) {
+      entry.bytes -= Buffer.byteLength(entry.frames[0]);
+      entry.frames.shift();
+      entry.stamps.shift();
+    }
+  }
+
   /**
    * Terminate a run's entry from the OUTER catch of the stream method (a failure
    * before/while wiring the pipe, so `done` will never arrive). Identity-checked
@@ -162,36 +298,77 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
   }
 
   /**
-   * Attach to a run's stream. Async only for the phase-2 Redis seam — the body
-   * runs synchronously so the replay snapshot and the subscriber registration
-   * happen in ONE tick with no await between them (invariant 4): a frame ingested
-   * concurrently cannot slip into the gap and be lost or duplicated.
+   * Attach to a run's stream from the client's step frontier `n` (its persisted
+   * `stepsPersisted`). Async only for the phase-2 Redis seam — the body runs
+   * synchronously so the tail SLICE and the subscriber registration happen in ONE
+   * tick with no await between them (invariant 4).
    *
    * Returns null (-> the caller answers 204) when:
-   *  - there is no entry, or it overflowed (replay is gone);
-   *  - expect=live with an anchor that does not match this run's assistant id
-   *    (invariant 6: a stripped tab must never replay a FOREIGN run's transcript);
-   *  - the run finished and the caller did not expect a live tail.
-   * A finished run with expect=live yields a replay-only attachment (no
-   * subscriber registered). Otherwise a paused subscriber is registered and the
-   * caller replays `replay`, then calls start() to drain and go live.
+   *  - there is no entry;
+   *  - the `anchor` does not match this run's assistant id (invariant 6);
+   *  - the ring does not cover the client's frontier (coverageFloor > n): a hole
+   *    from overflow, or the client's seed simply lagged behind a rotation. The
+   *    client then refetches (a larger n) and re-attaches.
+   *
+   * Otherwise the attachment's `replay` is a synthetic `start` frame (the run-fact
+   * on re-attach) followed by the buffered tail filtered to `stamp >= n`. For a
+   * FINISHED run this is replay-only (no subscriber) and ends after the replay —
+   * with n = N_final that tail is just the run's `finish` frame, so the client
+   * closes the stream. For a LIVE run a paused subscriber is registered; the
+   * caller writes the replay (respecting drain) then calls start() to drain the
+   * pending frames and go live.
    */
   async attach(
     chatId: string,
-    expectLive: boolean,
     anchor: string | undefined,
+    // The client's persisted step frontier. `null` = a NOT-tail-aware client (no
+    // `n` query param) — a legacy/parameterless tab that expects the old
+    // "finished -> 204 -> poll" contract; distinct from `0` (a tail-aware client
+    // with nothing persisted yet).
+    n: number | null,
     cb: RunStreamCallbacks,
   ): Promise<RunStreamAttachment | null> {
     const entry = this.entries.get(chatId);
-    if (!entry || entry.overflowed) return null;
+    if (!entry) return null;
     // Invariant 6: cross-run replay is forbidden. Before bind, assistantMessageId
     // is undefined and mismatches any anchor -> 204 -> client restore+poll path.
-    if (expectLive && anchor && entry.assistantMessageId !== anchor) return null;
-    if (entry.finished && !expectLive) return null;
-    if (entry.finished && expectLive) {
+    if (anchor && entry.assistantMessageId !== anchor) return null;
+    // #491 regression guard (#137/#161 dup): a NOT-tail-aware client (no `n`)
+    // resuming a FINISHED run must 204 and poll — the old `finished && !expectLive`
+    // gate. Without this, a missing `n` collapsing to frontier 0 would serve the
+    // WHOLE tail of a finished, NON-rotated run (coverageFloor 0), and a
+    // parameterless client that never stripped its transcript would APPEND that
+    // full replay onto the steps it already shows -> duplicated text. A tail-aware
+    // client (n present, incl. n=0) still gets the tail past its frontier.
+    if (entry.finished && n === null) return null;
+    // A finished entry with NOTHING in the ring (aborted before the first frame,
+    // or fully overflowed) has no tail to deliver -> 204 -> the client polls.
+    if (entry.finished && entry.frames.length === 0) return null;
+    // A LIVE run with no `n` (legacy parameterless) replays from step 0 (the old
+    // behavior); a tail-aware client resumes from its frontier.
+    const frontier = n ?? 0;
+    const floor = this.coverageFloor(entry);
+    if (floor > frontier) {
+      this.logger.warn(
+        `run-stream attach gap for run=${entry.runId}: coverageFloor=${floor} ` +
+          `> client frontier=${frontier} -> 204 (client refetches + re-attaches)`,
+      );
+      return null;
+    }
+
+    const startFrame = this.buildStartFrame(chatId, entry.runId);
+    const sliceTail = (): string[] => {
+      const out: string[] = [startFrame];
+      for (let i = 0; i < entry.frames.length; i++) {
+        if (entry.stamps[i] >= frontier) out.push(entry.frames[i]);
+      }
+      return out;
+    };
+
+    if (entry.finished) {
       // Replay-only: the run is done, no subscriber is registered.
       return {
-        replay: entry.frames.slice(),
+        replay: sliceTail(),
         finished: true,
         start: () => undefined,
         unsubscribe: () => undefined,
@@ -206,15 +383,12 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
       pendingBytes: 0,
       overflowed: false,
       pendingEnd: false,
+      minStamp: frontier,
     };
+    // Register + snapshot in the SAME synchronous block (invariant 4). No await
+    // separates them, so a concurrently ingested frame cannot be lost/duplicated.
     entry.subscribers.add(sub);
-    // Snapshot in the SAME synchronous block as the registration (invariant 4).
-    const replay = entry.frames.slice();
-    // CONTRACT: the caller MUST call start() in the SAME tick as this attach()
-    // returns — no await between them. While a subscriber is paused, every frame
-    // is buffered in sub.pending; a delayed start() lets a whole run accumulate
-    // there. The pendingBytes cap (see ingestFrame) is the structural backstop if
-    // that contract is ever broken (e.g. the phase-2 Redis await seam).
+    const replay = sliceTail();
     return {
       replay,
       finished: false,
@@ -263,24 +437,83 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
     this.entries.clear();
   }
 
-  /** Buffer + fan-out a single frame. See invariant/overflow semantics inline. */
+  /** The synthetic `start` frame the tail is prefixed with — the source of the
+   *  run-fact (runId/chatId) on re-attach. A `start` frame does NOT reset the
+   *  client's message parts (ai@6.0.207 createStreamingUIMessageState), so it is
+   *  safe to prepend even when the sliced tail begins mid-message. */
+  private buildStartFrame(chatId: string, runId: string): string {
+    return `data: ${JSON.stringify({
+      type: 'start',
+      messageMetadata: { runId, chatId },
+    })}\n\n`;
+  }
+
+  /**
+   * The smallest step FULLY present in the ring: its smallest retained stamp, or
+   * (when the leading step was only partially evicted by an overflow) one past it.
+   * When the ring is empty it is the current step (only the live tail is coming).
+   * An attach at frontier `n` is covered ⟺ coverageFloor <= n.
+   */
+  private coverageFloor(entry: Entry): number {
+    // Empty ring: only the live tail is coming. The floor is the current step,
+    // but never below persistedFloor — a confirmed persist can rotate the ring
+    // empty while currentStamp still lags a beat behind on another connection, so
+    // max() keeps the invariant STRUCTURAL (a client with n = persistedFloor is
+    // always covered) rather than timing-dependent.
+    if (entry.frames.length === 0)
+      return Math.max(entry.currentStamp, entry.persistedFloor);
+    const min = entry.stamps[0];
+    return entry.overflowThroughStamp >= min ? min + 1 : min;
+  }
+
+  /**
+   * Buffer (step-stamped) + fan-out a single frame. The stamp is the number of
+   * finish-step frames seen BEFORE this one; a finish-step frame carries the
+   * current value and THEN increments the counter (so its stamp equals the 0-based
+   * index of the step it closes). Only frames at/above persistedFloor are buffered
+   * (already-persisted steps are on disk); the ring is then trimmed to the byte
+   * cap, an unsafe eviction opening a gap. Fan-out is always live (filtered per
+   * subscriber by its frontier).
+   */
   private ingestFrame(entry: Entry, frame: string): void {
-    entry.bytes += Buffer.byteLength(frame);
-    if (!entry.overflowed) {
+    const size = Buffer.byteLength(frame);
+    const stamp = entry.currentStamp;
+    if (frame.startsWith(FINISH_STEP_FRAME_PREFIX)) {
+      entry.currentStamp = stamp + 1;
+    }
+
+    // Buffer for replay only if this step is not already persisted+rotated away.
+    if (stamp >= entry.persistedFloor) {
       entry.frames.push(frame);
-      if (entry.bytes > RUN_STREAM_MAX_BUFFER_BYTES) {
-        // The crossing frame was already counted AND (below) fanned out; only the
-        // replay buffer is dropped. After overflow no more frames are buffered,
-        // but live fan-out continues.
-        entry.overflowed = true;
-        entry.frames = [];
-        this.logger.warn(
-          `run-stream buffer overflow for run=${entry.runId}; ` +
-            `late attach will 204 until the run ends`,
-        );
+      entry.stamps.push(stamp);
+      entry.bytes += size;
+      // Enforce the ring cap. Evicting a not-yet-persisted frame (stamp >=
+      // persistedFloor) opens a GAP; a leftover persisted frame (< floor) is a
+      // safe drop. Keep evicting until the ring is back under the cap.
+      while (entry.bytes > this.maxBufferBytes && entry.frames.length > 0) {
+        const evStamp = entry.stamps[0];
+        entry.bytes -= Buffer.byteLength(entry.frames[0]);
+        entry.frames.shift();
+        entry.stamps.shift();
+        if (evStamp >= entry.persistedFloor) {
+          if (evStamp > entry.overflowThroughStamp)
+            entry.overflowThroughStamp = evStamp;
+          if (!entry.overflowed) {
+            entry.overflowed = true;
+            this.logger.warn(
+              `run-stream ring overflow for run=${entry.runId}: an un-persisted ` +
+                `step was evicted to stay under ${this.maxBufferBytes}B; a late ` +
+                `attach at an evicted step will 204 until a later persist confirms`,
+            );
+          }
+        }
       }
     }
+
+    // Fan out live, filtered to each subscriber's frontier (a subscriber only
+    // wants the tail past the step it already persisted).
     for (const sub of entry.subscribers) {
+      if (stamp < sub.minStamp) continue;
       if (sub.started) {
         try {
           sub.onFrame(frame);
@@ -289,12 +522,12 @@ export class AiChatStreamRegistryService implements OnModuleDestroy {
         }
       } else {
         sub.pending.push(frame);
-        sub.pendingBytes += Buffer.byteLength(frame);
-        if (sub.pendingBytes > SUBSCRIBER_MAX_BUFFERED_BYTES) {
+        sub.pendingBytes += size;
+        if (sub.pendingBytes > this.subscriberMaxBufferedBytes) {
           // The paused subscriber's buffer overflowed — only possible if start()
-          // was delayed past the same-tick contract (the phase-2 await seam).
-          // Drop it rather than buffer the whole run; on start() it degrades to an
-          // immediate end (a 204-equivalent) instead of replaying a partial.
+          // was delayed (the controller's drain-respecting tail write, or the
+          // phase-2 await seam). Drop it rather than buffer the whole run; on
+          // start() it degrades to an immediate end (a 204-equivalent).
           sub.overflowed = true;
           sub.pending = [];
           entry.subscribers.delete(sub);

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

@@ -1,19 +1,27 @@
 import {
   AiChatStreamRegistryService,
-  RUN_STREAM_MAX_BUFFER_BYTES,
+  AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES,
   RUN_STREAM_RETAIN_FINISHED_MS,
-  SUBSCRIBER_MAX_BUFFERED_BYTES,
   RunStreamCallbacks,
 } from './ai-chat-stream-registry.service';
 
 /**
- * Unit tests for the in-memory run-stream registry (#184 phase 1.5). The registry
- * is the whole of the resumable-transport contract: replay ordering, paused ->
- * live hand-off, overflow, retention, the anchor check (invariant 6), and the
- * mirror-the-done-path replace semantics (invariant 3). Every enumerated case in
- * the issue's task 1.5 has a test here.
+ * Unit tests for the in-memory run-stream registry (#184 phase 1.5, step-aligned
+ * retention #491). The registry is the whole of the resumable-transport contract:
+ * step-stamped retention, tail-only attach at the client's frontier N, the
+ * confirmed-persist ring rotation (and the anti-inversion rule), the memory bound,
+ * the overflow gap, paused -> live hand-off, retention, the anchor check
+ * (invariant 6), and the mirror-the-done-path replace semantics (invariant 3).
  */
 
+// Real ai@6 UI-message-stream SSE frames are `data: {json}\n\n`, one part each.
+const sse = (part: Record<string, unknown>): string =>
+  `data: ${JSON.stringify(part)}\n\n`;
+const finishStep = (): string => sse({ type: 'finish-step' });
+const textDelta = (id: string, delta: string): string =>
+  sse({ type: 'text-delta', id, delta });
+const finish = (): string => sse({ type: 'finish' });
+
 // A ReadableStream whose frames the test pushes explicitly, plus close/error.
 function makePushStream(): {
   stream: ReadableStream<string>;
@@ -58,6 +66,9 @@ function collector(): {
   };
 }
 
+// The tail past the synthetic start frame (replay[0] is always the start frame).
+const tail = (replay: string[]): string[] => replay.slice(1);
+
 describe('AiChatStreamRegistryService', () => {
   const CHAT = 'chat-1';
   let registry: AiChatStreamRegistryService;
@@ -71,7 +82,21 @@ describe('AiChatStreamRegistryService', () => {
     registry.onModuleDestroy();
   });
 
-  it('replays frames in arrival order (live attach)', async () => {
+  it('prepends a synthetic start frame carrying { runId, chatId }', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, 'assist-1', 0, c.cb))!;
+    const start = JSON.parse(att.replay[0].replace(/^data: /, '').trim());
+    expect(start.type).toBe('start');
+    expect(start.messageMetadata).toEqual({ runId: 'run-1', chatId: CHAT });
+  });
+
+  it('replays the buffered tail (from frontier 0) in arrival order (live attach)', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
@@ -81,13 +106,13 @@ describe('AiChatStreamRegistryService', () => {
     await flush();
 
     const c = collector();
-    const att = await registry.attach(CHAT, false, undefined, c.cb);
+    const att = await registry.attach(CHAT, 'assist-1', 0, c.cb);
     expect(att).not.toBeNull();
-    expect(att!.replay).toEqual(['a', 'b', 'c']);
+    expect(tail(att!.replay)).toEqual(['a', 'b', 'c']);
     expect(att!.finished).toBe(false);
   });
 
-  it('late attach gets the full prefix as replay plus the live tail', async () => {
+  it('late attach gets the buffered prefix as tail plus the live tail', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
@@ -96,17 +121,16 @@ describe('AiChatStreamRegistryService', () => {
     await flush();
 
     const c = collector();
-    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
-    expect(att.replay).toEqual(['a', 'b']);
+    const att = (await registry.attach(CHAT, 'assist-1', 0, c.cb))!;
+    expect(tail(att.replay)).toEqual(['a', 'b']);
     att.start();
-    // Live tail arrives after start().
     src.push('c');
     src.push('d');
     await flush();
     expect(c.frames).toEqual(['c', 'd']);
   });
 
-  it('a paused subscriber receives frames buffered during pause in order, then live (no loss/reorder)', async () => {
+  it('a paused subscriber receives frames buffered during pause in order, then live', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
@@ -114,81 +138,45 @@ describe('AiChatStreamRegistryService', () => {
     await flush();
 
     const c = collector();
-    // Attach (paused). Frames that arrive BEFORE start() must queue, not drop.
-    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
-    expect(att.replay).toEqual(['a']);
+    const att = (await registry.attach(CHAT, 'assist-1', 0, c.cb))!;
+    expect(tail(att.replay)).toEqual(['a']);
     src.push('b'); // arrives while paused -> pending
     src.push('c');
     await flush();
     expect(c.frames).toEqual([]); // nothing delivered yet (paused)
-    att.start(); // drains pending in order
+    att.start();
     expect(c.frames).toEqual(['b', 'c']);
-    src.push('d'); // now live
+    src.push('d');
     await flush();
     expect(c.frames).toEqual(['b', 'c', 'd']);
   });
 
   it('a run that finishes while a subscriber is paused ends it on start()', async () => {
     registry.open(CHAT, 'run-1');
+    registry.bind(CHAT, 'run-1', 'assist-1', makePushStream().stream);
     const c = collector();
-    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
-    // Terminate the run while the subscriber is still paused.
+    const att = (await registry.attach(CHAT, 'assist-1', 0, c.cb))!;
     registry.abortEntry(CHAT, 'run-1');
     expect(c.ended()).toBe(0); // paused: not ended yet
     att.start();
     expect(c.ended()).toBe(1); // start() drains + ends
   });
 
-  it('finished + expect=live returns a replay WITHOUT registering a subscriber', async () => {
-    registry.open(CHAT, 'run-1');
-    const src = makePushStream();
-    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
-    src.push('a');
-    src.push('b');
-    src.close();
-    await flush();
-
-    const c = collector();
-    const att = (await registry.attach(CHAT, true, undefined, c.cb))!;
-    expect(att.finished).toBe(true);
-    expect(att.replay).toEqual(['a', 'b']);
-    // No subscriber registered: start()/unsubscribe are no-ops and the entry has
-    // zero subscribers.
-    const entry = (registry as any).entries.get(CHAT);
-    expect(entry.subscribers.size).toBe(0);
-    att.start();
-    expect(c.frames).toEqual([]);
-  });
-
-  it('finished WITHOUT expect=live returns null', async () => {
-    registry.open(CHAT, 'run-1');
-    const src = makePushStream();
-    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
-    src.push('a');
-    src.close();
-    await flush();
-
-    const c = collector();
-    expect(await registry.attach(CHAT, false, undefined, c.cb)).toBeNull();
-  });
-
-  it('anchor mismatch with expect=live returns null (and null before bind sets assistantMessageId)', async () => {
+  it('anchor mismatch returns null (and null before bind sets assistantMessageId)', async () => {
     registry.open(CHAT, 'run-1');
     const c = collector();
     // Before bind: assistantMessageId is undefined -> mismatches any anchor.
-    expect(
-      await registry.attach(CHAT, true, 'assist-1', c.cb),
-    ).toBeNull();
+    expect(await registry.attach(CHAT, 'assist-1', 0, c.cb)).toBeNull();
 
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
     src.push('a');
     await flush();
     // Wrong anchor -> null (cross-run replay forbidden, invariant 6).
-    expect(await registry.attach(CHAT, true, 'other-id', c.cb)).toBeNull();
+    expect(await registry.attach(CHAT, 'other-id', 0, c.cb)).toBeNull();
   });
 
-  it('matching anchor with expect=live attaches', async () => {
+  it('matching anchor attaches', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
@@ -196,97 +184,60 @@ describe('AiChatStreamRegistryService', () => {
     await flush();
 
     const c = collector();
-    const att = await registry.attach(CHAT, true, 'assist-1', c.cb);
+    const att = await registry.attach(CHAT, 'assist-1', 0, c.cb);
     expect(att).not.toBeNull();
-    expect(att!.replay).toEqual(['a']);
+    expect(tail(att!.replay)).toEqual(['a']);
   });
 
-  it('overflow: attach returns null, but the LIVE subscriber keeps receiving (incl. the crossing frame)', async () => {
+  it('a throwing onFrame ejects only that subscriber; the ingest loop stays alive', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
 
-    // A live (started) subscriber attached before the flood.
+    const bad = collector();
+    const badAtt = (await registry.attach(CHAT, 'assist-1', 0, {
+      onFrame: () => {
+        throw new Error('boom');
+      },
+      onEnd: bad.cb.onEnd,
+    }))!;
+    badAtt.start();
+
+    const good = collector();
+    const goodAtt = (await registry.attach(CHAT, 'assist-1', 0, good.cb))!;
+    goodAtt.start();
+
+    src.push('a'); // bad throws on this frame -> ejected
+    src.push('b'); // good still receives both
+    await flush();
+
+    const entry = (registry as any).entries.get(CHAT);
+    expect(entry.subscribers.size).toBe(1);
+    expect(good.frames).toEqual(['a', 'b']);
+  });
+
+  it('open() over a LIVE entry ends started subscribers once; a late done never touches the new entry (invariant 3)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push('a');
+    await flush();
+
     const c = collector();
-    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
+    const att = (await registry.attach(CHAT, 'assist-1', 0, c.cb))!;
     att.start();
 
-    // Cap-relative so it survives a buffer-cap change (#430): a quarter-cap frame
-    // means 5 frames comfortably exceed the replay cap; the last one crosses.
-    const chunk = 'x'.repeat(Math.floor(RUN_STREAM_MAX_BUFFER_BYTES / 4));
-    for (let i = 0; i < 5; i++) src.push(chunk + i);
-    await flush();
-
-    const entry = (registry as any).entries.get(CHAT);
-    expect(entry.overflowed).toBe(true);
-    expect(entry.bytes).toBeGreaterThan(RUN_STREAM_MAX_BUFFER_BYTES);
-    // The live subscriber received ALL 5 frames, including the crossing one.
-    expect(c.frames).toHaveLength(5);
-    expect(c.frames[4]).toBe(chunk + 4);
-
-    // A NEW attach after overflow gets null (replay buffer is gone).
-    const c2 = collector();
-    expect(await registry.attach(CHAT, false, undefined, c2.cb)).toBeNull();
-  });
-
-  it('a paused subscriber whose pending buffer overflows is dropped and ends on start(); other subscribers keep receiving', async () => {
-    registry.open(CHAT, 'run-1');
-    const src = makePushStream();
-    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
-
-    // A: paused (start() deliberately delayed to simulate the phase-2 await seam).
-    const a = collector();
-    const attA = (await registry.attach(CHAT, false, undefined, a.cb))!;
-    // B: live (started) — its delivery must be unaffected by A's overflow.
-    const b = collector();
-    const attB = (await registry.attach(CHAT, false, undefined, b.cb))!;
-    attB.start();
-
-    // Cap-relative so it survives a buffer-cap change (#430): a quarter-of-the-
-    // per-subscriber-cap frame means 5 frames exceed A's paused-pending cap while
-    // B streams every frame live.
-    const chunk = 'x'.repeat(Math.floor(SUBSCRIBER_MAX_BUFFERED_BYTES / 4));
-    for (let i = 0; i < 5; i++) src.push(chunk + i);
-    await flush();
-
-    const entry = (registry as any).entries.get(CHAT);
-    // A was dropped from the subscriber set on overflow; B (started) remains.
-    expect(entry.subscribers.size).toBe(1);
-    expect(a.frames).toEqual([]); // paused + overflowed: nothing was delivered
-    // B received every frame live (delivery unaffected by A's overflow).
-    expect(b.frames).toHaveLength(5);
-
-    // A's start() (arriving late) degrades to an immediate end, not a partial replay.
-    attA.start();
-    expect(a.frames).toEqual([]);
-    expect(a.ended()).toBe(1);
-  });
-
-  it('open() over a LIVE entry ends started subscribers exactly once and a late done does not touch the new entry (invariant 3)', async () => {
-    registry.open(CHAT, 'run-1');
-    const src = makePushStream();
-    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
-    src.push('a');
-    await flush();
-
-    const c = collector();
-    const att = (await registry.attach(CHAT, false, undefined, c.cb))!;
-    att.start(); // started subscriber on run-1
-
-    // run-2 starts on the same chat while run-1's tee is still reading.
     registry.open(CHAT, 'run-2');
-    expect(c.ended()).toBe(1); // exactly one onEnd from the replace
+    expect(c.ended()).toBe(1);
 
     const newEntry = (registry as any).entries.get(CHAT);
     expect(newEntry.runId).toBe('run-2');
     expect(newEntry.finished).toBe(false);
 
-    // The old tee now completes: its late done must NOT double-end nor delete the
-    // new entry.
     src.push('b');
     src.close();
     await flush();
-    expect(c.ended()).toBe(1); // still exactly one
+    expect(c.ended()).toBe(1);
     const still = (registry as any).entries.get(CHAT);
     expect(still).toBe(newEntry);
     expect(still.runId).toBe('run-2');
@@ -299,7 +250,6 @@ describe('AiChatStreamRegistryService', () => {
     src.push('a');
     await flush();
     const entry = (registry as any).entries.get(CHAT);
-    // Frames were NOT ingested (bind bailed), assistantMessageId untouched.
     expect(entry.frames).toEqual([]);
     expect(entry.assistantMessageId).toBeUndefined();
   });
@@ -310,32 +260,276 @@ describe('AiChatStreamRegistryService', () => {
     const entry = (registry as any).entries.get(CHAT);
     expect(entry.finished).toBe(false);
   });
+});
 
-  it('a throwing onFrame ejects only that subscriber; the ingest loop stays alive', async () => {
+/**
+ * #491 step-stamped retention: the boundary detector, tail-only slicing at the
+ * client's frontier N, the confirmed-persist rotation (+ anti-inversion), the
+ * overflow gap, the memory bound, and the finished-retained tail. All observable
+ * against the REAL registry driven through open/bind/ingest.
+ */
+describe('AiChatStreamRegistryService step-aligned retention (#491)', () => {
+  const CHAT = 'chat-s';
+  let registry: AiChatStreamRegistryService;
+
+  beforeEach(() => {
+    registry = new AiChatStreamRegistryService();
+    jest.spyOn((registry as any).logger, 'warn').mockImplementation(() => {});
+  });
+  afterEach(() => registry.onModuleDestroy());
+
+  const entryOf = () => (registry as any).entries.get(CHAT);
+
+  it('stamps frames by finish-step count, aligned with stepsPersisted', async () => {
     registry.open(CHAT, 'run-1');
     const src = makePushStream();
     registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    // step 0 content, its finish-step, step 1 content, its finish-step, finish.
+    src.push(textDelta('t0', 'a')); // stamp 0
+    src.push(finishStep()); // stamp 0 (the finish-step frame carries the pre value)
+    src.push(textDelta('t1', 'b')); // stamp 1
+    src.push(finishStep()); // stamp 1
+    src.push(finish()); // stamp 2
+    await flush();
+    const e = entryOf();
+    expect(e.stamps).toEqual([0, 0, 1, 1, 2]);
+    expect(e.currentStamp).toBe(2);
+  });
 
-    const bad = collector();
-    const badAtt = (await registry.attach(CHAT, false, undefined, {
-      onFrame: () => {
-        throw new Error('boom');
-      },
-      onEnd: bad.cb.onEnd,
-    }))!;
-    badAtt.start();
+  it('does NOT treat a text delta that merely quotes "finish-step" as a boundary', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    // A model that literally types "type":"finish-step" — JSON-escaped in the frame.
+    src.push(textDelta('t0', '"type":"finish-step"'));
+    await flush();
+    expect(entryOf().currentStamp).toBe(0); // no false boundary
+  });
 
-    const good = collector();
-    const goodAtt = (await registry.attach(CHAT, false, undefined, good.cb))!;
-    goodAtt.start();
-
-    src.push('a'); // bad throws on this frame -> ejected
-    src.push('b'); // good still receives both
+  it('tail-only: attach at N slices frames with stamp >= N', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(textDelta('t1', 'b')); // 1
+    src.push(finishStep()); // 1
+    src.push(textDelta('t2', 'c')); // 2 (in-progress)
     await flush();
 
-    const entry = (registry as any).entries.get(CHAT);
-    expect(entry.subscribers.size).toBe(1); // bad ejected, good remains
-    expect(good.frames).toEqual(['a', 'b']);
+    const c = collector();
+    // Client persisted 2 steps -> wants the tail from step 2.
+    const att = (await registry.attach(CHAT, 'assist-1', 2, c.cb))!;
+    expect(tail(att.replay)).toEqual([textDelta('t2', 'c')]);
+  });
+
+  it('attach in the MIDDLE of a step (N between finish-steps) slices from that step', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(textDelta('t1', 'b1')); // 1
+    src.push(textDelta('t1', 'b2')); // 1 (still step 1, no finish-step yet)
+    await flush();
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, 'assist-1', 1, c.cb))!;
+    // Step 0's frames are dropped from the tail; the whole in-progress step 1 is kept.
+    expect(tail(att.replay)).toEqual([textDelta('t1', 'b1'), textDelta('t1', 'b2')]);
+  });
+
+  it('rotates the ring ONLY on a confirmed persist (drops stamp < N)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(textDelta('t1', 'b')); // 1
+    await flush();
+    expect(entryOf().stamps).toEqual([0, 0, 1]);
+
+    // Confirm step 0 persisted (stepsPersisted = 1) -> drop stamp < 1.
+    registry.confirmPersistedStep(CHAT, 'run-1', 1);
+    expect(entryOf().stamps).toEqual([1]);
+    expect(entryOf().persistedFloor).toBe(1);
+  });
+
+  it('persist FAILED but the ring still fits -> attach SUCCEEDS and the tail includes step N', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(textDelta('t1', 'b')); // 1 (step 1's persist FAILED -> no confirm)
+    await flush();
+    // No confirmPersistedStep for step 1: the ring still holds step 1.
+
+    const c = collector();
+    // Client's last successful persist was step 0 -> stepsPersisted = 1.
+    const att = await registry.attach(CHAT, 'assist-1', 1, c.cb);
+    expect(att).not.toBeNull();
+    expect(tail(att!.replay)).toEqual([textDelta('t1', 'b')]); // includes step 1
+  });
+
+  it('persist failed AND the ring overflowed past N -> 204 (coverage gap)', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    // Step 0: a fat step that blows past the cap with NO persist confirmation.
+    const big = 'x'.repeat(Math.floor(AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES / 2));
+    src.push(textDelta('t0', big)); // 0
+    src.push(textDelta('t0', big)); // 0
+    src.push(textDelta('t0', big)); // 0 -> overflow evicts stamp-0 frames
+    await flush();
+    const e = entryOf();
+    expect(e.overflowed).toBe(true);
+    expect(e.bytes).toBeLessThanOrEqual(registry.maxBufferBytes);
+
+    // A client at frontier 0 falls at/below an evicted step -> gap -> null.
+    const c = collector();
+    expect(await registry.attach(CHAT, 'assist-1', 0, c.cb)).toBeNull();
+  });
+
+  it('stale N (client seed lagged behind a rotation) -> 204; after a refetch (larger N) -> success', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(textDelta('t1', 'b')); // 1
+    src.push(finishStep()); // 1
+    src.push(textDelta('t2', 'c')); // 2
+    await flush();
+    // Server confirmed steps 0 and 1 -> rotate away stamp < 2.
+    registry.confirmPersistedStep(CHAT, 'run-1', 2);
+    expect(entryOf().stamps).toEqual([2]);
+
+    // A client whose seed still says stepsPersisted = 1 -> below minStamp -> 204.
+    const stale = collector();
+    expect(await registry.attach(CHAT, 'assist-1', 1, stale.cb)).toBeNull();
+
+    // It refetches (now stepsPersisted = 2) and re-attaches -> success.
+    const fresh = collector();
+    const att = await registry.attach(CHAT, 'assist-1', 2, fresh.cb);
+    expect(att).not.toBeNull();
+    expect(tail(att!.replay)).toEqual([textDelta('t2', 'c')]);
+  });
+
+  it('overflow gap CLEARS once a later persist rotates out the holey steps', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    const big = 'x'.repeat(Math.floor(AI_CHAT_RUN_STREAM_MAX_BUFFER_BYTES / 2));
+    src.push(textDelta('t0', big)); // 0
+    src.push(textDelta('t0', big)); // 0
+    src.push(finishStep()); // 0 (still stamp 0)
+    src.push(textDelta('t1', 'small')); // 1
+    src.push(finishStep()); // 1
+    src.push(textDelta('t2', 'c')); // 2
+    await flush();
+    expect(entryOf().overflowed).toBe(true);
+
+    // Late persist confirms steps 0..1 -> rotates out the holey step-0 frames.
+    registry.confirmPersistedStep(CHAT, 'run-1', 2);
+    // A client at frontier 2 is now cleanly covered (the hole was below it).
+    const c = collector();
+    const att = await registry.attach(CHAT, 'assist-1', 2, c.cb);
+    expect(att).not.toBeNull();
+    expect(tail(att!.replay)).toEqual([textDelta('t2', 'c')]);
+  });
+
+  it('finished-retained + N = N_final -> empty tail plus the finish frame', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(finish()); // 1  (N_final = 1)
+    src.close();
+    await flush();
+    // The last step's per-step persist confirmed stepsPersisted = 1.
+    registry.confirmPersistedStep(CHAT, 'run-1', 1);
+
+    const c = collector();
+    const att = (await registry.attach(CHAT, 'assist-1', 1, c.cb))!;
+    expect(att.finished).toBe(true);
+    // Empty step tail; just the finish frame so the client's SDK closes the stream.
+    expect(tail(att.replay)).toEqual([finish()]);
+    // No subscriber registered for a finished run.
+    expect(entryOf().subscribers.size).toBe(0);
+  });
+
+  it('#491 regression (#137/#161 dup): a PARAMETERLESS attach (n=null) to a finished NON-rotated run -> 204, but n=0 still gets the tail', async () => {
+    // A finished, non-rotated run: frames present, coverageFloor 0. A missing `n`
+    // (null — a legacy/parameterless tab that never stripped its transcript) must
+    // 204 -> poll, NOT receive the whole tail it would append (duplicate). A
+    // tail-aware client (n=0 present) still resumes.
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a')); // 0
+    src.push(finishStep()); // 0
+    src.push(finish()); // 1
+    src.close();
+    await flush();
+    // NOT rotated (no confirmPersistedStep) -> stamps[0]=0, coverageFloor=0.
+    // MUTATION-VERIFY: revert the `finished && n === null -> null` gate (default n
+    // to 0) and the parameterless attach below serves the full tail instead of 204.
+    expect(await registry.attach(CHAT, 'assist-1', null, collector().cb)).toBeNull();
+    // A tail-aware client at frontier 0 IS served (the distinction: null != 0).
+    const tailAware = await registry.attach(CHAT, 'assist-1', 0, collector().cb);
+    expect(tailAware).not.toBeNull();
+    expect(tailAware!.finished).toBe(true);
+  });
+
+  it('confirmPersistedStep is monotonic and identity-checked', async () => {
+    registry.open(CHAT, 'run-1');
+    const src = makePushStream();
+    registry.bind(CHAT, 'run-1', 'assist-1', src.stream);
+    src.push(textDelta('t0', 'a'));
+    src.push(finishStep());
+    src.push(textDelta('t1', 'b'));
+    await flush();
+    registry.confirmPersistedStep(CHAT, 'run-1', 1);
+    expect(entryOf().persistedFloor).toBe(1);
+    // A stale lower count is ignored.
+    registry.confirmPersistedStep(CHAT, 'run-1', 0);
+    expect(entryOf().persistedFloor).toBe(1);
+    // A foreign runId is ignored.
+    registry.confirmPersistedStep(CHAT, 'WRONG', 5);
+    expect(entryOf().persistedFloor).toBe(1);
+  });
+
+  it('MEMORY BOUND: 5 parallel marathon runs each stream well past 32MB; each ring stays <= the cap', async () => {
+    const cap = registry.maxBufferBytes;
+    const chats = ['m0', 'm1', 'm2', 'm3', 'm4'];
+    const srcs = chats.map((chat) => {
+      registry.open(chat, `run-${chat}`);
+      const s = makePushStream();
+      registry.bind(chat, `run-${chat}`, `assist-${chat}`, s.stream);
+      return s;
+    });
+    // ~256KB frames; 160 per chat = 40MB streamed each, well past the old 32MB.
+    // Interleave a finish-step every 8 frames so steps advance realistically. No
+    // persist confirmation -> the ONLY thing keeping memory bounded is the cap.
+    const frame = 'y'.repeat(256 * 1024);
+    for (let batch = 0; batch < 20; batch++) {
+      for (let i = 0; i < 8; i++) {
+        for (const s of srcs) s.push(textDelta('t', frame));
+      }
+      for (const s of srcs) s.push(finishStep());
+      await flush(); // drain the pump so queues never hold a whole run
+    }
+    let total = 0;
+    for (const chat of chats) {
+      const e = (registry as any).entries.get(chat);
+      expect(e.bytes).toBeLessThanOrEqual(cap);
+      total += e.bytes;
+    }
+    // Total retained across all 5 runs is bounded by 5x the per-run cap — the old
+    // registry would have retained ~5x40MB = 200MB here.
+    expect(total).toBeLessThanOrEqual(cap * chats.length);
   });
 });
 
@@ -361,7 +555,7 @@ describe('AiChatStreamRegistryService retention timers', () => {
 
   it('a finished entry is removed after the retention window', () => {
     registry.open(CHAT, 'run-1');
-    registry.abortEntry(CHAT, 'run-1'); // finalize -> retention armed
+    registry.abortEntry(CHAT, 'run-1');
     expect((registry as any).entries.get(CHAT)).toBeDefined();
     jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
     expect((registry as any).entries.get(CHAT)).toBeUndefined();
@@ -369,20 +563,18 @@ describe('AiChatStreamRegistryService retention timers', () => {
 
   it('retention deletes ONLY its own entry (invariant 2)', () => {
     registry.open(CHAT, 'run-1');
-    registry.abortEntry(CHAT, 'run-1'); // arm retention for entry A
-    // Simulate the race where the key was replaced without clearing A's timer.
+    registry.abortEntry(CHAT, 'run-1');
     const sentinel = { marker: true };
     (registry as any).entries.set(CHAT, sentinel);
     jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
-    // A's timer saw entries.get(CHAT) !== A, so it did NOT delete the successor.
     expect((registry as any).entries.get(CHAT)).toBe(sentinel);
   });
 
   it('open() over a retained entry clears its timer and the successor survives', () => {
     registry.open(CHAT, 'run-1');
-    registry.abortEntry(CHAT, 'run-1'); // retained, timer armed
+    registry.abortEntry(CHAT, 'run-1');
     const clearSpy = jest.spyOn(global, 'clearTimeout');
-    registry.open(CHAT, 'run-2'); // must clear run-1's retain timer
+    registry.open(CHAT, 'run-2');
     expect(clearSpy).toHaveBeenCalled();
     jest.advanceTimersByTime(RUN_STREAM_RETAIN_FINISHED_MS + 1);
     const entry = (registry as any).entries.get(CHAT);

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

@@ -8,10 +8,12 @@ import { SUBSCRIBER_MAX_BUFFERED_BYTES } from './ai-chat-stream-registry.service
 import type { User, Workspace } from '@docmost/db/types/entity.types';
 
 /**
- * Wiring spec for the #184 phase 1.5 attach endpoint
+ * Wiring spec for the #184 phase 1.5 attach endpoint (tail-only #491)
  * (`GET /ai-chat/runs/:chatId/stream`). Owner-gated via assertOwnedChat; the
- * registry is mocked so this exercises ONLY the controller's replay/live/204/
- * cleanup wiring against a fake raw socket. Constructor order is (aiChatService,
+ * registry is mocked so this exercises ONLY the controller's tail-write/live/204/
+ * cleanup wiring against a fake raw socket. The attach signature is now
+ * `(chatId, anchor, n, cb)` — the client hands its persisted step frontier `n`
+ * and its assistant row id `anchor`. Constructor order is (aiChatService,
  * aiChatRunService, aiChatRepo, aiChatMessageRepo, aiTranscription, pageRepo,
  * streamRegistry, environment).
  */
@@ -86,8 +88,8 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
       attach: jest.fn(
         (
           _chatId: string,
-          _live: boolean,
           _anchor: string | undefined,
+          _n: number,
           cb: RunStreamCallbacks,
         ) => {
           capturedCb = cb;
@@ -156,7 +158,7 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     expect(res.hijack).not.toHaveBeenCalled();
   });
 
-  it('threads expect=live and anchor through to the registry', async () => {
+  it('threads anchor and the numeric frontier n through to the registry', async () => {
     const { controller, streamRegistry } = makeController({
       chat: owned,
       attachment: null,
@@ -165,8 +167,8 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     const { req } = makeReq();
     await controller.attachRunStream(
       'c1',
-      'live',
       'anchor-1',
+      '2',
       req,
       res,
       user,
@@ -174,13 +176,44 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     );
     expect(streamRegistry.attach).toHaveBeenCalledWith(
       'c1',
-      true,
       'anchor-1',
+      2, // parsed to a number
       expect.anything(),
     );
   });
 
-  it('passes expect=false when the query is absent', async () => {
+  it('#491: an ABSENT/invalid n passes null (not 0) so a finished run 204s (not-tail-aware)', async () => {
+    // Distinguishing a MISSING `n` from `n=0` is the #137/#161 dup guard: a
+    // parameterless/legacy tab must be handed null (-> the registry 204s a finished
+    // run) rather than frontier 0 (which would serve a finished non-rotated run's
+    // whole tail). MUTATION-VERIFY: revert to `Number(n) || 0` and this asserts 0.
+    const { controller, streamRegistry } = makeController({
+      chat: owned,
+      attachment: null,
+    });
+    for (const bad of [undefined, '', 'abc']) {
+      streamRegistry.attach.mockClear();
+      const { res } = makeRawRes();
+      const { req } = makeReq();
+      await controller.attachRunStream(
+        'c1',
+        undefined,
+        bad,
+        req,
+        res,
+        user,
+        workspace,
+      );
+      expect(streamRegistry.attach).toHaveBeenCalledWith(
+        'c1',
+        undefined,
+        null,
+        expect.anything(),
+      );
+    }
+  });
+
+  it('#491: a PRESENT n=0 passes 0 (tail-aware, distinct from absent)', async () => {
     const { controller, streamRegistry } = makeController({
       chat: owned,
       attachment: null,
@@ -190,7 +223,7 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     await controller.attachRunStream(
       'c1',
       undefined,
-      undefined,
+      '0',
       req,
       res,
       user,
@@ -198,8 +231,8 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     );
     expect(streamRegistry.attach).toHaveBeenCalledWith(
       'c1',
-      false,
       undefined,
+      0,
       expect.anything(),
     );
   });
@@ -245,8 +278,8 @@ describe('AiChatController attach endpoint (#184 phase 1.5)', () => {
     const { req } = makeReq();
     await controller.attachRunStream(
       'c1',
-      'live',
       'a1',
+      '1',
       req,
       res,
       user,

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

@@ -0,0 +1,108 @@
+import { ForbiddenException } from '@nestjs/common';
+import { AiChatController } from './ai-chat.controller';
+import type { User, Workspace } from '@docmost/db/types/entity.types';
+
+/**
+ * Wiring spec for the #491 delta-poll endpoint (`POST /ai-chat/messages/delta`).
+ * Owner-gated via assertOwnedChat (same gate as the other reads), NOT flag-gated.
+ * The run fact rides IN the delta response (no separate /run poll). Hand-rolled
+ * mocks — no Nest graph, no DB. Constructor order: (aiChatService,
+ * aiChatRunService, aiChatRepo, aiChatMessageRepo, aiTranscription, pageRepo).
+ */
+describe('AiChatController POST /ai-chat/messages/delta (#491)', () => {
+  const user = { id: 'u1' } as User;
+  const workspace = { id: 'ws1' } as Workspace;
+
+  function makeController(opts: {
+    chat?: unknown;
+    delta?: { rows: unknown[]; cursor: string };
+    run?: unknown;
+  }) {
+    const aiChatRunService = {
+      getLatestForChat: jest.fn().mockResolvedValue(opts.run),
+    };
+    const aiChatRepo = {
+      findById: jest.fn().mockResolvedValue(opts.chat),
+    };
+    const aiChatMessageRepo = {
+      findByChatUpdatedAfter: jest
+        .fn()
+        .mockResolvedValue(opts.delta ?? { rows: [], cursor: 'C1' }),
+    };
+    const controller = new AiChatController(
+      {} as never,
+      aiChatRunService as never,
+      aiChatRepo as never,
+      aiChatMessageRepo as never,
+      {} as never,
+      {} as never,
+    );
+    return { controller, aiChatRunService, aiChatRepo, aiChatMessageRepo };
+  }
+
+  it('owner-gates: a chat the user does not own throws, never reaching the repo', async () => {
+    const { controller, aiChatMessageRepo, aiChatRunService } = makeController({
+      chat: { id: 'c1', creatorId: 'someone-else' },
+    });
+    await expect(
+      controller.getMessagesDelta({ chatId: 'c1' }, user, workspace),
+    ).rejects.toBeInstanceOf(ForbiddenException);
+    expect(aiChatMessageRepo.findByChatUpdatedAfter).not.toHaveBeenCalled();
+    expect(aiChatRunService.getLatestForChat).not.toHaveBeenCalled();
+  });
+
+  it('returns { rows, cursor, run:{id,status} } with the run fact inlined', async () => {
+    const rows = [{ id: 'm1' }];
+    const { controller } = makeController({
+      chat: { id: 'c1', creatorId: 'u1' },
+      delta: { rows, cursor: 'C2' },
+      run: { id: 'r1', status: 'running', error: 'ignored', stepCount: 3 },
+    });
+    const res = await controller.getMessagesDelta(
+      { chatId: 'c1', cursor: 'C1' },
+      user,
+      workspace,
+    );
+    expect(res).toEqual({
+      rows,
+      cursor: 'C2',
+      // ONLY id + status — never the whole run row.
+      run: { id: 'r1', status: 'running' },
+    });
+  });
+
+  it('run is null when the chat has never had a run', async () => {
+    const { controller } = makeController({
+      chat: { id: 'c1', creatorId: 'u1' },
+      run: undefined,
+    });
+    const res = await controller.getMessagesDelta(
+      { chatId: 'c1' },
+      user,
+      workspace,
+    );
+    expect(res.run).toBeNull();
+  });
+
+  it('passes cursor through, defaulting a missing cursor to null (first poll)', async () => {
+    const { controller, aiChatMessageRepo } = makeController({
+      chat: { id: 'c1', creatorId: 'u1' },
+    });
+    await controller.getMessagesDelta({ chatId: 'c1' }, user, workspace);
+    expect(aiChatMessageRepo.findByChatUpdatedAfter).toHaveBeenCalledWith(
+      'c1',
+      'ws1',
+      null,
+    );
+    await controller.getMessagesDelta(
+      { chatId: 'c1', cursor: 'CX' },
+      user,
+      workspace,
+    );
+    expect(aiChatMessageRepo.findByChatUpdatedAfter).toHaveBeenLastCalledWith(
+      'c1',
+      'ws1',
+      'CX',
+    );
+  });
+});

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

@@ -51,6 +51,7 @@ import {
   ChatIdDto,
   ExportChatDto,
   GeneratePageTitleDto,
+  GetChatDeltaDto,
   GetChatMessagesDto,
   GetRunDto,
   RenameChatDto,
@@ -63,6 +64,47 @@ import {
   SUBSCRIBER_MAX_BUFFERED_BYTES,
 } from './ai-chat-stream-registry.service';
 import { startSseHeartbeat } from './sse-resilience';
+
+/**
+ * Write the attach TAIL to the hijacked socket in chunks that RESPECT drain
+ * (#491): each `write()` that returns false (the kernel buffer is full) is awaited
+ * on the next 'drain' before continuing. The old code wrote the whole buffer
+ * synchronously, which — with the pre-#491 32MB ring — spiked memory (half the
+ * OOM). Bails immediately if the socket ended/errored mid-write. Frames that the
+ * paused registry subscriber buffers while this awaits are delivered by start().
+ */
+async function writeTailRespectingDrain(
+  raw: {
+    write(chunk: string): boolean;
+    writableEnded?: boolean;
+    destroyed?: boolean;
+    once(event: string, cb: () => void): unknown;
+    removeListener?(event: string, cb: () => void): unknown;
+  },
+  frames: string[],
+): Promise<void> {
+  for (const frame of frames) {
+    if (raw.writableEnded || raw.destroyed) return;
+    const ok = raw.write(frame);
+    if (!ok) {
+      // Kernel buffer full — wait for drain (or an early close/error) before the
+      // next chunk, so a slow reader never forces the whole tail into memory.
+      // Remove ALL three listeners once any fires, so a many-chunk tail with
+      // repeated backpressure never leaks (MaxListenersExceededWarning).
+      await new Promise<void>((resolve) => {
+        const finish = (): void => {
+          raw.removeListener?.('drain', finish);
+          raw.removeListener?.('close', finish);
+          raw.removeListener?.('error', finish);
+          resolve();
+        };
+        raw.once('drain', finish);
+        raw.once('close', finish);
+        raw.once('error', finish);
+      });
+    }
+  }
+}
 import { EnvironmentService } from '../../integrations/environment/environment.service';
 
 /**
@@ -149,6 +191,46 @@ export class AiChatController {
     );
   }
 
+  /**
+   * Delta poll (#491) — the degraded-poll fallback's payload. Returns the chat's
+   * message rows changed since `cursor` (a DB-clock timestamp from the previous
+   * poll), a FRESH cursor, AND the current run fact `{ id, status } | null`. This
+   * replaces the old degraded poll that refetched ALL infinite-query pages (full
+   * parts) every 2.5s: the client seeds once and thereafter merges only the
+   * deltas by id (the overlap window guarantees repeats — the merge is idempotent,
+   * see mergeById). The run fact rides IN the delta (a separate /run poll would
+   * double the poll QPS), so the client FSM gets the run's status on the same tick.
+   * Owner-gated via assertOwnedChat (same gate as the other read endpoints).
+   */
+  @HttpCode(HttpStatus.OK)
+  @Post('messages/delta')
+  async getMessagesDelta(
+    @Body() dto: GetChatDeltaDto,
+    @AuthUser() user: User,
+    @AuthWorkspace() workspace: Workspace,
+  ): Promise<{
+    rows: AiChatMessage[];
+    cursor: string;
+    run: { id: string; status: string } | null;
+  }> {
+    await this.assertOwnedChat(dto.chatId, user, workspace);
+    const { rows, cursor } =
+      await this.aiChatMessageRepo.findByChatUpdatedAfter(
+        dto.chatId,
+        workspace.id,
+        dto.cursor ?? null,
+      );
+    const run = await this.aiChatRunService.getLatestForChat(
+      dto.chatId,
+      workspace.id,
+    );
+    return {
+      rows,
+      cursor,
+      run: run ? { id: run.id, status: run.status } : null,
+    };
+  }
+
   /**
    * Export a chat to Markdown (#183). The DB is the single source of truth: the
    * whole transcript is loaded (oldest -> newest) and rendered server-side. Now
@@ -249,19 +331,25 @@ export class AiChatController {
   }
 
   /**
-   * Attach to a chat's live run stream (#184 phase 1.5). A late/reloaded tab
-   * replays the frames buffered so far and then follows the live tail as a normal
-   * streamer. Owner-gated via assertOwnedChat (same gate as getRun). When there is
-   * nothing to resume — no entry, a finished run without expect=live, an
-   * overflowed buffer, or an anchor that pins a DIFFERENT run — the endpoint
-   * answers 204, the ONLY "nothing to resume" signal the AI SDK's reconnect
-   * accepts (it maps 204 to a silent no-op). With AI_CHAT_RESUMABLE_STREAM off the
-   * registry is never populated, so attach always 204s.
+   * Attach to a chat's live run stream from the client's step frontier (#184 phase
+   * 1.5, tail-only #491). A late/reloaded tab hands the server the step count it
+   * has PERSISTED (`n` = the seeded row's `metadata.stepsPersisted`) and its
+   * assistant row id (`anchor`); the registry answers with the TAIL past step `n`
+   * (a synthetic `start` frame + the buffered frames stamped >= n) and then the
+   * live tail. Owner-gated via assertOwnedChat (same gate as getRun). When there
+   * is nothing to resume — no entry, a ring that does not cover the client's
+   * frontier (overflow gap, or the client's seed lagged a rotation), or an anchor
+   * that pins a DIFFERENT run (invariant 6) — the endpoint answers 204, the ONLY
+   * "nothing to resume" signal the AI SDK's reconnect accepts (it maps 204 to a
+   * silent no-op); the client then refetches (a larger n) and re-attaches. With
+   * AI_CHAT_RESUMABLE_STREAM off the registry is never populated, so attach always
+   * 204s.
    *
-   * `expect=live` opts into replaying a finished-but-retained run (safe only when
-   * the client stripped the streaming tail); `anchor` is the client's assistant
-   * row id, which must match this run's (invariant 6) or a foreign run's
-   * transcript would be replayed into the store.
+   * The step marker `n` comes ONLY from the client — the server never reads the
+   * row to derive it, because a server-side n from a stale seed would open a
+   * silent one-step hole. The tail is written to the socket in CHUNKS respecting
+   * drain (writeTailRespectingDrain): the old code synchronously blasted the whole
+   * buffer, which — with the old 32MB cap — was half the OOM.
    */
   @SkipTransform()
   @UseGuards(JwtAuthGuard, UserThrottlerGuard)
@@ -269,39 +357,49 @@ export class AiChatController {
   @Get('runs/:chatId/stream')
   async attachRunStream(
     @Param('chatId', new ParseUUIDPipe()) chatId: string,
-    @Query('expect') expect: string | undefined,
     @Query('anchor') anchor: string | undefined,
+    @Query('n') n: string | undefined,
     @Req() req: FastifyRequest,
     @Res() res: FastifyReply,
     @AuthUser() user: User,
     @AuthWorkspace() workspace: Workspace,
   ): Promise<void> {
     await this.assertOwnedChat(chatId, user, workspace); // same gate as getRun
+    // The client's persisted step frontier. #491: distinguish a MISSING/invalid `n`
+    // (null — a NOT-tail-aware, legacy/parameterless tab expecting the old
+    // "finished -> 204 -> poll" contract) from `n=0` (a tail-aware client with
+    // nothing persisted yet). Passing 0 for a missing `n` would serve a finished,
+    // non-rotated run's WHOLE tail and a parameterless client would append it onto
+    // the steps it already shows -> #137/#161 duplicate. null makes the registry
+    // 204 such a finished run (see attach); a tail-aware n=0 still resumes.
+    const frontier: number | null =
+      n === undefined || n === '' || !Number.isFinite(Number(n))
+        ? null
+        : Math.max(0, Number(n));
+    // The per-subscriber backpressure cap tracks the (env-tunable) ring cap.
+    const subscriberCap =
+      this.streamRegistry?.subscriberMaxBufferedBytes ??
+      SUBSCRIBER_MAX_BUFFERED_BYTES;
     let stopHeartbeat: () => void = () => undefined;
-    const attachment = await this.streamRegistry?.attach(
-      chatId,
-      expect === 'live',
-      anchor,
-      {
-        onFrame: (frame) => {
-          // Backpressure guard: 2x the replay cap, so the initial replay burst
-          // alone can never trip it; only a genuinely stalled socket can.
-          try {
-            if (res.raw.writableLength > SUBSCRIBER_MAX_BUFFERED_BYTES) {
-              res.raw.destroy(); // 'close' fires -> unsubscribe below
-              return;
-            }
-            if (!res.raw.writableEnded) res.raw.write(frame);
-          } catch {
-            res.raw.destroy();
+    const attachment = await this.streamRegistry?.attach(chatId, anchor, frontier, {
+      onFrame: (frame) => {
+        // Backpressure guard: 2x the ring cap, so the initial tail burst alone
+        // can never trip it; only a genuinely stalled socket can.
+        try {
+          if (res.raw.writableLength > subscriberCap) {
+            res.raw.destroy(); // 'close' fires -> unsubscribe below
+            return;
           }
-        },
-        onEnd: () => {
-          stopHeartbeat();
-          if (!res.raw.writableEnded) res.raw.end();
-        },
+          if (!res.raw.writableEnded) res.raw.write(frame);
+        } catch {
+          res.raw.destroy();
+        }
       },
-    );
+      onEnd: () => {
+        stopHeartbeat();
+        if (!res.raw.writableEnded) res.raw.end();
+      },
+    });
     if (!attachment) {
       res.status(204).send(); // the ONLY "nothing to resume" signal the SDK accepts
       return;
@@ -330,13 +428,16 @@ export class AiChatController {
         // deliberately NO Connection/Keep-Alive (hop-by-hop; Safari/HTTP2)
       });
       res.raw.flushHeaders?.();
-      for (const frame of attachment.replay) res.raw.write(frame);
+      // Write the tail in chunks respecting drain (not a synchronous blast, which
+      // was half the OOM). Frames the paused subscriber buffers meanwhile are
+      // drained by start() below; its cap is the backstop for a stalled socket.
+      await writeTailRespectingDrain(res.raw, attachment.replay);
       if (attachment.finished) {
-        res.raw.end();
+        if (!res.raw.writableEnded) res.raw.end();
         return;
       }
       stopHeartbeat = startSseHeartbeat(res.raw, 15_000);
-      attachment.start(); // drain pending accumulated during replay, go live
+      attachment.start(); // drain pending accumulated during the tail write, go live
     } catch {
       attachment.unsubscribe();
       stopHeartbeat();
@@ -418,6 +519,19 @@ 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 +546,66 @@ 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 +614,94 @@ 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 +717,10 @@ 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

@@ -155,6 +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 () => ({})) };
@@ -179,7 +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 = {
@@ -285,7 +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;
@@ -312,7 +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 () => {
@@ -332,7 +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 () => {
@@ -361,6 +369,51 @@ 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);
+  });
 });
 
 /**
@@ -415,6 +468,8 @@ describe('AiChatService.stream — begin-failure fails the turn (#184 F14 / #486
       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 () => ({})) };

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

@@ -13,6 +13,7 @@ import {
   compactToolOutput,
   assistantParts,
   serializeSteps,
+  type StepPartsCache,
   rowToUiMessage,
   prepareAgentStep,
   stepBudgetWarning,
@@ -28,10 +29,14 @@ 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';
 import type { McpClientsService } from './external-mcp/mcp-clients.service';
+import { resolveEffectiveReplayThreshold } from './history-budget';
 
 /**
  * Unit tests for compactToolOutput: the pure helper that shrinks tool outputs
@@ -114,6 +119,54 @@ 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 +284,320 @@ 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);
+  });
+
+  // #490 reactive recovery: a prior turn stamped `replayOverflow` must make the
+  // NEXT turn's effective budget the AGGRESSIVE 0.5x cut — that harder trim is
+  // what un-bricks a chat that just 400'd on the context window. This exercises
+  // the exact wiring the service uses: read the stamp, then scale the threshold.
+  it('#490: a prior replayOverflow drives the next turn to the 0.5x aggressive budget', () => {
+    const history = [
+      row('assistant', { replayOverflow: true }),
+      row('user', null),
+    ];
+    const priorOverflowed = lastAssistantReplayOverflow(history);
+    expect(priorOverflowed).toBe(true);
+    // Base budget 100k -> aggressive recovery halves it to 50k this turn.
+    expect(resolveEffectiveReplayThreshold(100_000, priorOverflowed)).toBe(50_000);
+    // Odd base floors, not rounds.
+    expect(resolveEffectiveReplayThreshold(99_999, true)).toBe(49_999);
+    // No prior overflow -> the base budget is used verbatim (no aggressive cut).
+    expect(resolveEffectiveReplayThreshold(100_000, false)).toBe(100_000);
+    // An explicit off-switch (null) is never overridden, even on recovery.
+    expect(resolveEffectiveReplayThreshold(null, true)).toBeNull();
+  });
 });
 
 describe('rowToUiMessage', () => {
@@ -618,6 +930,23 @@ 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 +1644,12 @@ 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 +1690,33 @@ 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 +1769,7 @@ 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 +1777,11 @@ 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 +1816,7 @@ 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 +1899,86 @@ 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 +2066,19 @@ 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 +2338,148 @@ 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.step-marker.spec.ts

@@ -0,0 +1,65 @@
+import { flushAssistant } from './ai-chat.service';
+
+/**
+ * #491 STEP MARKER — `metadata.stepsPersisted` is written by the SAME flush that
+ * builds `metadata.parts`, so the marker can never disagree with the persisted
+ * parts (the step-alignment anchor the resume stack builds on). These are
+ * PROPERTY tests: they assert the marker tracks the number of FINISHED steps for
+ * every flush shape.
+ */
+
+// A finished step carrying one line of text and one tool call/result.
+function step(i: number) {
+  return {
+    text: `step ${i}`,
+    toolCalls: [
+      { toolCallId: `c${i}`, toolName: 'getPage', input: { id: `p${i}` } },
+    ],
+    toolResults: [
+      { toolCallId: `c${i}`, toolName: 'getPage', output: { title: `T${i}` } },
+    ],
+  };
+}
+
+describe('flushAssistant step marker (#491)', () => {
+  it('seed (no steps) → stepsPersisted 0', () => {
+    const f = flushAssistant([], '', 'streaming');
+    expect(f.metadata.stepsPersisted).toBe(0);
+  });
+
+  it('PROPERTY: stepsPersisted equals the number of FINISHED steps, for any N', () => {
+    for (let n = 0; n <= 6; n++) {
+      const steps = Array.from({ length: n }, (_, i) => step(i));
+      const f = flushAssistant(steps, '', 'streaming');
+      expect(f.metadata.stepsPersisted).toBe(n);
+      // ...and the parts actually contain those N steps' text (marker agrees with
+      // the persisted parts — the atomicity the whole design relies on).
+      const parts = f.metadata.parts as Array<Record<string, unknown>>;
+      const textParts = parts.filter((p) => p.type === 'text');
+      expect(textParts).toHaveLength(n);
+    }
+  });
+
+  it('an in-progress trailing partial does NOT increment the marker', () => {
+    // 2 finished steps + a partial (not-yet-finished) trailing text: the marker
+    // counts only the CONFIRMED step boundaries, not the partial.
+    const f = flushAssistant([step(0), step(1)], 'partial third step', 'error', {
+      error: 'boom',
+    });
+    expect(f.metadata.stepsPersisted).toBe(2);
+    // The partial text IS persisted in parts (so the user sees it), but it is not a
+    // counted step.
+    const parts = f.metadata.parts as Array<Record<string, unknown>>;
+    expect(parts[parts.length - 1]).toEqual({
+      type: 'text',
+      text: 'partial third step',
+    });
+  });
+
+  it('terminal completed flush counts all finished steps', () => {
+    const f = flushAssistant([step(0), step(1), step(2)], '', 'completed', {
+      finishReason: 'stop',
+    });
+    expect(f.metadata.stepsPersisted).toBe(3);
+  });
+});

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

@@ -1,4 +1,10 @@
-import { IsOptional, IsString, MaxLength, MinLength } from 'class-validator';
+import {
+  IsISO8601,
+  IsOptional,
+  IsString,
+  MaxLength,
+  MinLength,
+} from 'class-validator';
 
 /** Identify a chat by id (workspace-scoped on the server). */
 export class ChatIdDto {
@@ -37,6 +43,24 @@ export class GetChatMessagesDto {
   cursor?: string;
 }
 
+/**
+ * Delta poll (#491): pull the chat's rows changed since `cursor` (a DB-clock
+ * timestamp from the previous poll) plus the current run fact — the degraded-poll
+ * fallback's payload, replacing the full infinite-query refetch. Omit `cursor` on
+ * the first poll (returns just a fresh cursor to start the chain).
+ */
+export class GetChatDeltaDto {
+  @IsString()
+  chatId: string;
+
+  // ISO-8601 timestamp echoed from the previous poll's response. Validated as
+  // ISO-8601 (not a bare string): a malformed cursor would otherwise reach the
+  // `::timestamptz` cast in findByChatUpdatedAfter and 500 instead of a clean 400.
+  @IsOptional()
+  @IsISO8601()
+  cursor?: string;
+}
+
 /** Resolve the chat bound to a document (the page's most-recent owned chat). */
 export class BoundChatDto {
   @IsString()

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 floor(0.7 x window) for a configured window (no cap)', () => {
+    // 0.7 x 60k = 42k
+    expect(resolveReplayBudget(60_000)).toEqual({
+      thresholdTokens: 42_000,
+      usedDefault: false,
+    });
+    // 0.7 x 1M = 700k — NOT capped (anti-brick vs the window, not a cost limiter).
+    expect(resolveReplayBudget(1_000_000)).toEqual({
+      thresholdTokens: 700_000,
+      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 +1,375 @@
+/**
+ * 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  -> `floor(fraction × window)` (NO cap — the budgeter is
+ *    anti-brick protection against the window itself, not a cost/economy limiter,
+ *    exactly as the codebase already treats maxOutputTokens; the reactive branch
+ *    still guarantees anti-brick regardless of how high this budget is)
+ *  - 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.floor(REPLAY_BUDGET_WINDOW_FRACTION * n),
+    usedDefault: false,
+  };
+}
+
+/**
+ * The effective replay threshold for THIS turn, given the base budget and whether
+ * the PREVIOUS turn hit a context-overflow 400 (the reactive-recovery signal,
+ * `metadata.replayOverflow`). On recovery the base budget is scaled down by
+ * {@link REPLAY_AGGRESSIVE_FRACTION}: the overflowing turn produced no usage
+ * signal, so the preventive estimate under-counted and a normal-threshold trim may
+ * not shrink enough to fit — this harder cut is what un-bricks the chat.
+ *
+ * A `null` base budget (trimming OFF) is passed through unchanged: an explicit
+ * off-switch is never overridden by the recovery path.
+ */
+export function resolveEffectiveReplayThreshold(
+  thresholdTokens: number | null,
+  priorOverflowed: boolean,
+): number | null {
+  if (!priorOverflowed || thresholdTokens == null) return thresholdTokens;
+  return Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION);
+}
+
+/**
+ * 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/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/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.delta.int-spec.ts

@@ -0,0 +1,280 @@
+import { randomUUID } from 'crypto';
+import { CamelCasePlugin, Kysely, sql } from 'kysely';
+import { PostgresJSDialect } from 'kysely-postgres-js';
+// NOT a default import: the project tsconfig is `module: commonjs` with NO
+// esModuleInterop, so `import postgres from 'postgres'` compiles to
+// `postgres_1.default(...)` and the CJS `postgres` export has no `.default` —
+// it threw in beforeAll, was swallowed as "DB unreachable", and SILENTLY voided
+// all six tests. Mirror the working integration harness (test/integration/db.ts).
+import * as postgres from 'postgres';
+import { AiChatMessageRepo } from './ai-chat-message.repo';
+import { AiChatRunRepo } from './ai-chat-run.repo';
+
+/**
+ * #491 delta-poll — OBSERVABLE-PROPERTY tests against a LIVE Postgres (the local
+ * gitmost test DB, docker `gitmost-test-pg` on :5432), not "rows through a mock"
+ * (a mock cannot observe the DB clock nor the overlap-window race — the very
+ * things that matter here). Drives the REAL repo methods (`findByChatUpdatedAfter`,
+ * the now()-stamped `update`) and asserts:
+ *   1. delta-relevant writes stamp `updatedAt` from the DB clock, not the app clock
+ *      (proven by faking the process clock far into the future and observing the
+ *      stamp stays on real DB time);
+ *   2. the poll returns only rows changed after the cursor, ordered, with a fresh
+ *      DB-clock cursor;
+ *   3. the "committed late but stamped earlier than the cursor" RACE is caught by
+ *      the overlap window (a naive `updatedAt > cursor` would MISS it);
+ *   4. the overlap GUARANTEES repeats across close polls — the contract behind the
+ *      client's idempotent merge (mergeById).
+ *
+ * INTEGRATION lane (`*.int-spec.ts`): runs under `test:int`, whose global-setup
+ * DROPS + RE-CREATES + MIGRATES `docmost_test`, so the real `ai_chat_messages` /
+ * `ai_chat_runs` tables EXIST here. (It was previously a `.spec.ts` defaulting to
+ * the UNmigrated dev `docmost`; in the CI unit lane — where `WAL_TEST_DATABASE_URL`
+ * is unset and only `test:int` migrates — that meant 5/6 ERROR
+ * `relation "ai_chat_messages" does not exist`, silently voiding coverage of the
+ * risky cursor/overlap logic. Renaming to `.int-spec.ts` + defaulting the DSN to
+ * `docmost_test` fixes the CI fidelity.)
+ *
+ * FK triggers are bypassed (`session_replication_role = replica`) so synthetic
+ * chat/workspace ids need no parent fixtures; a single pooled connection (max 1)
+ * keeps that session setting for every query. SKIPS cleanly when the DB is
+ * unreachable so a DB-less CI never breaks.
+ */
+const CONN =
+  process.env.WAL_TEST_DATABASE_URL ??
+  process.env.TEST_DATABASE_URL ??
+  'postgresql://docmost:docmost_dev_pw@localhost:5432/docmost_test';
+
+let db: Kysely<any>;
+let sqlClient: ReturnType<typeof postgres>;
+let msgRepo: AiChatMessageRepo;
+let runRepo: AiChatRunRepo;
+let reachable = false;
+
+const workspaceId = randomUUID();
+const chatId = randomUUID();
+
+beforeAll(async () => {
+  try {
+    sqlClient = postgres(CONN, { max: 1, onnotice: () => {} });
+    db = new Kysely<any>({
+      dialect: new PostgresJSDialect({ postgres: sqlClient }),
+      plugins: [new CamelCasePlugin()],
+    });
+    // Single connection keeps this session-scoped bypass for the whole suite.
+    await sql`set session_replication_role = replica`.execute(db);
+    await sql`select 1`.execute(db);
+    reachable = true;
+  } catch (err) {
+    reachable = false;
+    // A genuine connection failure (ECONNREFUSED etc.) is a legitimate skip on a
+    // DB-less CI. A PROGRAMMING error (bad import, typo, driver misuse) must NOT
+    // masquerade as "DB unreachable" and silently void the whole suite (that is
+    // exactly the bug that hid this spec's zero coverage) — rethrow it so the
+    // suite fails LOUDLY.
+    const msg = String((err as Error)?.message ?? err);
+    if (
+      !/ECONNREFUSED|ENOTFOUND|ETIMEDOUT|EHOSTUNREACH|connect|terminating|password|authentication|role .* does not exist|database .* does not exist/i.test(
+        msg,
+      )
+    ) {
+      throw err;
+    }
+  }
+  msgRepo = new AiChatMessageRepo(db as never);
+  runRepo = new AiChatRunRepo(db as never);
+});
+
+afterAll(async () => {
+  if (db) {
+    try {
+      await db
+        .deleteFrom('aiChatMessages')
+        .where('workspaceId', '=', workspaceId)
+        .execute();
+      await db
+        .deleteFrom('aiChatRuns')
+        .where('workspaceId', '=', workspaceId)
+        .execute();
+    } catch {
+      /* best-effort cleanup */
+    }
+    await db.destroy();
+  }
+});
+
+afterEach(() => {
+  jest.useRealTimers();
+});
+
+async function seedMessage(overrides: Record<string, unknown> = {}) {
+  return msgRepo.insert({
+    chatId,
+    workspaceId,
+    userId: null as never,
+    role: 'assistant',
+    content: 'x',
+    status: 'streaming',
+    ...overrides,
+  } as never);
+}
+
+async function dbNow(): Promise<string> {
+  const r = await sql<{ now: Date }>`select now() as now`.execute(db);
+  return r.rows[0].now.toISOString();
+}
+
+// Fake ONLY the Date object (so in-process `new Date()`/`Date.now()` jump), while
+// leaving every TIMER function real. Faking timers wholesale freezes postgres.js's
+// internal connection/query timers, so the awaited DB round-trip would hang the
+// test (and the afterAll cleanup) at the jest 5s cap. With Date-only faking the
+// query resolves normally, and we still prove the stamp is the DB clock (not the
+// skewed process clock).
+function fakeDateOnly(iso: string): void {
+  jest.useFakeTimers({
+    doNotFake: [
+      'hrtime',
+      'nextTick',
+      'performance',
+      'queueMicrotask',
+      'requestAnimationFrame',
+      'cancelAnimationFrame',
+      'requestIdleCallback',
+      'cancelIdleCallback',
+      'setImmediate',
+      'clearImmediate',
+      'setInterval',
+      'clearInterval',
+      'setTimeout',
+      'clearTimeout',
+    ],
+    now: new Date(iso),
+  });
+}
+
+const maybe = (name: string, fn: () => Promise<void>) =>
+  it(name, async () => {
+    if (!reachable) {
+      console.warn(`SKIP (${name}): test DB unreachable at ${CONN}`);
+      return;
+    }
+    await fn();
+  });
+
+describe('AiChatMessageRepo.findByChatUpdatedAfter (#491 delta poll)', () => {
+  maybe('null cursor returns no rows and a fresh DB-clock cursor', async () => {
+    const before = await dbNow();
+    const { rows, cursor } = await msgRepo.findByChatUpdatedAfter(
+      chatId,
+      workspaceId,
+      null,
+    );
+    expect(rows).toEqual([]);
+    expect(new Date(cursor).getTime()).toBeGreaterThanOrEqual(
+      new Date(before).getTime(),
+    );
+  });
+
+  maybe('returns only rows changed after the cursor', async () => {
+    const { cursor: c0 } = await msgRepo.findByChatUpdatedAfter(
+      chatId,
+      workspaceId,
+      null,
+    );
+    const m = await seedMessage();
+    const { rows, cursor: c1 } = await msgRepo.findByChatUpdatedAfter(
+      chatId,
+      workspaceId,
+      c0,
+    );
+    expect(rows.map((r) => r.id)).toContain(m.id);
+    // Cursor is monotonic (advances).
+    expect(new Date(c1).getTime()).toBeGreaterThanOrEqual(
+      new Date(c0).getTime(),
+    );
+  });
+
+  maybe(
+    'RACE: a row stamped BEFORE the cursor but seen after is caught by the overlap',
+    async () => {
+      // Cursor taken now; then a row appears whose updatedAt is 2s in the PAST
+      // (committed late on another connection but stamped earlier). A naive
+      // `updatedAt > cursor` would MISS it; the 5s overlap window catches it.
+      const cursor = await dbNow();
+      const m = await seedMessage();
+      await sql`update ai_chat_messages set updated_at = now() - interval '2 seconds' where id = ${m.id}`.execute(
+        db,
+      );
+      const { rows } = await msgRepo.findByChatUpdatedAfter(
+        chatId,
+        workspaceId,
+        cursor,
+      );
+      expect(rows.map((r) => r.id)).toContain(m.id);
+    },
+  );
+
+  maybe(
+    'overlap GUARANTEES repeats across close polls (idempotent-merge contract)',
+    async () => {
+      const { cursor: c0 } = await msgRepo.findByChatUpdatedAfter(
+        chatId,
+        workspaceId,
+        null,
+      );
+      const m = await seedMessage();
+      const first = await msgRepo.findByChatUpdatedAfter(
+        chatId,
+        workspaceId,
+        c0,
+      );
+      expect(first.rows.map((r) => r.id)).toContain(m.id);
+      // Immediately re-poll with the JUST-returned cursor: the row is still within
+      // the overlap window, so it is returned AGAIN — the client MUST dedupe by id.
+      const second = await msgRepo.findByChatUpdatedAfter(
+        chatId,
+        workspaceId,
+        first.cursor,
+      );
+      expect(second.rows.map((r) => r.id)).toContain(m.id);
+    },
+  );
+
+  maybe(
+    'update() stamps updatedAt from the DB clock, not the app clock',
+    async () => {
+      const m = await seedMessage();
+      // Skew the PROCESS clock ~73 years into the future (Date only). If the stamp
+      // came from `new Date()` the row would read year 2099; sql now() keeps it on
+      // DB time.
+      fakeDateOnly('2099-01-01T00:00:00Z');
+      const updated = await msgRepo.update(m.id, workspaceId, {
+        content: 'y',
+      });
+      jest.useRealTimers();
+      expect(updated).toBeDefined();
+      expect(new Date(updated!.updatedAt).getFullYear()).toBeLessThan(2099);
+    },
+  );
+
+  maybe(
+    'run update() also stamps updatedAt from the DB clock',
+    async () => {
+      const run = await runRepo.insert({
+        chatId,
+        workspaceId,
+        createdBy: null as never,
+        trigger: 'user',
+        status: 'running',
+        stepCount: 0,
+      } as never);
+      fakeDateOnly('2099-01-01T00:00:00Z');
+      const updated = await runRepo.update(run.id, workspaceId, {
+        stepCount: 1,
+      });
+      jest.useRealTimers();
+      expect(updated).toBeDefined();
+      expect(new Date(updated!.updatedAt).getFullYear()).toBeLessThan(2099);
+    },
+  );
+});

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 {
@@ -24,6 +25,20 @@ const SWEEP_STREAMING_STALE_MS = 10 * 60 * 1000; // 10 minutes
 // into memory; far above any realistic transcript length.
 const FIND_ALL_BY_CHAT_LIMIT = 5000;
 
+// Delta-poll overlap (#491): the poll query reaches this far BEHIND the client's
+// echoed cursor, so a row that committed with an `updatedAt` marginally before the
+// previous cursor was taken (on another autocommit connection) is still caught.
+// Sized well above realistic single-row commit skew; the client merge is
+// idempotent by id (mergeById), so the guaranteed repeats the overlap produces are
+// harmless.
+export const DELTA_POLL_OVERLAP_SECONDS = 5;
+
+// Hard cap on rows one delta poll returns — a safety bound (a poll should carry a
+// handful of just-changed rows, never a whole transcript). Ordered by (updatedAt,
+// id) asc, so on the pathological overflow the OLDEST changes win and the newest
+// are picked up by the next poll (its cursor did not advance past them).
+export const DELTA_POLL_MAX_ROWS = 500;
+
 @Injectable()
 export class AiChatMessageRepo {
   private readonly logger = new Logger(AiChatMessageRepo.name);
@@ -138,6 +153,72 @@ export class AiChatMessageRepo {
       .executeTakeFirst();
   }
 
+  /**
+   * Delta read (#491) for the degraded poll: the chat's messages whose row
+   * changed AFTER the client's `cursor`, plus a FRESH cursor taken from the DB
+   * clock. Replaces the old "refetch ALL infinite-query pages every 2.5s with
+   * full parts" poll — the client seeds once (findByChat) and thereafter pulls
+   * only the deltas and merges them by id (mergeById).
+   *
+   * Cursor: a DB-clock timestamp (now()) the client echoes back each poll. All
+   * delta-relevant writes stamp `updatedAt` with now() (see `update` /
+   * `finalizeOwner`), so this is a SINGLE monotonic axis. The query overlaps the
+   * cursor by DELTA_POLL_OVERLAP_SECONDS to catch a row committed with an
+   * `updatedAt` marginally BEFORE the previous cursor was taken on another
+   * connection (single-row autocommit UPDATEs; no long transactions). The overlap
+   * GUARANTEES occasional REPEATS, so the client merge MUST be idempotent by id.
+   *
+   * `cursor === null` (first poll after the full seed) returns NO rows — there is
+   * nothing "new" relative to a just-loaded seed — only the fresh cursor to start
+   * the delta chain. The fresh cursor is read AFTER the rows, so it is >= every
+   * returned row's `updatedAt` (they were read strictly earlier) — a row that
+   * commits between the rows-read and the cursor-read is at most
+   * DELTA_POLL_OVERLAP_SECONDS behind the returned cursor, so the next poll's
+   * overlap window always re-includes it (no miss).
+   */
+  async findByChatUpdatedAfter(
+    chatId: string,
+    workspaceId: string,
+    cursor: string | null,
+  ): Promise<{ rows: AiChatMessage[]; cursor: string }> {
+    if (cursor === null) {
+      const nowRow = await sql<{ now: Date }>`select now() as now`.execute(
+        this.db,
+      );
+      return { rows: [], cursor: nowRow.rows[0].now.toISOString() };
+    }
+    // Overlap the client cursor by DELTA_POLL_OVERLAP_SECONDS, computed in SQL off
+    // the echoed cursor so the whole comparison stays on the DB clock.
+    const rows = await this.db
+      .selectFrom('aiChatMessages')
+      .select(this.baseFields)
+      .where('chatId', '=', chatId)
+      .where('workspaceId', '=', workspaceId)
+      .where('deletedAt', 'is', null)
+      .where(
+        'updatedAt',
+        '>',
+        sql<Date>`${cursor}::timestamptz - make_interval(secs => ${DELTA_POLL_OVERLAP_SECONDS})`,
+      )
+      .orderBy('updatedAt', 'asc')
+      .orderBy('id', 'asc')
+      .limit(DELTA_POLL_MAX_ROWS)
+      .execute();
+    // When the page filled (pathological overflow), DO NOT advance the cursor to
+    // now(): that would skip the changed rows past the cap that this poll did not
+    // return. Resume from the last returned row's updatedAt instead (the next
+    // poll's overlap re-includes ties by id). In the normal case the fresh DB-clock
+    // now() is the cursor.
+    if (rows.length === DELTA_POLL_MAX_ROWS) {
+      return {
+        rows,
+        cursor: rows[rows.length - 1].updatedAt.toISOString(),
+      };
+    }
+    const nowRow = await sql<{ now: Date }>`select now() as now`.execute(this.db);
+    return { rows, cursor: nowRow.rows[0].now.toISOString() };
+  }
+
   async insert(
     insertable: InsertableAiChatMessage,
     trx?: KyselyTransaction,
@@ -171,7 +252,13 @@ export class AiChatMessageRepo {
     const db = dbOrTx(this.db, opts?.trx);
     let query = db
       .updateTable('aiChatMessages')
-      .set({ ...(patch as Record<string, unknown>), updatedAt: new Date() })
+      // #491: stamp `updatedAt` from the DB clock (sql now()), NOT the app clock
+      // (new Date()). The delta-poll cursor (findByChatUpdatedAfter) is a single
+      // DB-clock axis; a per-step 'streaming' UPDATE stamped with the app clock
+      // would be a SECOND, skewed clock source and could leave a row's updatedAt
+      // just under a cursor taken from now() on another connection — an
+      // independent source of delta MISSES. All delta-relevant writes use now().
+      .set({ ...(patch as Record<string, unknown>), updatedAt: sql`now()` })
       .where('id', '=', id)
       .where('workspaceId', '=', workspaceId);
     // Concurrency guard (#183 review): a per-step 'streaming' update must NEVER
@@ -188,6 +275,150 @@ 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')
+      // #491: DB-clock stamp (see `update`) — this terminal write flips the row's
+      // status, which the delta poll must observe on the shared now() cursor axis.
+      .set({ ...(patch as Record<string, unknown>), updatedAt: sql`now()` })
+      .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)`,
+        // #491: DB-clock stamp (see `update`) so a reconcile status flip lands on
+        // the same now() cursor axis the delta poll reads.
+        updatedAt: sql`now()`,
+      })
+      .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)`,
+        // #491: DB-clock stamp (see `update`). The staleness WHERE below stays on
+        // the app clock — a >minutes window makes the ms-scale skew irrelevant.
+        updatedAt: sql`now()`,
+      })
+      .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 +431,21 @@ 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)`,
+        // #491: DB-clock stamp (see `update`). Staleness WHERE stays app-clock.
+        updatedAt: sql`now()`,
+      })
       .where('status', '=', 'streaming')
       .where('updatedAt', '<', staleBefore)
       .returning('id')

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

@@ -62,10 +62,17 @@ describe('AiChatRunRepo.sweepRunning', () => {
     // ...but a fresh 'running' run (updatedAt = now) must NOT be skipped: no
     // updatedAt predicate at all on the boot path.
     expect(rec.wheres.some(([col]) => col === 'updatedAt')).toBe(false);
-    // It flips to 'aborted' and stamps finishedAt.
-    expect(rec.set).toEqual(
-      expect.objectContaining({ status: 'aborted', finishedAt: expect.any(Date) }),
-    );
+    // It flips to 'aborted' and stamps finishedAt + updatedAt. #491: the stamps
+    // are now DB-clock `sql now()` expressions (raw builders), NOT app-clock
+    // `new Date()`, so the run row shares the delta poll's single now() cursor axis
+    // — assert they are present and are the sql raw-builder objects (not a Date,
+    // not undefined).
+    expect(rec.set?.status).toBe('aborted');
+    for (const stamp of ['finishedAt', 'updatedAt'] as const) {
+      expect(rec.set?.[stamp]).toBeDefined();
+      expect(rec.set?.[stamp]).not.toBeInstanceOf(Date);
+      expect(typeof rec.set?.[stamp]).toBe('object');
+    }
   });
 
   it('phase-2 path: an explicit staleMs reintroduces the updatedAt window', async () => {

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

@@ -136,13 +136,53 @@ export class AiChatRunRepo {
     const db = dbOrTx(this.db, trx);
     return db
       .updateTable('aiChatRuns')
-      .set({ ...(patch as Record<string, unknown>), updatedAt: new Date() })
+      // #491: DB-clock stamp (sql now()) so the run row shares the delta poll's
+      // single now() cursor axis with the assistant message rows — a run-status
+      // change (the run fact the delta carries) must never sit on a skewed app
+      // clock relative to the message updatedAt cursor.
+      .set({ ...(patch as Record<string, unknown>), updatedAt: sql`now()` })
       .where('id', '=', id)
       .where('workspaceId', '=', workspaceId)
       .returning(this.baseFields)
       .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);
+    return db
+      .updateTable('aiChatRuns')
+      .set({
+        status: patch.status,
+        error: patch.error,
+        // #491: DB-clock stamps (finished_at + updated_at) so the terminal run
+        // fact lands on the delta poll's now() cursor axis.
+        finishedAt: sql`now()`,
+        updatedAt: sql`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
@@ -157,7 +197,8 @@ export class AiChatRunRepo {
     const db = dbOrTx(this.db, trx);
     return db
       .updateTable('aiChatRuns')
-      .set({ stopRequestedAt: new Date(), updatedAt: new Date() })
+      // #491: DB-clock stamps (see `update`).
+      .set({ stopRequestedAt: sql`now()`, updatedAt: sql`now()` })
       .where('id', '=', id)
       .where('workspaceId', '=', workspaceId)
       .where('status', 'in', ACTIVE_RUN_STATUSES as unknown as string[])
@@ -184,18 +225,44 @@ 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,
   ): Promise<number> {
     const db = dbOrTx(this.db, trx);
-    const now = new Date();
     let query = db
       .updateTable('aiChatRuns')
       .set({
         status: 'aborted',
-        finishedAt: now,
-        updatedAt: now,
+        // #491: DB-clock stamps (see `update`). The staleness WHERE below stays on
+        // the app clock — a >minutes window makes the ms-scale skew irrelevant.
+        finishedAt: sql`now()`,
+        updatedAt: sql`now()`,
         error: sql`coalesce(error, ${'Run interrupted by a server restart.'})`,
       })
       .where('status', 'in', ACTIVE_RUN_STATUSES as unknown as string[]);
@@ -203,7 +270,7 @@ export class AiChatRunRepo {
     // sibling replica's live run is never aborted. Omitted on the phase-1 boot
     // sweep -> unconditional.
     if (typeof opts.staleMs === 'number') {
-      const staleBefore = new Date(now.getTime() - opts.staleMs);
+      const staleBefore = new Date(Date.now() - opts.staleMs);
       query = query.where('updatedAt', '<', staleBefore);
     }
     const rows = await query.returning('id').execute();

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

@@ -30,11 +30,13 @@ import {
  * tees the SSE frames into it via `consumeSseStream` while stamping the DB row id
  * via `generateMessageId` (both gated on runId + the resumable flag).
  *
- * Proven here: a finished run's replay is the full frame sequence incl `[DONE]`
- * with `start.messageId` == the seeded DB row id; the anchor check (invariant 6);
- * an attach opened BEFORE the first frame follows the live stream from frame 0; an
- * explicit stop surfaces `{"type":"abort"}` + `[DONE]` + end to the subscriber;
- * and the legacy (non-run) path tees nothing.
+ * Proven here (tail-only #491): a finished run attached at its persisted frontier
+ * N_final delivers only the TAIL past N (a synthetic `start` carrying the run-fact
+ * + the terminal `finish`/`[DONE]`) — the step content below N lives in the seeded
+ * DB row, NOT the ring; the anchor check (invariant 6); an attach opened BEFORE the
+ * first frame follows the live stream from frame 0; an explicit stop surfaces
+ * `{"type":"abort"}` + `[DONE]` + end to the subscriber; and the legacy (non-run)
+ * path tees nothing.
  */
 
 const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
@@ -133,14 +135,16 @@ function liveSink(): {
   };
 }
 
-// The SSE `start` frame carries the message id; pull it out of a `data: {...}`.
-function parseStartMessageId(frames: string[]): string | undefined {
+// Parse the first `start` frame's JSON out of a `data: {...}` sequence.
+function parseStartFrame(
+  frames: string[],
+): { messageId?: string; messageMetadata?: any } | undefined {
   for (const f of frames) {
     const m = /^data: (\{.*\})\s*$/m.exec(f.trim());
     if (!m) continue;
     try {
       const json = JSON.parse(m[1]);
-      if (json.type === 'start') return json.messageId;
+      if (json.type === 'start') return json;
     } catch {
       /* not this frame */
     }
@@ -270,7 +274,7 @@ describe('AiChatService run-stream attach [integration]', () => {
     await destroyTestDb();
   });
 
-  it('run-wrapped: replay is the full frame sequence incl [DONE], start.messageId == the seeded DB row id', async () => {
+  it('run-wrapped, tail-only: a finished run at N_final delivers the run-fact start + finish/[DONE]; the step content lives in the seeded row', async () => {
     const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
     const registry = new AiChatStreamRegistryService();
     const runService = new AiChatRunService(runRepo, {
@@ -300,27 +304,37 @@ describe('AiChatService run-stream attach [integration]', () => {
         );
       });
       const rowId = await assistantRowId(chatId);
+      // The client reads its persisted step frontier N from the seeded row.
+      const row: any = await msgRepo.findById(rowId, workspaceId);
+      const nFinal = row.metadata.stepsPersisted as number;
+      expect(nFinal).toBe(1); // a single finished step
+      // The step content is in the SEEDED row (parts/content), not the ring.
+      expect(JSON.stringify(row.metadata.parts)).toContain('Hello');
 
-      // Finished-run replay with expect=live + the correct anchor.
+      // Attach at N_final with the correct anchor: the tail past step 1 is just
+      // the terminal frames; step 0's 'Hello' is BELOW the frontier (seeded).
       const sink = liveSink();
-      const att = await registry.attach(chatId, true, rowId, sink.cb);
+      const att = await registry.attach(chatId, rowId, nFinal, sink.cb);
       expect(att).not.toBeNull();
       expect(att!.finished).toBe(true);
-      // The tee captured frames (consumeSseStream was wired).
-      expect(att!.replay.length).toBeGreaterThan(0);
-      // generateMessageId stamped the DB row id onto the streamed start frame.
-      expect(parseStartMessageId(att!.replay)).toBe(rowId);
-      // The full sequence includes the streamed text and the terminal marker.
-      const joined = att!.replay.join('');
-      expect(joined).toContain('Hello');
+      // The synthetic start frame carries the run-fact (runId/chatId), the source
+      // of the run-fact on re-attach.
+      const start = parseStartFrame(att!.replay);
+      expect(start?.messageMetadata).toMatchObject({
+        runId: box.runId,
+        chatId,
+      });
+      // The terminal marker is delivered so the client's SDK closes the stream.
       expect(att!.replay.some((f) => f.includes('[DONE]'))).toBe(true);
+      // 'Hello' (step 0, below the frontier) is NOT re-streamed — it is seeded.
+      expect(att!.replay.some((f) => f.includes('Hello'))).toBe(false);
     } finally {
       registry.onModuleDestroy();
       await cleanup();
     }
   });
 
-  it('anchor mismatch with expect=live returns null (invariant 6)', async () => {
+  it('anchor mismatch returns null (invariant 6)', async () => {
     const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
     const registry = new AiChatStreamRegistryService();
     const runService = new AiChatRunService(runRepo, {
@@ -347,7 +361,7 @@ describe('AiChatService run-stream attach [integration]', () => {
       const sink = liveSink();
       // A foreign anchor must NOT replay this run's transcript.
       expect(
-        await registry.attach(chatId, true, 'a-different-run-row', sink.cb),
+        await registry.attach(chatId, 'a-different-run-row', 1, sink.cb),
       ).toBeNull();
     } finally {
       registry.onModuleDestroy();
@@ -376,8 +390,11 @@ describe('AiChatService run-stream attach [integration]', () => {
     try {
       // Attach while the entry exists (opened at begin) but before any frame.
       const sink = liveSink();
-      const att = (await registry.attach(chatId, false, undefined, sink.cb))!;
-      expect(att.replay).toEqual([]); // nothing streamed yet -> replay from 0
+      const att = (await registry.attach(chatId, undefined, 0, sink.cb))!;
+      // Nothing streamed yet -> the tail is just the synthetic start frame; the
+      // whole live stream (start..DONE) follows via onFrame after start().
+      expect(att.replay).toHaveLength(1);
+      expect(att.replay[0]).toContain('"type":"start"');
       att.start(); // go live (drains nothing, then follows)
 
       // Now emit the whole turn.
@@ -448,7 +465,7 @@ describe('AiChatService run-stream attach [integration]', () => {
     });
     try {
       const sink = liveSink();
-      const att = (await registry.attach(chatId, false, undefined, sink.cb))!;
+      const att = (await registry.attach(chatId, undefined, 0, sink.cb))!;
       att.start();
 
       // Give streamText a beat to begin consuming the partial output.
@@ -526,7 +543,9 @@ describe('AiChatService run-stream attach [integration]', () => {
       expect(entry).toBeDefined();
       expect(entry.finished).toBe(true);
       const sink = liveSink();
-      expect(await registry.attach(chatId, false, undefined, sink.cb)).toBeNull();
+      // Finished with an EMPTY ring (aborted before any frame) -> null -> the
+      // client degrades to poll instead of hanging on an empty stream.
+      expect(await registry.attach(chatId, undefined, 0, sink.cb)).toBeNull();
     } finally {
       registry.onModuleDestroy();
       await cleanup();
@@ -556,8 +575,8 @@ describe('AiChatService run-stream attach [integration]', () => {
       });
       const sink = liveSink();
       // No entry was ever opened; attach always yields null.
-      expect(await registry.attach(chatId, false, undefined, sink.cb)).toBeNull();
-      expect(await registry.attach(chatId, true, 'anything', sink.cb)).toBeNull();
+      expect(await registry.attach(chatId, undefined, 0, sink.cb)).toBeNull();
+      expect(await registry.attach(chatId, 'anything', 1, sink.cb)).toBeNull();
     } finally {
       registry.onModuleDestroy();
       await cleanup();

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

apps/server/test/jest-integration.json

@@ -22,6 +22,7 @@
     "^@docmost/db/(.*)$": "<rootDir>/src/database/$1",
     "^@docmost/transactional/(.*)$": "<rootDir>/src/integrations/transactional/$1",
     "^@docmost/ee/(.*)$": "<rootDir>/src/ee/$1",
+    "^@docmost/token-estimate$": "<rootDir>/../../packages/token-estimate/src/index.ts",
     "^src/(.*)$": "<rootDir>/src/$1"
   }
 }

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/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

@@ -170,6 +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
@@ -571,6 +608,10 @@ export abstract class DocmostClientContext {
           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.
@@ -602,6 +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,
@@ -709,7 +755,15 @@ export abstract class DocmostClientContext {
     // #486: on a rejected collab-WS handshake, invalidate + refresh the token and
     // retry the write once (symmetric to the HTTP-401 reauth path).
     return this.writeWithCollabAuthRetry(collabToken, (token) =>
-      mutatePageContent(pageUuid, token, apiUrl, transform),
+      // #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,
+      ),
     );
   }
 
@@ -733,7 +787,14 @@ export abstract class DocmostClientContext {
     // #486: on a rejected collab-WS handshake, invalidate + refresh the token and
     // retry the write once (symmetric to the HTTP-401 reauth path).
     return this.writeWithCollabAuthRetry(collabToken, (token) =>
-      replacePageContent(pageUuid, doc, token, apiUrl),
+      // #487: same pre-commit safe-point as mutatePage, for full-document writes.
+      replacePageContent(
+        pageUuid,
+        doc,
+        token,
+        apiUrl,
+        this.toolAbortSignal ?? undefined,
+      ),
     );
   }
 

packages/mcp/src/index.ts

@@ -30,6 +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/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/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 ' +
@@ -919,6 +951,7 @@ 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 ' +
@@ -965,6 +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 ' +
@@ -1009,6 +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 ' +
@@ -1038,6 +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 ' +
@@ -1088,6 +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.',
@@ -1181,6 +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.',
@@ -1202,6 +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.',
@@ -1235,6 +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 ' +
@@ -1291,6 +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 ' +
@@ -1325,6 +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 ' +
@@ -1373,6 +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
@@ -1505,6 +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.
@@ -1532,6 +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.
@@ -1572,6 +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.
@@ -1651,6 +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 ' +
@@ -1684,6 +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 ' +
@@ -1707,6 +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). ' +
@@ -1747,6 +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 ' +
@@ -1784,6 +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 ' +
@@ -1828,6 +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 ' +
@@ -1866,6 +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 / ' +
@@ -1899,6 +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 ' +
@@ -1969,6 +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 ' +
@@ -2020,6 +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 ' +
@@ -2078,6 +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 ' +
@@ -2202,6 +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: ' +
@@ -2248,6 +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 ' +
@@ -2294,6 +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: ' +
@@ -2323,3 +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/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

@@ -442,3 +442,139 @@ 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",
+  );
+});
+
+// -----------------------------------------------------------------------------
+// 7) checkNewComments partial failure (#490): the concurrent scan is resilient —
+//    if ONE page's /comments fetch rejects (deleted mid-scan, a transient 500),
+//    that page is skipped and the WHOLE scan still resolves with every other
+//    page's fresh comments. A single failing fetch must never reject the batch
+//    (Promise.all in mapWithConcurrency would otherwise abort all of it) nor
+//    corrupt the deterministic order of the pages that DID succeed.
+// -----------------------------------------------------------------------------
+test("checkNewComments skips a page whose fetch fails and still reports the rest", async () => {
+  const NODES = [{ id: "parent", title: "Parent", parentPageId: null, hasChildren: true }];
+  for (let i = 0; i < 5; i++) {
+    NODES.push({ id: `k${i}`, title: `Kid ${i}`, parentPageId: "parent", hasChildren: false });
+  }
+  // The one page whose comment fetch blows up (500 -> listComments rejects).
+  const FAILING = "k2";
+
+  const { baseURL } = await spawn(async (req, res) => {
+    const raw = await readBody(req);
+    if (handleLogin(req, res)) return;
+    if (req.url === "/api/pages/tree") {
+      sendJson(res, 200, { success: true, data: { items: NODES } });
+      return;
+    }
+    if (req.url === "/api/comments") {
+      const body = JSON.parse(raw || "{}");
+      if (body.pageId === FAILING) {
+        // A transient server error on exactly one page's fetch.
+        sendJson(res, 500, { success: false, message: "boom" });
+        return;
+      }
+      sendJson(res, 200, {
+        success: true,
+        data: {
+          items: [
+            { id: `c-${body.pageId}`, createdAt: "2030-01-01T00:00:00.000Z", content: null },
+          ],
+          meta: { nextCursor: null },
+        },
+      });
+      return;
+    }
+    sendJson(res, 404, {});
+  });
+
+  const client = new DocmostClient(baseURL, "user@example.com", "pw");
+  // Must RESOLVE (not reject) despite one page's fetch failing.
+  const result = await client.checkNewComments(
+    "space-1",
+    "2020-01-01T00:00:00.000Z",
+    "parent",
+  );
+
+  // Every page in scope was still scanned (the failing one counts as checked).
+  assert.equal(result.checkedPages, 6, "all pages scanned incl. the failing one");
+  // The failing page contributes nothing; the other 5 each report one comment.
+  assert.equal(result.pagesWithNewComments, 5, "the failing page is dropped");
+  assert.equal(result.totalNewComments, 5, "only the succeeding pages' comments");
+  const reportedIds = result.comments.map((r) => r.pageId);
+  assert.ok(!reportedIds.includes(FAILING), "the failing page is absent from results");
+  // Order of the survivors is still the deterministic enumeration order (the hole
+  // left by the failing page is closed without reordering the rest).
+  assert.deepEqual(
+    reportedIds,
+    ["parent", "k0", "k1", "k3", "k4"],
+    "survivors keep enumeration order with the failing page removed",
+  );
+});

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/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"]
+}

pnpm-lock.yaml

@@ -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
@@ -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':