refactor(mcp): убрать мёртвый runtime-слой page-id, оставить бренд PageId

Ревью #516: isPageId/asPageId/isSlugId/asSlugId/SLUG_ID_RE + типы SlugId/PageRef имели НОЛЬ вызовов в монорепе — только барные ре-экспорты в client.ts и кейсы теста. Несущий смысл — только compile-time бренд PageId (минтится as PageId в resolvePageId, единственном узле канонизации; asPageId туда намеренно не звался). Докстринг заявлял «asPageId() guards the untrusted PUBLIC boundary» — но никто не звал: спекулятивный вес. Удалил мёртвые runtime-символы + типы SlugId/PageRef + ре-экспорты + их тест-файл. Оставил тип PageId + касты. Поправил врущий коммент в resolvePageId (бренд — чистый compile-time маркер, гарантия — что этот узел единственный производитель). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
test(db): executeTx — покрыть rollback-ветку after-commit-хука
2026-07-12 01:02:13 +03:00 · 2026-07-12 01:02:13 +03:00 · 2026-07-12 01:02:13 +03:00 · 2026-07-12 01:02:13 +03:00 · 2026-07-12 01:02:13 +03:00 · 2026-07-12 01:02:13 +03:00
57 changed files with 2188 additions and 359 deletions
@@ -124,9 +124,17 @@ jobs:
          exit "$FAILED"

      # A GENUINE counterexample: fast-check printed a shrunk minimal case and its
-      # reproducing seed into property-output.txt. File a dedup-guarded issue whose
-      # title prefix is UNIQUE to counterexamples, so an infra failure (handled by
-      # the next step under a different title) can never poison this dedup.
+      # reproducing seed into property-output.txt. File a dedup-guarded issue.
+      #
+      # Dedup is keyed on a HASH of the SHRUNK COUNTEREXAMPLE (the minimal failing
+      # input), NOT on the issue title prefix. Keying on the prefix would let a
+      # single open issue swallow every OTHER counterexample (a different bug B whose
+      # title shares the prefix would be treated as a duplicate and stay silent until
+      # the first issue is closed). Hashing the shrunk example instead means two
+      # DIFFERENT counterexamples get two DIFFERENT issues, while a re-find of the
+      # SAME counterexample still dedupes onto the existing one. The infra-failure
+      # step (below) still keys on its own distinct title, so it can never poison
+      # this dedup either.
      - name: File counterexample issue
        # always() is REQUIRED: the fuzz step exits nonzero on a failing shard,
        # so a bare `if:` (implicitly success() && ...) would skip this step
@@ -146,25 +154,48 @@ jobs:
            echo "No fast-check counterexample signature — infra failure, handled by the next step."
            exit 0
          fi
-          TITLE="${TITLE_PREFIX} (seed=${FAIL_SEED})"
+          # Extract the SHRUNK counterexample block: the "Counterexample:" line(s)
+          # up to (but excluding) the "Shrunk N time(s)" / "Got error" line. This is
+          # the minimal failing INPUT and is STABLE across the different seeds/paths
+          # that reach the same bug — unlike the seed, path, or shrink count (which
+          # precede/follow this block and vary run-to-run) and unlike the whole
+          # output (which embeds those varying parts). Hashing THIS is what makes the
+          # dedup identity the bug itself rather than an incidental run detail.
+          CE_TEXT=$(awk '/Counterexample:/{c=1} /Shrunk [0-9]+ time|Got error/{c=0} c{print}' property-output.txt)
+          if [ -z "$CE_TEXT" ]; then
+            # No parseable shrunk block (unexpected — the signature check above
+            # already confirmed fast-check output). Fall back to the reproducing
+            # seed so we still emit a stable identity instead of silently deduping.
+            CE_TEXT="seed:${FAIL_SEED}"
+          fi
+          # Stable short id: first 12 hex chars of sha256 over the counterexample.
+          CE_HASH=$(printf '%s' "$CE_TEXT" | sha256sum | cut -c1-12)
+          # Machine-readable marker embedded in the issue body; the open-issue search
+          # below matches on it (and on the hash in the title) so identity travels
+          # with the issue regardless of any human title edits.
+          CE_MARKER="<!-- counterexample-hash: ${CE_HASH} -->"
+          export CE_HASH CE_MARKER
+          TITLE="${TITLE_PREFIX} [${CE_HASH}] (seed=${FAIL_SEED})"

-          # Best-effort dedup: skip if an open issue with the counterexample title
-          # prefix already exists. A failure of this check must NOT block creation.
+          # Dedup on the counterexample hash: skip only if an OPEN issue already
+          # carries this exact hash (in its title or its body marker). A different
+          # counterexample has a different hash and is NOT deduped. A failure of this
+          # check must NOT block creation.
          EXISTING=""
          if EXISTING=$(curl -sS \
              -H "Authorization: token ${GITHUB_TOKEN}" \
              "${GITHUB_API_URL}/repos/${GITHUB_REPOSITORY}/issues?state=open&limit=100"); then
            if printf '%s' "$EXISTING" \
-                | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{let a;try{a=JSON.parse(s)}catch{process.exit(1)}if(!Array.isArray(a))process.exit(1);const p=process.env.TITLE_PREFIX;process.exit(a.some(i=>typeof i.title==="string"&&i.title.startsWith(p))?0:1)})'; then
-              echo "An open '${TITLE_PREFIX}' issue already exists — skipping creation."
+                | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>{let a;try{a=JSON.parse(s)}catch{process.exit(1)}if(!Array.isArray(a))process.exit(1);const h=process.env.CE_HASH,m=process.env.CE_MARKER;process.exit(a.some(i=>(typeof i.title==="string"&&i.title.includes(h))||(typeof i.body==="string"&&i.body.includes(m)))?0:1)})'; then
+              echo "An open issue for counterexample ${CE_HASH} already exists — skipping creation."
              exit 0
            fi
          fi

          # Build the JSON body with the test output SAFELY escaped (never hand-
          # interpolate the counterexample into JSON).
-          BODY_TEXT=$(printf 'A nightly property fuzz SHARD failed with a fast-check counterexample.\n\n- failing shard seed: `%s`\n- NUM_RUNS (per shard): `%s`\n- run: %s\n\nReproduce locally:\n\n```\nPROPERTY_SEED=%s PROPERTY_NUM_RUNS=%s pnpm --filter @docmost/prosemirror-markdown exec vitest run test/generative/\n```\n\nfast-check shrinks the failure to a minimal counterexample. Commit it as a permanent fixture under `packages/prosemirror-markdown/test/fixtures/counterexamples/` + a case in `counterexamples.test.ts`, then FIX the converter (do not weaken a property). See `packages/prosemirror-markdown/README.md`.\n\nTail of the test output (contains the shrunk counterexample):\n\n```\n%s\n```\n' \
-            "$FAIL_SEED" "$NUM_RUNS" "$RUN_URL" "$FAIL_SEED" "$NUM_RUNS" "$(tail -n 120 property-output.txt)")
+          BODY_TEXT=$(printf 'A nightly property fuzz SHARD failed with a fast-check counterexample.\n\n- counterexample hash: `%s`\n- failing shard seed: `%s`\n- NUM_RUNS (per shard): `%s`\n- run: %s\n\nReproduce locally:\n\n```\nPROPERTY_SEED=%s PROPERTY_NUM_RUNS=%s pnpm --filter @docmost/prosemirror-markdown exec vitest run test/generative/\n```\n\nfast-check shrinks the failure to a minimal counterexample. Commit it as a permanent fixture under `packages/prosemirror-markdown/test/fixtures/counterexamples/` + a case in `counterexamples.test.ts`, then FIX the converter (do not weaken a property). See `packages/prosemirror-markdown/README.md`.\n\nTail of the test output (contains the shrunk counterexample):\n\n```\n%s\n```\n\n%s\n' \
+            "$CE_HASH" "$FAIL_SEED" "$NUM_RUNS" "$RUN_URL" "$FAIL_SEED" "$NUM_RUNS" "$(tail -n 120 property-output.txt)" "$CE_MARKER")

          jq -n --arg title "$TITLE" --arg body "$BODY_TEXT" \
            '{title: $title, body: $body}' > payload.json
@@ -471,6 +471,8 @@ Vite SPA. Code is organized by feature under `apps/client/src/features/*` (mirro
 - The version string shown in the UI comes from `APP_VERSION` (CI/Docker) or `git describe --tags --always` (local), resolved in `vite.config.ts` — not from `package.json`.
 - Server TS config is permissive (`noImplicitAny: false`, `strictNullChecks: false`, `no-explicit-any` lint disabled). Follow the existing relaxed style rather than tightening types broadly.
 - Dependency versions are heavily pinned via `pnpm.overrides` and `pnpm.patchedDependencies` (`scimmy`, `yjs`, `ai`) in the root `package.json`. Don't bump pinned/patched deps casually; the patches and overrides exist for compatibility/security reasons. The `ai@6.0.134` patch carries TWO independent server fixes, each with its own tripwire test: (1) it disables the SDK's O(n²) cumulative `partialOutput` accumulation when no output strategy is requested (server heap OOM on long agent runs, #184; tripwire: `apps/server/src/integrations/ai/ai-sdk-partial-output.patch.spec.ts`); (2) it fixes `writeToServerResponse`'s drain-hang — the loop awaited only `"drain"` under backpressure, so a mid-write client disconnect parked the pipe forever and leaked the reader/buffers until restart; it now races `"drain"` against `"close"`/`"error"`, cancels the reader on disconnect, and swallows the fire-and-forget read rejection (#486; tripwire: `apps/server/src/integrations/ai/ai-sdk-drain-hang.patch.spec.ts`). Both tripwires assert BOTH installed dist builds carry their patch marker. The patch MUST be re-created via `pnpm patch` when bumping `ai`.
+  - **Upstream tracking (report the analysis upstream, don't just carry it):** both `ai` fixes and the hocuspocus one are candidates for upstreaming so we can eventually drop the local patch — the analysis is already written up in each patch's `PATCH(...)` header comments. File (a) an upstream **issue** on `vercel/ai` for the O(n²) cumulative `partialOutput` accumulation (heap OOM), (b) an upstream **issue** on `vercel/ai` for the `writeToServerResponse` drain-hang, and (c) an upstream **PR** on `@hocuspocus/server` for the connect-vs-unload race (local marker `PATCH(gitmost #401)` in `patches/@hocuspocus__server@3.4.4.patch`). Do NOT edit the patch files to add links — the patch bytes feed `patch_hash` in `pnpm-lock.yaml` (`ai@6.0.134` → `e8c599b3…`), so any content change there desyncs the lockfile pin and breaks `pnpm install`; keep upstream references here instead.
+  - **`ai` version is split across the monorepo and MUST be aligned deliberately, NOT casually:** the server pins `ai@6.0.134` (patched, exact — the `patchedDependencies` key forces that version), while the client declares `ai@6.0.207` (unpatched — the server-side `writeToServerResponse`/`partialOutput` fixes are dead code in the browser, so the mismatch is currently benign but is real drift). Alignment is a **planned, install-gated step**, never a bare `package.json` edit: (1) choose the target version; (2) re-create ALL THREE patch hunks (partialOutput publish-each, the `DefaultStreamTextResult` lazy-`output` wiring, and the drain-hang race) against the target dist via `pnpm patch` — the line offsets shift between versions, so the current patch WILL fail to apply as-is; (3) run a full `pnpm install` so the lockfile + new `patch_hash` regenerate together; (4) confirm both tripwire specs still find their markers. `pnpm install` FAILS HARD on an unapplied patch — that failure is the guardrail, so treat the port as a deliberate plan rather than discovering it as a deploy-time surprise.
 - **The MCP tool inventory in `SERVER_INSTRUCTIONS` is GENERATED from the registry** (`packages/mcp/src/server-instructions.ts`: `buildToolInventory()` over `SHARED_TOOL_SPECS`) and spliced into the hand-written routing prose (`ROUTING_PROSE`). So adding/renaming/removing a **shared** spec in `packages/mcp/src/tool-specs.ts` auto-updates the `<tool_inventory>` — no manual `SERVER_INSTRUCTIONS` edit needed. Only an **inline** MCP-only tool (those registered via `server.registerTool(...)` in `index.ts`, not through the registry) needs a one-line entry in `INLINE_MCP_INVENTORY`. Enforced by `packages/mcp/test/unit/tool-inventory.test.mjs`, which fails when a registered tool is missing from the generated inventory (there is no `EXCEPTIONS` opt-out anymore — every tool must appear). Update `ROUTING_PROSE` when a tool's *intent guidance* (when-to-use) changes. `packages/mcp/build/` is gitignored and rebuilt in CI/Docker via `pnpm build` (same convention as `git-sync`/`prosemirror-markdown`) — never commit it; rebuild locally after editing to run the tests.

 ## CI / release
@@ -304,6 +304,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
  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)
+- **Vendor `ai` patch: upstream-tracking + version-alignment plan documented.**
+  The two local `ai@6.0.134` fixes (O(n²) `partialOutput` heap-OOM; the
+  `writeToServerResponse` drain-hang) and the hocuspocus connect-vs-unload race
+  now have explicit upstream-reporting and `ai`-version-alignment steps recorded
+  in `AGENTS.md` (client `ai@6.0.207` vs server `ai@6.0.134`-patched drift). The
+  patch bytes are unchanged — they feed the lockfile `patch_hash`, so the
+  alignment is called out as an install-gated plan rather than a bare version
+  bump. No runtime change.
 - **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
@@ -256,6 +256,9 @@
  "Invite link": "Ссылка для приглашения",
  "Copy": "Копировать",
  "Copy to space": "Копировать в пространство",
+  "Copy chat": "Копировать чат",
+  "Dock to sidebar": "Закрепить в боковой панели",
+  "Undock": "Открепить",
  "Copied": "Скопировано",
  "Failed to export chat": "Не удалось экспортировать чат",
  "Duplicate": "Дублировать",
@@ -285,6 +288,9 @@
  "Alt text": "Альтернативный текст",
  "Describe this for accessibility.": "Опишите это для специальных возможностей.",
  "Add a description": "Добавить описание",
+  "Caption": "Подпись",
+  "Add a caption": "Добавить подпись",
+  "Shown below the image.": "Отображается под изображением.",
  "Justify": "По ширине",
  "Merge cells": "Объединить ячейки",
  "Split cell": "Разделить ячейку",
@@ -388,22 +394,6 @@
  "Quote": "Цитата",
  "Image": "Изображение",
  "Audio": "Аудио",
-  "Transcribe": "Транскрибировать",
-  "Transcribing…": "Транскрибация…",
-  "No speech detected": "Речь не распознана",
-  "Transcription failed": "Не удалось распознать речь",
-  "Voice dictation is not configured": "Голосовой ввод не настроен",
-  "Start dictation": "Начать диктовку",
-  "Stop recording": "Остановить запись",
-  "Microphone access denied": "Доступ к микрофону запрещён",
-  "No microphone found": "Микрофон не найден",
-  "Microphone is unavailable or already in use": "Микрофон недоступен или уже используется",
-  "Could not start recording": "Не удалось начать запись",
-  "Audio recording is not available in this browser/context": "Запись аудио недоступна в этом браузере/контексте",
-  "Dictation": "Диктовка",
-  "Dictation becomes available once the page finishes connecting": "Диктовка станет доступна после подключения к документу",
-  "No connection to the collaboration server — dictation unavailable": "Нет связи с сервером совместного редактирования — диктовка недоступна",
-  "This page is read-only": "Страница открыта только для чтения",
  "Embed PDF": "Встроить PDF",
  "Upload and embed a PDF file.": "Загрузите и встроите PDF-файл.",
  "Embed as PDF": "Встроить как PDF",
@@ -419,9 +409,6 @@
  "Footnote {{number}}": "Сноска {{number}}",
  "Go to footnote": "Перейти к сноске",
  "Back to reference": "Вернуться к ссылке",
-  "Back to references": "Вернуться к ссылкам",
-  "Back to reference {{label}}": "Вернуться к ссылке {{label}}",
-  "Empty footnote": "Пустая сноска",
  "Math inline": "Строчная формула",
  "Insert inline math equation.": "Вставить математическое выражение в строку.",
  "Math block": "Блок формулы",
@@ -447,6 +434,9 @@
  "{{count}} command available_other": "Доступно {{count}} команд",
  "{{count}} result available_one": "Доступен 1 результат",
  "{{count}} result available_other": "Доступно {{count}} результатов",
+  "{{count}} result found_one": "Найден {{count}} результат",
+  "{{count}} result found_few": "Найдено {{count}} результата",
+  "{{count}} result found_other": "Найдено {{count}} результатов",
  "Equal columns": "Равные столбцы",
  "Left sidebar": "Левая боковая панель",
  "Right sidebar": "Правая боковая панель",
@@ -456,6 +446,7 @@
  "Names do not match": "Названия не совпадают",
  "Today, {{time}}": "Сегодня, {{time}}",
  "Yesterday, {{time}}": "Вчера, {{time}}",
+  "now": "сейчас",
  "Space created successfully": "Пространство успешно создано",
  "Space updated successfully": "Пространство успешно обновлено",
  "Space deleted successfully": "Пространство успешно удалено",
@@ -559,6 +550,7 @@
  "Add 2FA method": "Добавить метод 2FA",
  "Backup codes": "Резервные коды",
  "Disable": "Отключить",
+  "disabled": "отключено",
  "Invalid verification code": "Недействительный код подтверждения",
  "New backup codes have been generated": "Новые резервные коды сгенерированы",
  "Failed to regenerate backup codes": "Не удалось заново сгенерировать резервные коды",
@@ -702,62 +694,6 @@
  "AI search": "Поиск ИИ",
  "AI Answer": "Ответ ИИ",
  "Ask AI": "Спросить ИИ",
-  "AI agent": "AI-агент",
-  "Take a look at the current document": "Посмотри текущий документ",
-  "Start automatically": "Запускать автоматически",
-  "When on, picking this role sends a launch message and starts the chat. When off, the role is selected and you type the first message yourself.": "Когда включено, выбор этой роли отправляет стартовое сообщение и начинает чат. Когда выключено, роль выбирается, а первое сообщение вы вводите сами.",
-  "Launch message": "Стартовое сообщение",
-  "Sent automatically when this role is picked. Leave empty to use the default text. Ignored when “Start automatically” is off.": "Отправляется автоматически при выборе этой роли. Оставьте пустым, чтобы использовать текст по умолчанию. Игнорируется, когда «Запускать автоматически» выключено.",
-  "AI agent is typing…": "AI-агент печатает…",
-  "{{name}} is typing…": "{{name}} печатает…",
-  "Thinking…": "Думаю…",
-  "Thinking… · {{count}} tokens": "Думаю… · {{count}} токенов",
-  "Thinking… · {{count}} tokens_one": "Думаю… · {{count}} токен",
-  "Thinking… · {{count}} tokens_few": "Думаю… · {{count}} токена",
-  "Thinking… · {{count}} tokens_many": "Думаю… · {{count}} токенов",
-  "Thinking · {{count}} tokens": "Размышления · {{count}} токенов",
-  "Thinking · {{count}} tokens_one": "Размышления · {{count}} токен",
-  "Thinking · {{count}} tokens_few": "Размышления · {{count}} токена",
-  "Thinking · {{count}} tokens_many": "Размышления · {{count}} токенов",
-  "Agent role": "Роль агента",
-  "AI chat": "AI-чат",
-  "AI chat is disabled for this workspace.": "AI-чат отключён для этого рабочего пространства.",
-  "Ask a question about this documentation.": "Задайте вопрос об этой документации.",
-  "Ask a question…": "Задайте вопрос…",
-  "Ask the AI agent anything about your workspace.": "Спросите AI-агента о чём угодно по вашему рабочему пространству.",
-  "Ask the AI agent…": "Спросите AI-агента…",
-  "Copy chat": "Копировать чат",
-  "Dock to sidebar": "Закрепить в боковой панели",
-  "Undock": "Открепить",
-  "Created successfully": "Успешно создано",
-  "Context size / model limit": "Размер контекста / лимит модели",
-  "Context window (tokens)": "Окно контекста (токены)",
-  "Shown as used / total in the chat header. Leave empty to hide the limit.": "Показывается в шапке чата как использовано / всего. Пусто — лимит скрыт.",
-  "Delete this chat?": "Удалить этот чат?",
-  "Deleted successfully": "Успешно удалено",
-  "AI agent «{{role}}» on behalf of {{person}}": "AI-агент «{{role}}» от имени {{person}}",
-  "AI agent {{name}}": "AI-агент {{name}}",
-  "Failed to delete chat": "Не удалось удалить чат",
-  "Failed to rename chat": "Не удалось переименовать чат",
-  "Failed": "Ошибка",
-  "OK · {{n}}": "OK · {{n}}",
-  "Test": "Тест",
-  "No tools available": "Инструменты недоступны",
-  "Available tools": "Доступные инструменты",
-  "Minimize": "Свернуть",
-  "No chats yet.": "Чатов пока нет.",
-  "Send": "Отправить",
-  "Send when the agent finishes": "Отправить, когда агент закончит",
-  "Queue message": "Поставить в очередь",
-  "Remove queued message": "Убрать из очереди",
-  "Send now": "Отправить сейчас",
-  "Interrupt and send now": "Прервать и отправить сейчас",
-  "Something went wrong": "Что-то пошло не так",
-  "Stop": "Стоп",
-  "The AI agent could not respond. Please try again.": "AI-агент не смог ответить. Попробуйте ещё раз.",
-  "The AI provider is not configured. Ask an administrator to set it up.": "AI-провайдер не настроен. Попросите администратора настроить его.",
-  "Universal assistant": "Универсальный ассистент",
-  "You": "Вы",
  "AI is thinking...": "ИИ обрабатывает запрос...",
  "Thinking": "Думаю",
  "Ask a question...": "Задайте вопрос...",
@@ -784,8 +720,40 @@
  "Manage API keys for all users in the workspace. View the <anchor>API documentation</anchor> for usage details.": "Управляйте API-ключами для всех пользователей в рабочем пространстве. Смотрите <anchor>документацию по API</anchor> для получения информации об использовании.",
  "View the <anchor>API documentation</anchor> for usage details.": "Смотрите <anchor>документацию по API</anchor> для получения информации об использовании.",
  "View the <anchor>MCP documentation</anchor>.": "Смотрите <anchor>документацию по MCP</anchor>.",
-  "Instructions": "Инструкции",
+  "AI / Models": "ИИ / Модели",
+  "AI / External tools (MCP)": "ИИ / Внешние инструменты (MCP)",
+  "Add server": "Добавить сервер",
+  "Edit server": "Изменить сервер",
+  "Delete server": "Удалить сервер",
+  "Are you sure you want to delete this MCP server?": "Вы уверены, что хотите удалить этот MCP-сервер?",
+  "No external servers configured": "Внешние серверы не настроены",
+  "Server name": "Имя сервера",
+  "Transport": "Транспорт",
+  "URL": "URL",
+  "Authorization header": "Заголовок авторизации",
+  "Tool allowlist": "Список разрешённых инструментов",
+  "Optional. Leave empty to allow all tools the server exposes.": "Необязательно. Оставьте пустым, чтобы разрешить все инструменты, которые предоставляет сервер.",
  "Optional guidance for the agent on how and when to use this server's tools. Injected into the system prompt. The server's tools are namespaced as \"<server name>_*\".": "Необязательное указание агенту, как и когда использовать инструменты этого сервера. Добавляется в системный промпт. Инструменты сервера именуются с префиксом «<имя сервера>_*».",
+  "Test": "Тест",
+  "Available tools": "Доступные инструменты",
+  "No tools available": "Инструменты недоступны",
+  "Failed": "Ошибка",
+  "OK · {{n}}": "OK · {{n}}",
+  "Created successfully": "Успешно создано",
+  "Deleted successfully": "Успешно удалено",
+  "Clear": "Очистить",
+  "Provider": "Провайдер",
+  "•••• set": "•••• задан",
+  "Clear key": "Очистить ключ",
+  "Base URL": "Базовый URL",
+  "Chat model": "Модель чата",
+  "Embedding model": "Модель эмбеддингов",
+  "System message": "Системное сообщение",
+  "A built-in safety framework is always appended.": "Встроенный набор правил безопасности всегда добавляется автоматически.",
+  "Test connection": "Проверить соединение",
+  "Connection successful": "Соединение установлено",
+  "Connection failed": "Не удалось установить соединение",
+  "Only workspace admins can manage AI provider settings.": "Управлять настройками провайдера ИИ могут только администраторы рабочего пространства.",
  "Sources": "Источники",
  "AI Answers not available for attachments": "Ответы ИИ недоступны для вложений",
  "No answer available": "Ответ недоступен",
@@ -1013,6 +981,7 @@
  "Try again": "Попробовать снова",
  "Untitled chat": "Чат без названия",
  "No document": "Без документа",
+  "You": "Вы",
  "What can I help you with?": "Чем я могу вам помочь?",
  "Are you sure you want to revoke this {{credential}}": "Вы уверены, что хотите отозвать этот {{credential}}",
  "Automatically provision users and groups from your identity provider via SCIM.": "Автоматически предоставляйте доступ пользователям и группам из вашего провайдера удостоверений через SCIM.",
@@ -1041,6 +1010,9 @@
  "Page menu": "Меню страницы",
  "Expand": "Развернуть",
  "Collapse": "Свернуть",
+  "Expand all": "Развернуть все",
+  "Collapse all": "Свернуть все",
+  "Couldn't expand the tree: {{reason}}": "Не удалось развернуть дерево: {{reason}}",
  "Comment menu": "Меню комментария",
  "Group menu": "Меню группы",
  "Show hidden breadcrumbs": "Показать скрытые хлебные крошки",
@@ -1077,7 +1049,7 @@
  "Search pages and spaces...": "Поиск страниц и пространств...",
  "No results found": "Результаты не найдены",
  "You don't have permission to create pages here": "У вас нет прав на создание страниц здесь",
-  "Chat menu": "Меню чата",
+  "Chat menu for {{title}}": "Меню чата для {{title}}",
  "API key menu": "Меню API-ключа",
  "Jump to comment selection": "Перейти к выбору комментария",
  "Slash commands": "Команды со слешем",
@@ -1131,6 +1103,9 @@
  "Undo": "Отменить",
  "Redo": "Повторить",
  "Backlinks": "Обратные ссылки",
+  "Back to references": "Вернуться к ссылкам",
+  "Back to reference {{label}}": "Вернуться к ссылке {{label}}",
+  "Empty footnote": "Пустая сноска",
  "Last updated by": "Последний изменивший",
  "Last updated": "Последнее обновление",
  "Stats": "Статистика",
@@ -1164,6 +1139,7 @@
  "Page title": "Заголовок страницы",
  "Page content": "Содержимое страницы",
  "Member actions": "Действия с участником",
+  "Member actions for {{name}}": "Действия с участником {{name}}",
  "Toggle password visibility": "Переключить видимость пароля",
  "Send comment": "Отправить комментарий",
  "Token actions": "Действия с токеном",
@@ -1183,11 +1159,187 @@
  "Removed from favorites": "Удалено из избранного",
  "Added {{name}} to favorites": "{{name}} добавлено в избранное",
  "Removed {{name}} from favorites": "{{name}} удалено из избранного",
+  "Label added": "Метка добавлена",
+  "Label removed": "Метка удалена",
+  "Image updated": "Изображение обновлено",
+  "Unsupported image type": "Неподдерживаемый тип изображения",
+  "Member deactivated": "Участник деактивирован",
+  "Member activated": "Участник активирован",
+  "Name is required": "Укажите имя",
+  "Name must be 40 characters or fewer": "Имя должно содержать не более 40 символов",
+  "Group name must be at least 2 characters": "Название группы должно содержать не менее 2 символов",
+  "Group name must be 100 characters or fewer": "Название группы должно содержать не более 100 символов",
+  "Description must be 500 characters or fewer": "Описание должно содержать не более 500 символов",
+  "Invalid invitation link": "Недействительная ссылка-приглашение",
  "Page menu for {{name}}": "Меню страницы для {{name}}",
  "Create subpage of {{name}}": "Создать подстраницу для {{name}}",
+  "AI chat": "AI-чат",
+  "Ask a question about this documentation.": "Задайте вопрос об этой документации.",
+  "Ask a question…": "Задайте вопрос…",
+  "Thinking…": "Думаю…",
+  "Thinking… · {{count}} tokens": "Думаю… · {{count}} токенов",
+  "Thinking… · {{count}} tokens_one": "Думаю… · {{count}} токен",
+  "Thinking… · {{count}} tokens_few": "Думаю… · {{count}} токена",
+  "Thinking… · {{count}} tokens_many": "Думаю… · {{count}} токенов",
+  "Thinking… · {{count}} tokens_other": "Думаю… · {{count}} токенов",
+  "Thinking · {{count}} tokens": "Размышления · {{count}} токенов",
+  "Thinking · {{count}} tokens_one": "Размышления · {{count}} токен",
+  "Thinking · {{count}} tokens_few": "Размышления · {{count}} токена",
+  "Thinking · {{count}} tokens_many": "Размышления · {{count}} токенов",
+  "Thinking · {{count}} tokens_other": "Размышления · {{count}} токенов",
+  "The assistant is unavailable right now. Please try again.": "Ассистент сейчас недоступен. Попробуйте ещё раз.",
+  "Public share assistant": "Ассистент публичного доступа",
+  "Let anonymous visitors of public shares ask an AI assistant scoped to that share's pages. You pay for the tokens.": "Позвольте анонимным посетителям публичных ссылок обращаться к ИИ-ассистенту в рамках страниц этой публикации. Токены оплачиваете вы.",
+  "Public assistant model": "Модель публичного ассистента",
+  "Defaults to the chat model": "По умолчанию используется модель чата",
+  "Optional cheaper model id for the public assistant. Empty uses the chat model above.": "Необязательный более дешёвый идентификатор модели для публичного ассистента. Если пусто, используется модель чата выше.",
+  "Assistant identity": "Личность ассистента",
+  "Pick an agent role whose persona the public assistant adopts. The safety rules always still apply.": "Выберите роль агента, чью личность примет публичный ассистент. Правила безопасности всегда остаются в силе.",
+  "Built-in assistant persona": "Встроенная личность ассистента",
+  "Minimize": "Свернуть",
+  "Context size / model limit": "Размер контекста / лимит модели",
+  "Context window (tokens)": "Окно контекста (токены)",
+  "Shown as used / total in the chat header. Leave empty to hide the limit.": "Показывается в шапке чата как использовано / всего. Пусто — лимит скрыт.",
+  "AI agent": "AI-агент",
+  "Take a look at the current document": "Посмотри текущий документ",
+  "AI agent is typing…": "AI-агент печатает…",
+  "{{name}} is typing…": "{{name}} печатает…",
+  "Send": "Отправить",
+  "Send when the agent finishes": "Отправить, когда агент закончит",
+  "Queue message": "Поставить в очередь",
+  "Remove queued message": "Убрать из очереди",
+  "Send now": "Отправить сейчас",
+  "Interrupt and send now": "Прервать и отправить сейчас",
+  "Stop": "Стоп",
+  "Response stopped.": "Ответ остановлен.",
+  "Connection lost — the answer was interrupted.": "Соединение потеряно — ответ был прерван.",
+  "Response stopped (manually or the connection dropped).": "Ответ остановлен (вручную или из-за разрыва соединения).",
+  "Chat menu": "Меню чата",
+  "No chats yet.": "Чатов пока нет.",
+  "Delete this chat?": "Удалить этот чат?",
+  "Ask the AI agent…": "Спросите AI-агента…",
+  "Ask the AI agent anything about your workspace.": "Спросите AI-агента о чём угодно по вашему рабочему пространству.",
+  "Failed to rename chat": "Не удалось переименовать чат",
+  "Failed to delete chat": "Не удалось удалить чат",
+  "Something went wrong": "Что-то пошло не так",
+  "AI chat is disabled for this workspace.": "AI-чат отключён для этого рабочего пространства.",
+  "The AI provider is not configured. Ask an administrator to set it up.": "AI-провайдер не настроен. Попросите администратора настроить его.",
+  "The AI agent could not respond. Please try again.": "AI-агент не смог ответить. Попробуйте ещё раз.",
+  "Searched pages": "Поиск по страницам",
+  "Read page": "Прочитана страница",
+  "Created page": "Создана страница",
+  "Updated page": "Обновлена страница",
+  "Renamed page": "Переименована страница",
+  "Moved page": "Перемещена страница",
+  "Deleted page (to trash)": "Удалена страница (в корзину)",
+  "Commented": "Добавлен комментарий",
+  "Resolved comment": "Комментарий решён",
+  "Ran tool {{name}}": "Выполнен инструмент {{name}}",
+  "AI agent «{{role}}» on behalf of {{person}}": "AI-агент «{{role}}» от имени {{person}}",
+  "AI agent {{name}}": "AI-агент {{name}}",
+  "Endpoints": "Эндпоинты",
+  "where we fetch models": "откуда мы получаем модели",
+  "All endpoints are OpenAI-compatible. Point the Base URL at OpenAI, OpenRouter, a local Ollama, or any self-hosted server.": "Все эндпоинты совместимы с OpenAI. Укажите в базовом URL адрес OpenAI, OpenRouter, локального Ollama или любого self-hosted сервера.",
+  "Chat / LLM": "Чат / LLM",
+  "root": "корневой",
+  "Semantic search": "Семантический поиск",
+  "Voice / STT": "Голос / STT",
+  "Voice dictation": "Голосовой ввод",
+  "Streaming dictation": "Потоковый голосовой ввод",
+  "Transcribe as you speak, cutting on pauses": "Транскрибирование по мере речи, с разбивкой на паузах",
+  "Voice dictation is not available yet.": "Голосовой ввод пока недоступен.",
+  "Test endpoint": "Проверить эндпоинт",
+  "Save and test": "Сохранить и проверить",
+  "Save endpoints": "Сохранить эндпоинты",
+  "Configured and enabled": "Настроено и включено",
+  "Configured but disabled": "Настроено, но отключено",
+  "Enabled but not configured": "Включено, но не настроено",
+  "Not configured": "Не настроено",
+  "External tools": "Внешние инструменты",
+  "Gitmost as MCP client": "Gitmost как MCP-клиент",
+  "Servers the agent calls out to.": "Серверы, к которым обращается агент.",
+  "MCP server": "MCP-сервер",
+  "expose the workspace": "открыть доступ к рабочему пространству",
+  "Enable MCP server": "Включить MCP-сервер",
+  "Exposes the workspace as an MCP server at /mcp — this provides a capability, it doesn't consume a model.": "Открывает рабочее пространство как MCP-сервер по адресу /mcp — это предоставляет возможность, а не потребляет модель.",
+  "Resolves to {{url}}": "Разрешается в {{url}}",
+  "Model": "Модель",
+  "Done": "Готово",
+  "shared prompt · safety framework appended automatically": "общий промпт · правила безопасности добавляются автоматически",
+  "/v1/chat/completions · root endpoint — Embeddings and Voice inherit its URL and key": "/v1/chat/completions · корневой эндпоинт — Эмбеддинги и Голос наследуют его URL и ключ",
+  "/v1/embeddings · embeds pages so semantic search can find them": "/v1/embeddings · создаёт эмбеддинги страниц, чтобы их находил семантический поиск",
+  "/v1/audio/transcriptions · works with local whisper (speaches / faster-whisper-server)": "/v1/audio/transcriptions · работает с локальным whisper (speaches / faster-whisper-server)",
+  "Vector search · requires pgvector": "Векторный поиск · требуется pgvector",
+  "Embedding API key": "API-ключ для эмбеддингов",
+  "Embeddings": "Эмбеддинги",
+  "Leave empty to use the chat API key": "Оставьте пустым, чтобы использовать API-ключ чата",
+  "Leave empty to use the chat base URL": "Оставьте пустым, чтобы использовать базовый URL чата",
+  "Reindex now": "Переиндексировать сейчас",
+  "Start dictation": "Начать диктовку",
+  "Stop recording": "Остановить запись",
+  "Transcribing…": "Транскрибация…",
+  "Microphone access denied": "Доступ к микрофону запрещён",
+  "No microphone found": "Микрофон не найден",
+  "Could not start recording": "Не удалось начать запись",
+  "Transcription failed": "Не удалось распознать речь",
+  "Transcribe": "Транскрибировать",
+  "No speech detected": "Речь не распознана",
+  "Voice dictation is not configured": "Голосовой ввод не настроен",
+  "Microphone is unavailable or already in use": "Микрофон недоступен или уже используется",
+  "Audio recording is not available in this browser/context": "Запись аудио недоступна в этом браузере/контексте",
+  "Dictation": "Диктовка",
+  "Dictation becomes available once the page finishes connecting": "Диктовка станет доступна после подключения к документу",
+  "No connection to the collaboration server — dictation unavailable": "Нет связи с сервером совместного редактирования — диктовка недоступна",
+  "This page is read-only": "Страница открыта только для чтения",
+  "Request format": "Формат запроса",
+  "How transcription requests are sent to the endpoint": "Как запросы на транскрибирование отправляются на эндпоинт",
+  "OpenAI-compatible (multipart/form-data)": "Совместимо с OpenAI (multipart/form-data)",
+  "OpenRouter (JSON, base64 audio)": "OpenRouter (JSON, аудио в base64)",
  "Dictation language": "Язык диктовки",
  "Auto-detect": "Автоопределение",
  "Spoken language hint sent to the transcription model. Auto-detect lets the model decide.": "Подсказка языка речи для модели транскрипции. «Автоопределение» оставляет выбор за моделью.",
+  "Agent role": "Роль агента",
+  "Universal assistant": "Универсальный ассистент",
+  "Add role": "Добавить роль",
+  "Edit role": "Изменить роль",
+  "Role name": "Название роли",
+  "e.g. Proofreader": "напр. Корректор",
+  "Optional. Shown as the chat badge.": "Необязательно. Отображается как значок чата.",
+  "Optional. A short note about what this role does.": "Необязательно. Краткое описание того, что делает эта роль.",
+  "Instructions": "Инструкции",
+  "The built-in safety framework is always added automatically.": "Встроенный набор правил безопасности всегда добавляется автоматически.",
+  "Model provider override": "Переопределение провайдера модели",
+  "Optional. Defaults to the workspace provider.": "Необязательно. По умолчанию используется провайдер рабочего пространства.",
+  "Model override": "Переопределение модели",
+  "Optional. Defaults to the workspace model.": "Необязательно. По умолчанию используется модель рабочего пространства.",
+  "e.g. gpt-4o-mini": "напр. gpt-4o-mini",
+  "If you choose a different provider, it must already be configured in AI settings.": "Если вы выбираете другого провайдера, он уже должен быть настроен в настройках ИИ.",
+  "Start automatically": "Запускать автоматически",
+  "When on, picking this role sends a launch message and starts the chat. When off, the role is selected and you type the first message yourself.": "Когда включено, выбор этой роли отправляет стартовое сообщение и начинает чат. Когда выключено, роль выбирается, а первое сообщение вы вводите сами.",
+  "Launch message": "Стартовое сообщение",
+  "Sent automatically when this role is picked. Leave empty to use the default text. Ignored when “Start automatically” is off.": "Отправляется автоматически при выборе этой роли. Оставьте пустым, чтобы использовать текст по умолчанию. Игнорируется, когда «Запускать автоматически» выключено.",
+  "Agent roles": "Роли агента",
+  "Reusable presets that shape the agent's behavior (and optionally its model). Picked when starting a new chat.": "Многоразовые пресеты, определяющие поведение агента (и, при желании, его модель). Выбираются при запуске нового чата.",
+  "No roles configured": "Роли не настроены",
+  "Delete role": "Удалить роль",
+  "Are you sure you want to delete this role?": "Вы уверены, что хотите удалить эту роль?",
+  "HTML embed": "HTML-вставка",
+  "Edit HTML embed": "Изменить HTML-вставку",
+  "HTML embed is disabled in this workspace": "HTML-вставки отключены в этом рабочем пространстве",
+  "Click to add HTML / CSS / JS": "Нажмите, чтобы добавить HTML / CSS / JS",
+  "This HTML/CSS/JS runs in a sandboxed frame and cannot access the viewer's session, cookies, or API.": "Этот HTML/CSS/JS выполняется в изолированном фрейме и не имеет доступа к сессии, cookie или API просматривающего.",
+  "<script>...</script>": "<script>...</script>",
+  "Height (px, blank = auto)": "Высота (px, пусто = авто)",
+  "advanced": "дополнительно",
+  "Enable HTML embed": "Включить HTML-вставки",
+  "Allow members to insert raw HTML/CSS/JavaScript blocks. The block renders in a sandboxed frame and cannot access the viewer's session, cookies, or API. Off by default.": "Разрешить участникам вставлять блоки с необработанным HTML/CSS/JavaScript. Блок отображается в изолированном фрейме и не имеет доступа к сессии, cookie или API просматривающего. По умолчанию выключено.",
+  "When enabled, any member can insert an HTML embed block. The toggle just enables or disables the block type workspace-wide.": "Когда включено, любой участник может вставить блок HTML-вставки. Переключатель просто включает или отключает этот тип блока во всём рабочем пространстве.",
+  "Embeds run inside a sandboxed iframe with a separate origin, so they cannot read or modify the page they are embedded in.": "Вставки выполняются в изолированном iframe с отдельным источником, поэтому они не могут читать или изменять страницу, в которую встроены.",
+  "Turning this off hides existing embeds (they render as a disabled placeholder) and stops serving them on public share pages.": "Отключение этой опции скрывает существующие вставки (они отображаются как отключённая заглушка) и прекращает их показ на публичных страницах.",
+  "Analytics / tracker": "Аналитика / трекер",
+  "Injected verbatim into the <head> of PUBLIC SHARE pages only (same-origin). For analytics snippets (Google Analytics, Yandex.Metrika, etc.). Admin only.": "Вставляется дословно в <head> только ПУБЛИЧНЫХ страниц (тот же источник). Для сниппетов аналитики (Google Analytics, Яндекс.Метрика и т. п.). Только для администраторов.",
+  "Go to login page": "Перейти на страницу входа",
+  "Move to space": "Переместить в пространство",
  "Float left (wrap text)": "Обтекание слева",
  "Float right (wrap text)": "Обтекание справа",
  "Inline (side by side)": "В ряд",
@@ -1199,6 +1351,7 @@
  "Showing {{count}} subpages_one": "Показано {{count}} подстраница",
  "Showing {{count}} subpages_few": "Показано {{count}} подстраницы",
  "Showing {{count}} subpages_many": "Показано {{count}} подстраниц",
+  "Showing {{count}} subpages_other": "Показано {{count}} подстраниц",
  "Protocol": "Протокол",
  "How chat requests are sent and how reasoning is surfaced": "Как отправляются запросы чата и как показывается reasoning",
  "OpenAI-compatible (surfaces reasoning)": "OpenAI-совместимый (показывает reasoning)",
@@ -1268,7 +1421,6 @@
  "Retry": "Повторить",
  "The catalog is empty": "Каталог пуст",
  "No role bundles are published for this language yet. Try switching the content language.": "Для этого языка ещё не опубликовано ни одного набора ролей. Попробуйте сменить язык контента.",
-  "No roles configured": "Роли не настроены",
  "Already up to date": "Уже актуальна",
  "Updated to the latest version": "Обновлено до последней версии",
  "This role is no longer in the catalog": "Эта роль больше не представлена в каталоге",
@@ -1,5 +1,5 @@
 import { describe, it, expect } from "vitest";
-import { isChunkLoadError } from "./chunk-load-error-boundary";
+import { isChunkLoadError, shouldAutoReload } from "./chunk-load-error-boundary";

 // The detector decides whether a caught render error is a stale-deploy chunk-404
 // (→ auto-reload to fetch the new manifest) vs a genuine app error (→ generic
@@ -35,3 +35,31 @@ describe("isChunkLoadError", () => {
    expect(isChunkLoadError(err)).toBe(false);
  });
 });
+
+// The window gate replaces the old one-shot flag: it must permit recovery across
+// several deploys in one tab (each > window apart) while still stopping an infinite
+// reload loop when a lazy chunk is permanently broken (a second failure < window).
+describe("shouldAutoReload", () => {
+  const WINDOW = 5 * 60 * 1000;
+  const NOW = 1_000_000_000_000;
+
+  it("allows a reload when we have never auto-reloaded", () => {
+    expect(shouldAutoReload(NOW, null, WINDOW)).toBe(true);
+  });
+
+  it("allows a reload when the last one was 6 minutes ago (outside the window)", () => {
+    expect(shouldAutoReload(NOW, NOW - 6 * 60 * 1000, WINDOW)).toBe(true);
+  });
+
+  it("blocks a reload when the last one was 1 minute ago (inside the window)", () => {
+    expect(shouldAutoReload(NOW, NOW - 1 * 60 * 1000, WINDOW)).toBe(false);
+  });
+
+  it("blocks a reload exactly at the window boundary (not strictly older)", () => {
+    expect(shouldAutoReload(NOW, NOW - WINDOW, WINDOW)).toBe(false);
+  });
+
+  it("allows a reload when the stored timestamp is unparseable (NaN)", () => {
+    expect(shouldAutoReload(NOW, NaN, WINDOW)).toBe(true);
+  });
+});
@@ -2,7 +2,25 @@ import { ReactNode } from "react";
 import { ErrorBoundary } from "react-error-boundary";
 import { Button, Center, Stack, Text } from "@mantine/core";

-const RELOAD_FLAG = "chunk-reload-attempted";
+// sessionStorage key holding the epoch-ms timestamp of the last automatic reload.
+const RELOAD_AT_KEY = "chunk-reload-at";
+// Allow at most one automatic reload per this window. A stale-deploy 404 is cured
+// by a single reload, so anything inside the window is treated as a reload loop
+// (permanently-broken chunk) and falls through to the manual UI. A window (rather
+// than a one-shot flag) lets a SECOND deploy in the same tab's lifetime recover too.
+const RELOAD_WINDOW_MS = 5 * 60 * 1000;
+
+// Pure window decision, unit-tested in isolation: auto-reload only if we have never
+// auto-reloaded (lastReloadAt null/NaN) or the last one was strictly older than the
+// window. Anything inside the window is suppressed to break an infinite reload loop.
+export function shouldAutoReload(
+  now: number,
+  lastReloadAt: number | null,
+  windowMs: number,
+): boolean {
+  if (lastReloadAt === null || Number.isNaN(lastReloadAt)) return true;
+  return now - lastReloadAt > windowMs;
+}

 // Heuristic detection of a failed dynamic import. Since the code-splitting work,
 // every route (plus Aside / AiChatWindow) is React.lazy: when a new deploy
@@ -24,12 +42,16 @@ export function isChunkLoadError(error: unknown): boolean {
 function handleError(error: unknown) {
  if (!isChunkLoadError(error)) return;
  // A stale-chunk 404 is cured by a full reload that re-fetches index.html and
-  // the new chunk manifest. Auto-reload once, guarding against a reload loop
-  // (e.g. a genuinely missing chunk) with a one-shot sessionStorage flag. If the
-  // flag is already set we fall through to the manual recovery UI below.
+  // the new chunk manifest. Auto-reload at most once per RELOAD_WINDOW_MS: this
+  // recovers across multiple deploys in a single tab's lifetime, yet a
+  // permanently-broken lazy chunk (which would loop) is stopped after the first
+  // reload and falls through to the manual recovery UI below.
  try {
-    if (sessionStorage.getItem(RELOAD_FLAG)) return;
-    sessionStorage.setItem(RELOAD_FLAG, "1");
+    const raw = sessionStorage.getItem(RELOAD_AT_KEY);
+    const lastReloadAt = raw === null ? null : Number.parseInt(raw, 10);
+    const now = Date.now();
+    if (!shouldAutoReload(now, lastReloadAt, RELOAD_WINDOW_MS)) return;
+    sessionStorage.setItem(RELOAD_AT_KEY, String(now));
  } catch {
    // sessionStorage unavailable (private mode / disabled): skip the automatic
    // reload rather than risk an unguarded loop; the fallback UI still recovers.
@@ -89,6 +89,23 @@ describe("describeChatError", () => {
    expect(view.title).not.toBe("AI provider not configured");
  });

+  it("classifies a token-degeneration abort under the SAME 'Response stopped.' marker the live view shows (#495)", () => {
+    // The exact reason the server persists in metadata.error on a degeneration
+    // abort (ai-chat.service OUTPUT_DEGENERATION_ERROR). Live, this event shows
+    // the neutral "Response stopped." notice; the persisted banner MUST match it
+    // so live and refetch never disagree.
+    const view = describeChatError(
+      "Output degeneration detected (repeated token loop)",
+      t,
+    );
+    expect(view.title).toBe("Response stopped.");
+    expect(view.detail).toBe(
+      "The answer was stopped automatically because the model fell into a repeated output loop.",
+    );
+    // Regression guard: it must NOT fall through to the generic heading.
+    expect(view.title).not.toBe("Something went wrong");
+  });
+
  it("classifies a dropped connection (ECONNRESET) as a lost-connection error", () => {
    expect(
      describeChatError("Cannot connect to API: read ECONNRESET", t).title,
@@ -77,6 +77,22 @@ export function describeChatError(
    };
  }

+  // Our own token-degeneration abort (#444): the server aborts a runaway
+  // repetition loop and persists this exact reason in metadata.error. LIVE, the
+  // same abort surfaces as the neutral "Response stopped." notice (the client
+  // cannot tell it from a manual Stop mid-stream), so the persisted banner must
+  // read the SAME "Response stopped." marker — otherwise the live view and a
+  // later refetch show two different texts for one event. The detail explains the
+  // loop-guard cause without contradicting the shared heading.
+  if (/output degeneration detected|repeated token loop/i.test(msg)) {
+    return {
+      title: t("Response stopped."),
+      detail: t(
+        "The answer was stopped automatically because the model fell into a repeated output loop.",
+      ),
+    };
+  }
+
  if (/"statusCode"\s*:\s*403\b/.test(msg)) {
    return {
      title: t("AI chat is disabled"),
@@ -0,0 +1,65 @@
+import { describe, it, expect } from "vitest";
+import * as Y from "yjs";
+import { yHistoryAvailability } from "./use-toolbar-state.ts";
+
+// Undo/redo availability is derived from the Yjs UndoManager's PRIVATE
+// `undoStack` / `redoStack` fields (see use-toolbar-state.ts for why we read the
+// stack lengths directly instead of the expensive `editor.can().undo()` dry-run).
+// These tests lock in the behavior AND pin the library shape so a yjs / y-undo
+// upgrade that renames/restructures those internals fails loudly here rather than
+// silently enabling/disabling the toolbar buttons in production.
+describe("yHistoryAvailability", () => {
+  it("reports availability from the stack lengths", () => {
+    expect(yHistoryAvailability({ undoStack: [], redoStack: [] })).toEqual({
+      canUndo: false,
+      canRedo: false,
+    });
+    expect(
+      yHistoryAvailability({ undoStack: [{}], redoStack: [] }),
+    ).toEqual({ canUndo: true, canRedo: false });
+    expect(
+      yHistoryAvailability({ undoStack: [{}], redoStack: [{}, {}] }),
+    ).toEqual({ canUndo: true, canRedo: true });
+  });
+
+  it("returns null when the private stack shape is unrecognized (upgrade guard)", () => {
+    // Simulates a yjs / y-undo upgrade that renames or restructures the private
+    // fields: the caller then falls back to the safe prosemirror-history default
+    // instead of throwing on `.length` of undefined or reading garbage.
+    expect(yHistoryAvailability(undefined)).toBeNull();
+    expect(yHistoryAvailability(null)).toBeNull();
+    expect(yHistoryAvailability({})).toBeNull();
+    expect(yHistoryAvailability({ undoStack: 5, redoStack: 5 })).toBeNull();
+    // Only one stack present (partial rename) is still not trusted.
+    expect(yHistoryAvailability({ undoStack: [] })).toBeNull();
+  });
+
+  it("pin-test: a real yjs UndoManager still exposes undoStack/redoStack arrays", () => {
+    const doc = new Y.Doc();
+    const text = doc.getText("prosemirror");
+    const undoManager = new Y.UndoManager(text);
+
+    // Fresh manager: both stacks empty -> nothing to undo/redo.
+    expect(yHistoryAvailability(undoManager)).toEqual({
+      canUndo: false,
+      canRedo: false,
+    });
+
+    // A tracked edit must push onto the private undoStack. If a future yjs
+    // renames these fields, yHistoryAvailability(undoManager) returns null and
+    // the expectation below fails loudly.
+    text.insert(0, "hello");
+    undoManager.stopCapturing();
+    expect(yHistoryAvailability(undoManager)).toEqual({
+      canUndo: true,
+      canRedo: false,
+    });
+
+    // Undoing moves the item to the redoStack -> redo becomes available.
+    undoManager.undo();
+    expect(yHistoryAvailability(undoManager)).toEqual({
+      canUndo: false,
+      canRedo: true,
+    });
+  });
+});
@@ -35,6 +35,30 @@ export interface ToolbarState {
 // When neither history backend is installed (the pre-sync static editor —
 // mainExtensions only, undoRedo disabled), both fall through to 0 -> false,
 // matching the previous `safeCan` behavior.
+// Reads the Yjs UndoManager's undo/redo availability from its stack lengths.
+//
+// `undoStack` / `redoStack` are PRIVATE y-undo / yjs internals, so we touch them
+// defensively: a yjs or y-undo upgrade that renames or restructures these fields
+// must not silently mis-drive the toolbar buttons (nor throw on `.length` of
+// `undefined`). We only trust them when they are actually arrays; otherwise this
+// returns null and the caller falls back to a safe default. The pin-test in
+// use-toolbar-state.test.ts asserts the current library shape, so an upgrade that
+// breaks this contract fails loudly there instead of failing silently in the UI.
+export function yHistoryAvailability(
+  undoManager: unknown,
+): { canUndo: boolean; canRedo: boolean } | null {
+  if (!undoManager || typeof undoManager !== "object") return null;
+  const { undoStack, redoStack } = undoManager as {
+    undoStack?: unknown;
+    redoStack?: unknown;
+  };
+  if (!Array.isArray(undoStack) || !Array.isArray(redoStack)) return null;
+  return {
+    canUndo: undoStack.length > 0,
+    canRedo: redoStack.length > 0,
+  };
+}
+
 function historyAvailability(editor: Editor): {
  canUndo: boolean;
  canRedo: boolean;
@@ -43,16 +67,14 @@ function historyAvailability(editor: Editor): {

  // Collaboration history (Yjs) takes precedence when present.
  const yState = yUndoPluginKey.getState(state) as
-    | { undoManager?: { undoStack: unknown[]; redoStack: unknown[] } }
+    | { undoManager?: unknown }
    | undefined;
-  if (yState?.undoManager) {
-    return {
-      canUndo: yState.undoManager.undoStack.length > 0,
-      canRedo: yState.undoManager.redoStack.length > 0,
-    };
-  }
+  const yAvail = yHistoryAvailability(yState?.undoManager);
+  if (yAvail) return yAvail;

  // Plain prosemirror-history (returns 0 when the history plugin is absent).
+  // This is also the safe default when a Yjs UndoManager is present but its
+  // private stack shape is no longer recognized (yHistoryAvailability -> null).
  return {
    canUndo: undoDepth(state) > 0,
    canRedo: redoDepth(state) > 0,
@@ -13,8 +13,7 @@ let currentAlias: IShareAlias | null = null;
 let availabilityResult: {
  valid: boolean;
  available: boolean;
-  currentPageId: string | null;
-} = { valid: true, available: true, currentPageId: null };
+} = { valid: true, available: true };

 vi.mock("@/features/share/queries/share-query.ts", () => ({
  useShareAliasForPageQuery: () => ({ data: currentAlias }),
@@ -56,7 +55,7 @@ describe("ShareAliasSection — taken-name handling is never a dead end", () =>
  beforeEach(() => {
    setMutateAsync.mockReset();
    currentAlias = null;
-    availabilityResult = { valid: true, available: true, currentPageId: null };
+    availabilityResult = { valid: true, available: true };
  });

  it("shows a 'will move it here' HINT (not a terminal error) when the name belongs to another page, and keeps Save enabled", async () => {
@@ -65,7 +64,6 @@ describe("ShareAliasSection — taken-name handling is never a dead end", () =>
    availabilityResult = {
      valid: true,
      available: false,
-      currentPageId: "page-X",
    };

    renderSection("page-Y");
@@ -97,7 +95,6 @@ describe("ShareAliasSection — taken-name handling is never a dead end", () =>
    availabilityResult = {
      valid: true,
      available: false,
-      currentPageId: "page-X",
    };
    // The server rejects the un-confirmed save asking the client to confirm.
    setMutateAsync.mockRejectedValueOnce({
@@ -106,7 +103,6 @@ describe("ShareAliasSection — taken-name handling is never a dead end", () =>
        status: 409,
        data: {
          code: "ALIAS_REASSIGN_REQUIRED",
-          currentPageId: "page-X",
          currentPageTitle: "Alias Test Page X",
        },
      },
@@ -48,7 +48,6 @@ export default function ShareAliasSection({
  const [availability, setAvailability] = useState<{
    valid: boolean;
    available: boolean;
-    currentPageId: string | null;
  } | null>(null);
  const [reassign, setReassign] = useState<{
    alias: string;
@@ -76,7 +75,6 @@ export default function ShareAliasSection({
        setAvailability({
          valid: res.valid,
          available: res.available,
-          currentPageId: res.currentPageId,
        });
      } catch {
        setAvailability(null);
@@ -108,7 +108,6 @@ export interface IShareAliasAvailability {
  alias: string;
  valid: boolean;
  available: boolean;
-  currentPageId: string | null;
 }

 export interface ISharedPageTree {
@@ -6,6 +6,8 @@ import {
  nextReindexPollInterval,
  isReindexComplete,
  isReindexButtonLoading,
+  reindexRunKey,
+  isNewReindexRun,
 } from './ai-provider-settings';

 describe('resolveCardStatus', () => {
@@ -221,6 +223,128 @@ describe('isReindexComplete', () => {
  });
 });

+describe('reindexRunKey', () => {
+  it('is null when the status carries no run identity', () => {
+    expect(reindexRunKey(undefined)).toBeNull();
+    expect(
+      reindexRunKey({ reindexing: false, indexedPages: 5, totalPages: 5 }),
+    ).toBeNull();
+  });
+
+  it('is null for a legacy/degraded record with an empty runId', () => {
+    // The server sends runId='' for a record written before the field existed;
+    // the client must treat that as "no identity" (fall back to prior behaviour).
+    expect(
+      reindexRunKey({
+        reindexing: true,
+        indexedPages: 0,
+        totalPages: 10,
+        runId: '',
+        reindexStartedAt: 1000,
+      }),
+    ).toBeNull();
+  });
+
+  it('folds runId and startedAt into one stable key', () => {
+    expect(
+      reindexRunKey({
+        reindexing: true,
+        indexedPages: 0,
+        totalPages: 10,
+        runId: 'run-a',
+        reindexStartedAt: 1000,
+      }),
+    ).toBe('run-a:1000');
+  });
+
+  it('changes when the runId changes for the same startedAt', () => {
+    const a = reindexRunKey({
+      reindexing: true,
+      indexedPages: 0,
+      totalPages: 10,
+      runId: 'run-a',
+      reindexStartedAt: 1000,
+    });
+    const b = reindexRunKey({
+      reindexing: true,
+      indexedPages: 0,
+      totalPages: 10,
+      runId: 'run-b',
+      reindexStartedAt: 1000,
+    });
+    expect(a).not.toBe(b);
+  });
+
+  it('changes when the same runId restarts at a new startedAt', () => {
+    const a = reindexRunKey({
+      reindexing: true,
+      indexedPages: 0,
+      totalPages: 10,
+      runId: 'run-a',
+      reindexStartedAt: 1000,
+    });
+    const b = reindexRunKey({
+      reindexing: true,
+      indexedPages: 0,
+      totalPages: 10,
+      runId: 'run-a',
+      reindexStartedAt: 2000,
+    });
+    expect(a).not.toBe(b);
+  });
+});
+
+describe('isNewReindexRun (poll keying on runId)', () => {
+  // Derive the status shape from the helper itself so the test needs no export
+  // of the component-internal ReindexStatus type.
+  type ReindexStatusLike = NonNullable<Parameters<typeof reindexRunKey>[0]>;
+  const run = (runId: string, startedAt: number): ReindexStatusLike => ({
+    reindexing: true,
+    indexedPages: 0,
+    totalPages: 10,
+    runId,
+    reindexStartedAt: startedAt,
+  });
+
+  it('first identity after none latched is a NEW run', () => {
+    expect(isNewReindexRun(null, run('run-a', 1000))).toBe(true);
+  });
+
+  it('the SAME identity is not a new run (same run being watched)', () => {
+    const key = reindexRunKey(run('run-a', 1000));
+    expect(isNewReindexRun(key, run('run-a', 1000))).toBe(false);
+  });
+
+  it('a DIFFERENT runId is a new run (reset per-run poll state)', () => {
+    const key = reindexRunKey(run('run-a', 1000));
+    expect(isNewReindexRun(key, run('run-b', 1000))).toBe(true);
+  });
+
+  it('an identity-less poll (no runId / cleared record) is never a new run', () => {
+    const key = reindexRunKey(run('run-a', 1000));
+    expect(
+      isNewReindexRun(key, {
+        reindexing: false,
+        indexedPages: 10,
+        totalPages: 10,
+      }),
+    ).toBe(false);
+  });
+
+  it('a legacy empty-runId poll does not spuriously reset a latched run', () => {
+    const key = reindexRunKey(run('run-a', 1000));
+    expect(
+      isNewReindexRun(key, {
+        reindexing: true,
+        indexedPages: 3,
+        totalPages: 10,
+        runId: '',
+        reindexStartedAt: 1000,
+      }),
+    ).toBe(false);
+  });
+});
+
 describe('isReindexButtonLoading', () => {
  it('loads while the POST mutation is pending', () => {
    expect(
@@ -173,9 +173,43 @@ export function resolveKeyField(
 // Subset of the status payload that drives the reindex poll decisions.
 type ReindexStatus = Pick<
  IAiSettings,
-  "reindexing" | "indexedPages" | "totalPages"
+  "reindexing" | "indexedPages" | "totalPages" | "runId" | "reindexStartedAt"
 >;

+/**
+ * A stable per-RUN key for the reindex poll: `runId:startedAt`, or `null` when
+ * the status carries no run identity (no active run, or a legacy/degraded
+ * server record with an empty runId). Two polls of the SAME run share a key; a
+ * new run mints a fresh runId and so a different key.
+ *
+ * This is the single place the client turns the server's run identity into the
+ * value it keys on — it removes the "is this the same run I've been watching or
+ * a brand-new one?" ambiguity that made a class of reindex-status bugs (a stale
+ * pre-reindex snapshot vs a fresh run) get fixed twice (#262). `startedAt` is
+ * folded in so a run that somehow reuses a runId but restarted is still new.
+ */
+export function reindexRunKey(status: ReindexStatus | undefined): string | null {
+  const runId = status?.runId;
+  if (!runId) return null;
+  return `${runId}:${status?.reindexStartedAt ?? ""}`;
+}
+
+/**
+ * Decide whether the latest poll represents a NEW reindex run relative to the
+ * run key the client last latched (`prevKey`, `null` if none yet). True only
+ * when the status carries an identity AND it differs from the latched one — the
+ * signal to reset any per-run poll state (the "seen active" latch / progress the
+ * UI held). The same identity (or no identity) is NOT a new run, so an unchanged
+ * or identity-less poll never resets mid-run.
+ */
+export function isNewReindexRun(
+  prevKey: string | null,
+  status: ReindexStatus | undefined,
+): boolean {
+  const key = reindexRunKey(status);
+  return key !== null && key !== prevKey;
+}
+
 /**
 * Decide the TanStack Query `refetchInterval` while a reindex may be running.
 * Returns the poll interval (ms) to keep polling, or `false` to stop.
@@ -320,6 +354,13 @@ export default function AiProviderSettings() {
  // counter at 0 until a manual reload. A ref (not state) because it must not
  // trigger a render and is only ever read where `reindexing` is already false.
  const reindexSeenActiveRef = useRef(false);
+  // The run identity (runId:startedAt) the current poll window is keyed on. When
+  // a poll reports a DIFFERENT runId the server has started a NEW run, so we
+  // re-latch to it and reset `reindexSeenActiveRef` — a fresh run must never
+  // inherit the previous run's "seen active"/completion state (which would stop
+  // polling immediately or read the old run's counters as this run's). null =
+  // no run keyed yet (steady state, or a legacy record without a runId).
+  const reindexRunKeyRef = useRef<string | null>(null);

  // Only admins may read the (masked) AI settings; the server enforces this too.
  const { data: settings, isLoading } = useAiSettingsQuery(isAdmin, (query) =>
@@ -336,6 +377,14 @@ export default function AiProviderSettings() {
  // unmount because the deadline state goes away with the component.
  useEffect(() => {
    if (reindexDeadline === null) return;
+    // Key the poll on the run identity: if this poll carries a runId different
+    // from the one we latched, the server started a NEW run, so adopt it and
+    // drop the per-run "seen active" latch (a fresh run must not inherit the
+    // previous run's completion state). Same runId => same run, leave it alone.
+    if (isNewReindexRun(reindexRunKeyRef.current, settings)) {
+      reindexRunKeyRef.current = reindexRunKey(settings);
+      reindexSeenActiveRef.current = false;
+    }
    // Latch "we have seen the active run" the moment a poll reports it, so the
    // completion check below (and the refetchInterval's) only fires once the run
    // has genuinely started — never on the stale pre-reindex snapshot.
@@ -1220,6 +1269,10 @@ export default function AiProviderSettings() {
                  // immediately.
                  onSuccess: () => {
                    reindexSeenActiveRef.current = false;
+                    // Forget the previous run's identity so the first poll of
+                    // this window (carrying the new run's runId) is recognized
+                    // as a new run and keyed afresh.
+                    reindexRunKeyRef.current = null;
                    setReindexDeadline(Date.now() + REINDEX_POLL_CAP_MS);
                  },
                })
@@ -51,6 +51,14 @@ export interface IAiSettings {
  // True while a full workspace reindex is actively running; the counts above
  // then reflect the live run progress (done climbs 0 -> total).
  reindexing?: boolean;
+  // Identity of the ACTIVE reindex run (present only while `reindexing`). The
+  // poll keys on `runId`: a changed value means a NEW run (reset the per-run
+  // poll state the UI latched), the same value is the run already being watched.
+  // Absent/empty ('') => no identity available; the client keeps prior behaviour.
+  runId?: string;
+  // Epoch-ms the active run started; paired with `runId` so a restart with a
+  // recycled id is still detected as a new run.
+  reindexStartedAt?: number;
 }

 // Update payload. Key semantics (same for `apiKey` and `embeddingApiKey`):
@@ -1,5 +1,5 @@
 import { describe, it, expect } from "vitest";
-import { templateRoute } from "./route-template";
+import { templateRoute, KNOWN_ROUTE_TEMPLATES } from "./route-template";

 describe("templateRoute", () => {
  it("templates a space page path (never leaks slugs)", () => {
@@ -32,4 +32,30 @@ describe("templateRoute", () => {
    expect(templateRoute("/weird/unknown/thing")).toBe("other");
    expect(templateRoute("/s/team/p/slug/extra/segments")).toBe("other");
  });
+
+  // The server's /api/telemetry/vitals mirror (ALLOWED_ROUTE_TEMPLATES) drops any
+  // route outside KNOWN_ROUTE_TEMPLATES, so templateRoute must NEVER emit a label
+  // that is not in that dictionary — otherwise legit client metrics get dropped.
+  it("only ever emits labels contained in KNOWN_ROUTE_TEMPLATES (#495)", () => {
+    const samples = [
+      "/",
+      "/home",
+      "/settings/members",
+      "/settings/groups/g-1",
+      "/s/team",
+      "/s/team/trash",
+      "/s/team/p/slug",
+      "/p/slug",
+      "/share/abc",
+      "/share/abc/p/slug",
+      "/share/p/slug",
+      "/labels/urgent",
+      "/invites/inv-1",
+      "/weird/unknown/thing", // -> "other"
+      "/deep/unmatched/x/y/z", // -> "other"
+    ];
+    for (const path of samples) {
+      expect(KNOWN_ROUTE_TEMPLATES.has(templateRoute(path))).toBe(true);
+    }
+  });
 });
@@ -44,6 +44,22 @@ const STATIC_ROUTES = new Set<string>([
  '/settings/sharing',
 ]);

+/**
+ * The COMPLETE, finite vocabulary `templateRoute` can ever emit: the two
+ * synthetic labels (`/` and `other`), the static routes, and the dynamic
+ * templates. Exported so the public `/api/telemetry/vitals` endpoint can reject
+ * any `route` outside this dictionary server-side (the endpoint is anonymous, so
+ * an un-checked `route` is a free-text write surface). The server keeps a mirror
+ * (`ALLOWED_ROUTE_TEMPLATES` in client-metrics.constants.ts) — this is the
+ * canonical source; keep them in lockstep.
+ */
+export const KNOWN_ROUTE_TEMPLATES: ReadonlySet<string> = new Set<string>([
+  '/',
+  'other',
+  ...STATIC_ROUTES,
+  ...ROUTE_PATTERNS.map((p) => p.template),
+]);
+
 export function templateRoute(pathname: string): string {
  // Normalise a trailing slash (except root).
  const path =
@@ -117,9 +117,14 @@ const FINAL_STEP_NUDGE =
 // NO text at all (#444, mitigates the "empty turn" the lockdown used to prevent
 // when the toggle is OFF). Makes the exhausted-without-answer state explicit to
 // the user and, on replay, to the model on the next turn.
+// The persisted content is the app's base locale (en-US) — which is ALSO the
+// i18n key the client localizes through `t()` — instead of a hardcoded Russian
+// string (it used to render Russian for every locale, and fed Russian back to
+// the model on replay). Keep it a plain, model-readable English sentence so the
+// next turn's replay reads cleanly; the client resolves the locale.
 const STEP_LIMIT_NO_ANSWER_MARKER =
-  '(Достигнут лимит шагов — итоговый ответ не сформулирован; работа могла ' +
-  'остаться незавершённой. Напишите «продолжай», чтобы агент продолжил.)';
+  '(Step limit reached — no final answer was produced; the work may be ' +
+  'unfinished. Reply "continue" to let the agent carry on.)';

 // Reason recorded in ai_chat_runs.error / the assistant row when the token-
 // degeneration detector (#444) aborts a run. Distinct from a user Stop (no error)
@@ -1514,6 +1519,13 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
          system,
          messages,
          tools,
+          // Pin the AI SDK per-request retry budget explicitly instead of relying
+          // on its default (which is also 2). Connection arithmetic per turn:
+          // (1 + maxRetries=2) × (1 + AI_STREAM_PRE_RESPONSE_RETRIES) network
+          // connects worst-case — the two retry layers compose, so making the SDK
+          // side explicit keeps that ceiling visible and pinned against SDK-default
+          // drift.
+          maxRetries: 2,
          // No maxOutputTokens cap on the agent: tool-call arguments (e.g. a full
          // page body for the write tools) are emitted as OUTPUT tokens, so a fixed
          // cap would truncate complex tool calls mid-argument. Let the model use its
@@ -9,9 +9,73 @@ import {
  DEGENERATION_CHECK_STEP,
  REPEATED_LINES_THRESHOLD,
  MIN_PERIOD_REPEATS,
+  degenerationThresholds,
 } from './output-degeneration';
 import { AiChatService } from './ai-chat.service';

+// Part A (#495 iter10): the detector thresholds are env-tunable. These drive the
+// resolver against real repeat-count shapes and mutation-verify that the env
+// override actually changes the trigger point (not a vacuous read).
+describe('degeneration thresholds are env-configurable', () => {
+  const VARS = [
+    'AI_CHAT_DEGENERATION_REPEATED_LINES',
+    'AI_CHAT_DEGENERATION_PERIOD_MAX_LEN',
+    'AI_CHAT_DEGENERATION_PERIOD_MIN_REPEATS',
+    'AI_CHAT_DEGENERATION_CHECK_STEP',
+  ];
+  const saved: Record<string, string | undefined> = {};
+  beforeEach(() => {
+    for (const v of VARS) saved[v] = process.env[v];
+  });
+  afterEach(() => {
+    for (const v of VARS) {
+      if (saved[v] === undefined) delete process.env[v];
+      else process.env[v] = saved[v];
+    }
+  });
+
+  it('defaults to the compiled constants when unset', () => {
+    for (const v of VARS) delete process.env[v];
+    expect(degenerationThresholds()).toEqual({
+      repeatedLines: REPEATED_LINES_THRESHOLD,
+      maxPeriodLen: 150,
+      minPeriodRepeats: MIN_PERIOD_REPEATS,
+      checkStep: DEGENERATION_CHECK_STEP,
+    });
+  });
+
+  it('falls back to the default on blank / invalid / non-positive values', () => {
+    for (const bad of ['', '  ', 'abc', '0', '-3', '1.5']) {
+      process.env.AI_CHAT_DEGENERATION_REPEATED_LINES = bad;
+      // '1.5' floors to 1 (still ≥1, valid); every other bad value → default.
+      const expected = bad === '1.5' ? 1 : REPEATED_LINES_THRESHOLD;
+      expect(degenerationThresholds().repeatedLines).toBe(expected);
+    }
+  });
+
+  it('a RAISED check-step suppresses a burst the default would have flagged', () => {
+    // A ~3.3KB periodic burst is periodic-degenerate, but shouldCheckDegeneration
+    // is the throttle gate. Default checkStep=2000 arms on it; raising the step
+    // above the burst size means the throttle never re-fires for it.
+    const burstLen = 'loadTools.\n'.repeat(300).length; // ~3300
+    delete process.env.AI_CHAT_DEGENERATION_CHECK_STEP;
+    expect(shouldCheckDegeneration(burstLen, 0)).toBe(true); // default 2000
+    process.env.AI_CHAT_DEGENERATION_CHECK_STEP = String(burstLen + 1);
+    expect(shouldCheckDegeneration(burstLen, 0)).toBe(false); // raised gate
+  });
+
+  it('a LOWERED repeated-lines threshold trips on a shorter identical-line run', () => {
+    // 8 identical lines: below the default 25 (rule 1) and below the periodic
+    // rule's 20 repeats — so isDegenerateOutput is false by default.
+    const shortRun = 'x\n'.repeat(8);
+    delete process.env.AI_CHAT_DEGENERATION_REPEATED_LINES;
+    expect(isDegenerateOutput(shortRun)).toBe(false);
+    // Lower rule 1 to 5 → the 8-line run now trips.
+    process.env.AI_CHAT_DEGENERATION_REPEATED_LINES = '5';
+    expect(isDegenerateOutput(shortRun)).toBe(true);
+  });
+});
+
 // Mock ONLY streamText so we can capture the onChunk/onStepFinish callbacks the
 // service registers and drive them by hand; every other `ai` export the service
 // uses (convertToModelMessages, stepCountIs, …) stays real.
@@ -23,6 +23,54 @@ export const MAX_PERIOD_LEN = 150;
 /** Rule 2: minimum number of consecutive block repeats to trigger. */
 export const MIN_PERIOD_REPEATS = 20;

+/**
+ * Read a positive-integer threshold from an env var, falling back to `fallback`
+ * on unset/blank/invalid/non-positive. Mirrors the `AI_STREAM_PRE_RESPONSE_RETRIES`
+ * resolver in `ai-streaming-fetch.ts`: read the RAW string first so a blank value
+ * is treated as "unset" (→ fallback) rather than coercing to 0. Thresholds must
+ * stay ≥ 1 — a 0/negative would make the detector fire on any text (or never), so
+ * a bad value degrades to the safe compiled default instead. Env-tunable so an
+ * operator can retune the anti-babble guard (#444) without a redeploy, following
+ * the `AI_CHAT_FINAL_STEP_LOCKDOWN` toggle convention.
+ */
+function envThreshold(name: string, fallback: number): number {
+  const rawStr = process.env[name];
+  if (rawStr === undefined || rawStr.trim() === '') return fallback;
+  const raw = Number(rawStr);
+  return Number.isFinite(raw) && raw >= 1 ? Math.floor(raw) : fallback;
+}
+
+/**
+ * Resolve the degeneration-detector thresholds from the environment, each
+ * defaulting to the compiled constant above. Read fresh per call (not cached at
+ * import) so a test — or a runtime env change — takes effect deterministically.
+ */
+export function degenerationThresholds(): {
+  repeatedLines: number;
+  maxPeriodLen: number;
+  minPeriodRepeats: number;
+  checkStep: number;
+} {
+  return {
+    repeatedLines: envThreshold(
+      'AI_CHAT_DEGENERATION_REPEATED_LINES',
+      REPEATED_LINES_THRESHOLD,
+    ),
+    maxPeriodLen: envThreshold(
+      'AI_CHAT_DEGENERATION_PERIOD_MAX_LEN',
+      MAX_PERIOD_LEN,
+    ),
+    minPeriodRepeats: envThreshold(
+      'AI_CHAT_DEGENERATION_PERIOD_MIN_REPEATS',
+      MIN_PERIOD_REPEATS,
+    ),
+    checkStep: envThreshold(
+      'AI_CHAT_DEGENERATION_CHECK_STEP',
+      DEGENERATION_CHECK_STEP,
+    ),
+  };
+}
+
 /**
 * Rule 1 — ≥`REPEATED_LINES_THRESHOLD` consecutive IDENTICAL non-empty lines at
 * the tail. Catches the classic newline-delimited loop ("loadTools.\n" ×N).
@@ -128,7 +176,11 @@ export function hasPeriodicTail(
 * Pure — the caller owns the abort side effect.
 */
 export function isDegenerateOutput(text: string): boolean {
-  return hasRepeatedLineRun(text) || hasPeriodicTail(text);
+  const cfg = degenerationThresholds();
+  return (
+    hasRepeatedLineRun(text, cfg.repeatedLines) ||
+    hasPeriodicTail(text, cfg.maxPeriodLen, cfg.minPeriodRepeats)
+  );
 }

 /**
@@ -154,7 +206,7 @@ export function shouldCheckDegeneration(
  textLen: number,
  lastCheckLen: number,
 ): boolean {
-  return textLen - lastCheckLen >= DEGENERATION_CHECK_STEP;
+  return textLen - lastCheckLen >= degenerationThresholds().checkStep;
 }

 /**
@@ -307,6 +307,10 @@ export class PublicShareChatService {
        system,
        messages: modelMessages,
        tools,
+        // Pin the AI SDK per-request retry budget explicitly (matches the SDK
+        // default of 2). Connection arithmetic: (1 + maxRetries) × (1 +
+        // AI_STREAM_PRE_RESPONSE_RETRIES) worst-case connects per turn.
+        maxRetries: 2,
        // Bound the agent loop for anonymous callers.
        stopWhen: stepCountIs(5),
        // Cap per-request output so one anonymous call cannot run up the provider
@@ -17,9 +17,10 @@ import { MovePageDto } from './move-page.dto';
 //     a valid ordering key the server itself generated would be refused on move.
 //
 // The tests below assert the CORRECT contract: any key the generator can produce
-// must satisfy the DTO. The genuinely-failing case is marked `test.failing` so the
-// suite stays green while locking the bug; it flips red (alerting us) once the DTO
-// bounds are widened to cover the generator's real range.
+// must satisfy the DTO. FIXED (#495 item 9): the DTO now validates `position` by
+// CHARSET ([0-9A-Za-z], the generator's base-62 alphabet) instead of the wrong
+// @MaxLength(12) length bound, so dense between-inserts are accepted; the former
+// `test.failing` bug-lock is now a passing assertion.

 function constraintErrors(position: unknown) {
  const dto = plainToInstance(MovePageDto, {
@@ -47,24 +48,33 @@ describe('MovePageDto.position vs generateJitteredKeyBetween parity', () => {
    expect(hasError(errors, 'position')).toBe(false);
  });

-  // BUG LOCK: dense between-inserts produce keys longer than 12 chars, which
-  // MaxLength(12) rejects even though they are valid ordering keys. This SHOULD
-  // pass; it currently fails. Flips green when the DTO bound is fixed.
-  test.failing(
-    'accepts dense between-inserted keys (currently rejected by MaxLength(12))',
-    async () => {
-      let lo = generateJitteredKeyBetween(null, null);
-      let hi = generateJitteredKeyBetween(lo, null);
-      // Repeatedly insert just above `lo`, shrinking the gap so the key grows.
-      let longest = lo;
-      for (let i = 0; i < 40; i++) {
-        const mid = generateJitteredKeyBetween(lo, hi);
-        if (mid.length > longest.length) longest = mid;
-        hi = mid;
-      }
-      expect(longest.length).toBeGreaterThan(12); // sanity: we produced a long key
-      const errors = await constraintErrors(longest);
-      expect(hasError(errors, 'position')).toBe(false);
-    },
-  );
+  // FIXED: dense between-inserts produce keys longer than 12 chars, which the old
+  // MaxLength(12) rejected even though they are valid ordering keys. Now accepted.
+  it('accepts dense between-inserted keys longer than 12 chars', async () => {
+    let lo = generateJitteredKeyBetween(null, null);
+    let hi = generateJitteredKeyBetween(lo, null);
+    // Repeatedly insert just above `lo`, shrinking the gap so the key grows. The
+    // generator is JITTERED (random), so use enough iterations that the longest
+    // key reliably clears the old 12-char bound: measured min-over-50-trials is
+    // ~11 at 40 iterations (flaky) but ~36 at 200 (robust margin).
+    let longest = lo;
+    for (let i = 0; i < 200; i++) {
+      const mid = generateJitteredKeyBetween(lo, hi);
+      if (mid.length > longest.length) longest = mid;
+      hi = mid;
+    }
+    expect(longest.length).toBeGreaterThan(12); // sanity: we produced a long key
+    const errors = await constraintErrors(longest);
+    expect(hasError(errors, 'position')).toBe(false);
+  });
+
+  // The charset guard replaces the length bound: reject anything outside the
+  // generator's [0-9A-Za-z] alphabet (control chars, separators, injection) and
+  // the empty string, while still accepting every real key.
+  it('rejects a position with characters outside the fractional-index alphabet', async () => {
+    for (const bad of ['a0/b', 'a b', 'a\n0', 'a.b', '', "a';--"]) {
+      const errors = await constraintErrors(bad);
+      expect(hasError(errors, 'position')).toBe(true);
+    }
+  });
 });
@@ -1,8 +1,8 @@
 import {
  IsString,
  IsOptional,
-  MinLength,
  MaxLength,
+  Matches,
  IsNotEmpty,
 } from 'class-validator';

@@ -10,9 +10,19 @@ export class MovePageDto {
  @IsString()
  pageId: string;

+  // `position` is a fractional-indexing key from `generateJitteredKeyBetween`
+  // (the SAME generator page.service uses). Validate by CHARSET, not length: the
+  // generator's default base-62 alphabet is [0-9A-Za-z], and DENSE between-inserts
+  // legitimately grow a key well past a dozen chars (measured >40), so the old
+  // @MinLength(5)/@MaxLength(12) bounds wrongly 400'd valid ordering keys the
+  // server itself produced (Gitea #139 item 6). The charset regex rejects control
+  // chars / separators / injection, and a generous MaxLength stays only as a
+  // DoS guard — far above any realistic key, so it never rejects a real move.
  @IsString()
-  @MinLength(5)
-  @MaxLength(12)
+  @Matches(/^[0-9A-Za-z]+$/, {
+    message: 'position must be a fractional-index key ([0-9A-Za-z])',
+  })
+  @MaxLength(256)
  position: string;

  @IsOptional()
@@ -0,0 +1,49 @@
+import 'reflect-metadata';
+import { plainToInstance } from 'class-transformer';
+import { validate } from 'class-validator';
+import { PageIdDto } from './page.dto';
+
+// #435: PageIdDto.pageId carries a page's DOUBLE identity (internal UUID OR
+// public 10-char slugId), both as bare strings. The DTO must accept exactly
+// those two FORMATS and reject a malformed / swapped identity at the boundary.
+async function pageIdErrors(pageId: unknown) {
+  const dto = plainToInstance(PageIdDto, { pageId });
+  const errors = await validate(dto as object);
+  return errors.some((e) => e.property === 'pageId');
+}
+
+const UUID = '019f499a-9f8c-7d68-b7be-ce100d7c6c56';
+const SLUG = 'aB3xQ7kR2p';
+
+describe('PageIdDto pageId format validation', () => {
+  it('accepts a canonical page UUID', async () => {
+    expect(await pageIdErrors(UUID)).toBe(false);
+  });
+
+  it('accepts a 10-char slugId', async () => {
+    expect(await pageIdErrors(SLUG)).toBe(false);
+  });
+
+  it('rejects a truncated / wrong-length slug', async () => {
+    expect(await pageIdErrors('aB3xQ7kR2')).toBe(true); // 9 chars
+    expect(await pageIdErrors('aB3xQ7kR2pX')).toBe(true); // 11 chars
+  });
+
+  it('rejects a slug with an illegal character', async () => {
+    expect(await pageIdErrors('aB3xQ7kR2!')).toBe(true);
+  });
+
+  it('rejects a full URL / path-shaped identity (not the bare id)', async () => {
+    expect(await pageIdErrors(`my-page-title-${SLUG}`)).toBe(true);
+    expect(await pageIdErrors(`https://x/p/${SLUG}`)).toBe(true);
+  });
+
+  it('rejects a malformed UUID', async () => {
+    expect(await pageIdErrors('019f499a-9f8c-7d68-b7be')).toBe(true);
+    expect(await pageIdErrors('not-a-uuid-at-all-really')).toBe(true);
+  });
+
+  it('rejects an empty string', async () => {
+    expect(await pageIdErrors('')).toBe(true);
+  });
+});
@@ -0,0 +1,39 @@
+import { applyDecorators } from '@nestjs/common';
+import { Matches, ValidationOptions } from 'class-validator';
+
+/**
+ * A page identity at the API boundary is EITHER the internal page UUID or the
+ * public 10-char slugId (page.repo.findById matches a non-UUID input as a
+ * slugId). Both arrive as bare strings, which is exactly how the two got swapped
+ * silently (incident family #435). This regex pins the two accepted FORMATS so a
+ * malformed / cross-wired identity (a truncated slug, a full URL, an email, an
+ * id from another entity kind that isn't even shaped like either) is rejected at
+ * the boundary instead of falling through to a confusing 404.
+ *
+ *   - UUID:    canonical 8-4-4-4-12 hex (version-agnostic — page ids are UUIDv7,
+ *              so only the shape/length is enforced, matching the MCP's UUID_RE
+ *              and the server's isValidUUID acceptance).
+ *   - slugId:  exactly 10 chars over [0-9A-Za-z] (nanoid `generateSlugId`).
+ *
+ * The two are disjoint (a UUID is 36 chars WITH dashes, a slugId 10 chars
+ * WITHOUT), so a value can only satisfy one branch.
+ */
+export const PAGE_ID_OR_SLUG_ID_REGEX =
+  /^(?:[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}|[0-9A-Za-z]{10})$/i;
+
+/**
+ * Validate that a DTO string field is a well-formed page identity (page UUID OR
+ * 10-char slugId). Composed decorator so the same format rule is applied
+ * consistently wherever a DTO accepts a `pageId` that may be either form.
+ */
+export function IsPageIdOrSlugId(
+  validationOptions?: ValidationOptions,
+): PropertyDecorator {
+  return applyDecorators(
+    Matches(PAGE_ID_OR_SLUG_ID_REGEX, {
+      message:
+        'pageId must be a page UUID or a 10-character slugId',
+      ...validationOptions,
+    }),
+  );
+}
@@ -9,10 +9,16 @@ import {
 import { Transform } from 'class-transformer';

 import { ContentFormat } from './create-page.dto';
+import { IsPageIdOrSlugId } from './page-identity.validator';

 export class PageIdDto {
  @IsString()
  @IsNotEmpty()
+  // Format-validate the double identity (#435): accept only a page UUID or a
+  // 10-char slugId so a malformed / swapped identity is rejected at the boundary
+  // rather than passed to the repo as a bare string. Base for PageInfoDto,
+  // DeletePageDto, BacklinksListDto, AddLabelsDto/RemoveLabelDto, etc.
+  @IsPageIdOrSlugId()
  pageId: string;
 }

@@ -1,19 +1,38 @@
 import { TemporaryNoteCleanupService } from '../temporary-note-cleanup.service';

 /**
- * Chainable Kysely stub that records every `.where(...)` call so the test can
- * assert the sweep only selects armed, expired, not-yet-trashed notes. The
- * terminal `.execute()` resolves the configured expired rows (the batch SELECT);
- * `.executeTakeFirst()` resolves the per-row deadline re-read done just before
- * each `removePage`. By default the re-read reports the note as still armed and
- * still expired (epoch deadline < now), so the sweep proceeds to delete it;
- * tests override `reReadFirst` to simulate a concurrent "Make permanent".
+ * Chainable Kysely stub for the temporary-note sweep.
+ *
+ * `this.db` serves the non-locking candidate SELECT (selectFrom/select/where/
+ * limit/execute -> the configured expired rows) AND `.transaction().execute(cb)`,
+ * which runs `cb` with a separate `trx` builder. The `trx` builder serves the
+ * per-row LOCKED re-check (selectFrom/select/where/forUpdate/skipLocked/
+ * executeTakeFirst). `lockedRows` drives what that locked re-check returns per
+ * candidate — an id/creator/workspace row means "still expired, delete it";
+ * `undefined` means the predicate no longer matched (made permanent / re-armed /
+ * already trashed) or the row was SKIP-LOCKED by another worker, so it is skipped.
 */
-function makeDbStub(expiredRows: any[]) {
+function makeDbStub(expiredRows: any[], lockedRows?: any[]) {
  const whereCalls: any[][] = [];
-  const reReadFirst = jest
-    .fn()
-    .mockResolvedValue({ temporaryExpiresAt: new Date(0), deletedAt: null });
+  const locked = [
+    ...(lockedRows ??
+      expiredRows.map((r) => ({
+        id: r.id,
+        creatorId: r.creatorId,
+        workspaceId: r.workspaceId,
+      }))),
+  ];
+  const lockedTakeFirst = jest.fn(() => Promise.resolve(locked.shift()));
+  const forUpdate = jest.fn(() => trxBuilder);
+  const skipLocked = jest.fn(() => trxBuilder);
+  const trxBuilder: any = {
+    selectFrom: jest.fn(() => trxBuilder),
+    select: jest.fn(() => trxBuilder),
+    where: jest.fn(() => trxBuilder),
+    forUpdate,
+    skipLocked,
+    executeTakeFirst: lockedTakeFirst,
+  };
  const builder: any = {
    selectFrom: jest.fn(() => builder),
    select: jest.fn(() => builder),
@@ -23,9 +42,11 @@ function makeDbStub(expiredRows: any[]) {
    }),
    limit: jest.fn(() => builder),
    execute: jest.fn().mockResolvedValue(expiredRows),
-    executeTakeFirst: reReadFirst,
+    transaction: jest.fn(() => ({
+      execute: (cb: (trx: any) => Promise<any>) => Promise.resolve(cb(trxBuilder)),
+    })),
  };
-  return { builder, whereCalls, reReadFirst };
+  return { builder, whereCalls, lockedTakeFirst, forUpdate, skipLocked };
 }

 describe('TemporaryNoteCleanupService.sweepExpiredTemporaryNotes', () => {
@@ -52,20 +73,36 @@ describe('TemporaryNoteCleanupService.sweepExpiredTemporaryNotes', () => {
    expect(builder.limit.mock.calls[0][0]).toBeGreaterThan(0);
  });

-  it('soft-deletes each expired note via removePage, attributed to its creator', async () => {
+  it('soft-deletes each expired note via removePage under a row lock, attributed to its creator', async () => {
    const expired = [
      { id: 'p1', creatorId: 'u1', workspaceId: 'w1' },
      { id: 'p2', creatorId: 'u2', workspaceId: 'w1' },
    ];
-    const { builder } = makeDbStub(expired);
+    const { builder, forUpdate, skipLocked } = makeDbStub(expired);
    const pageRepo = { removePage: jest.fn().mockResolvedValue(undefined) } as any;
    const service = new TemporaryNoteCleanupService(builder, pageRepo);

    await service.sweepExpiredTemporaryNotes();

    expect(pageRepo.removePage).toHaveBeenCalledTimes(2);
-    expect(pageRepo.removePage).toHaveBeenNthCalledWith(1, 'p1', 'u1', 'w1');
-    expect(pageRepo.removePage).toHaveBeenNthCalledWith(2, 'p2', 'u2', 'w1');
+    // The 4th arg is the locking transaction — the delete runs inside it.
+    expect(pageRepo.removePage).toHaveBeenNthCalledWith(
+      1,
+      'p1',
+      'u1',
+      'w1',
+      expect.anything(),
+    );
+    expect(pageRepo.removePage).toHaveBeenNthCalledWith(
+      2,
+      'p2',
+      'u2',
+      'w1',
+      expect.anything(),
+    );
+    // The re-check acquired a FOR UPDATE SKIP LOCKED lock (once per candidate).
+    expect(forUpdate).toHaveBeenCalledTimes(2);
+    expect(skipLocked).toHaveBeenCalledTimes(2);
  });

  it('continues past a failing note (one bad removePage does not abort the sweep)', async () => {
@@ -86,60 +123,29 @@ describe('TemporaryNoteCleanupService.sweepExpiredTemporaryNotes', () => {
      service.sweepExpiredTemporaryNotes(),
    ).resolves.toBeUndefined();
    expect(pageRepo.removePage).toHaveBeenCalledTimes(2);
-    expect(pageRepo.removePage).toHaveBeenNthCalledWith(2, 'good', 'u2', 'w1');
+    expect(pageRepo.removePage).toHaveBeenNthCalledWith(
+      2,
+      'good',
+      'u2',
+      'w1',
+      expect.anything(),
+    );
  });

-  it('does NOT trash a note made permanent in the race window', async () => {
-    // The batch SELECT saw the note as expired, but before its turn in the loop
-    // the user clicked "Make permanent" (temporary_expires_at -> null). The
-    // deadline re-read must catch this and skip the delete so the keep wins.
+  it('does NOT trash a note made permanent / re-armed / already trashed (locked re-check returns nothing)', async () => {
+    // The batch SELECT saw the note as expired, but by the time the LOCKED
+    // re-check runs the row no longer matches the still-armed+expired+not-trashed
+    // predicate (make-permanent, re-arm to a future deadline, or already trashed),
+    // OR another worker holds the row (SKIP LOCKED). In every case the locked
+    // SELECT returns nothing and the delete is skipped so the keep/other worker wins.
    const expired = [{ id: 'p1', creatorId: 'u1', workspaceId: 'w1' }];
-    const { builder, reReadFirst } = makeDbStub(expired);
-    reReadFirst.mockResolvedValueOnce({
-      temporaryExpiresAt: null,
-      deletedAt: null,
-    });
+    const { builder, lockedTakeFirst } = makeDbStub(expired, [undefined]);
    const pageRepo = { removePage: jest.fn() } as any;
    const service = new TemporaryNoteCleanupService(builder, pageRepo);

    await service.sweepExpiredTemporaryNotes();

-    expect(reReadFirst).toHaveBeenCalledTimes(1);
-    expect(pageRepo.removePage).not.toHaveBeenCalled();
-  });
-
-  it('skips a note already trashed since the batch SELECT', async () => {
-    const expired = [{ id: 'p1', creatorId: 'u1', workspaceId: 'w1' }];
-    const { builder, reReadFirst } = makeDbStub(expired);
-    reReadFirst.mockResolvedValueOnce({
-      temporaryExpiresAt: new Date(0),
-      deletedAt: new Date(),
-    });
-    const pageRepo = { removePage: jest.fn() } as any;
-    const service = new TemporaryNoteCleanupService(builder, pageRepo);
-
-    await service.sweepExpiredTemporaryNotes();
-
-    expect(pageRepo.removePage).not.toHaveBeenCalled();
-  });
-
-  it('does NOT trash a note re-armed to a future deadline in the race window', async () => {
-    // The batch SELECT saw the note as expired, but before its turn in the loop
-    // the user disarmed it and re-armed it to a fresh, still-future deadline
-    // (temporary_expires_at -> now + 1h). The deadline re-read must catch that
-    // the note is no longer expired and skip the delete so the keep wins.
-    const expired = [{ id: 'p1', creatorId: 'u1', workspaceId: 'w1' }];
-    const { builder, reReadFirst } = makeDbStub(expired);
-    reReadFirst.mockResolvedValueOnce({
-      temporaryExpiresAt: new Date(Date.now() + 60 * 60 * 1000),
-      deletedAt: null,
-    });
-    const pageRepo = { removePage: jest.fn() } as any;
-    const service = new TemporaryNoteCleanupService(builder, pageRepo);
-
-    await service.sweepExpiredTemporaryNotes();
-
-    expect(reReadFirst).toHaveBeenCalledTimes(1);
+    expect(lockedTakeFirst).toHaveBeenCalledTimes(1);
    expect(pageRepo.removePage).not.toHaveBeenCalled();
  });

@@ -151,4 +157,31 @@ describe('TemporaryNoteCleanupService.sweepExpiredTemporaryNotes', () => {
    await service.sweepExpiredTemporaryNotes();
    expect(pageRepo.removePage).not.toHaveBeenCalled();
  });
+
+  it('sweeps once on application bootstrap (catches notes expired during downtime)', async () => {
+    const expired = [{ id: 'p1', creatorId: 'u1', workspaceId: 'w1' }];
+    const { builder } = makeDbStub(expired);
+    const pageRepo = { removePage: jest.fn().mockResolvedValue(undefined) } as any;
+    const service = new TemporaryNoteCleanupService(builder, pageRepo);
+
+    await service.onApplicationBootstrap();
+
+    expect(pageRepo.removePage).toHaveBeenCalledTimes(1);
+    expect(pageRepo.removePage).toHaveBeenCalledWith(
+      'p1',
+      'u1',
+      'w1',
+      expect.anything(),
+    );
+  });
+
+  it('a startup-sweep failure never blocks application boot', async () => {
+    const { builder } = makeDbStub([]);
+    // Make the candidate SELECT throw to simulate a boot-time DB hiccup.
+    builder.execute.mockRejectedValueOnce(new Error('db not ready'));
+    const pageRepo = { removePage: jest.fn() } as any;
+    const service = new TemporaryNoteCleanupService(builder, pageRepo);
+
+    await expect(service.onApplicationBootstrap()).resolves.toBeUndefined();
+  });
 });
@@ -1,8 +1,13 @@
-import { Injectable, Logger } from '@nestjs/common';
+import {
+  Injectable,
+  Logger,
+  OnApplicationBootstrap,
+} from '@nestjs/common';
 import { Interval } from '@nestjs/schedule';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB } from '@docmost/db/types/kysely.types';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
+import { executeTx } from '@docmost/db/utils';

 /**
 * Background sweeper for temporary notes ("structure or die"). A note whose
@@ -11,7 +16,7 @@ import { PageRepo } from '@docmost/db/repos/page/page.repo';
 * TrashCleanupService; `@nestjs/schedule` is already enabled globally.
 */
@Injectable()
-export class TemporaryNoteCleanupService {
+export class TemporaryNoteCleanupService implements OnApplicationBootstrap {
  private readonly logger = new Logger(TemporaryNoteCleanupService.name);

  // Cap a single sweep so a large backlog (e.g. many notes created during
@@ -24,6 +29,20 @@ export class TemporaryNoteCleanupService {
    private readonly pageRepo: PageRepo,
  ) {}

+  // Sweep once at startup so notes that expired during downtime are trashed
+  // right away instead of waiting up to an hour for the first @Interval tick.
+  // Best-effort: never let a startup-sweep failure block application boot.
+  async onApplicationBootstrap() {
+    try {
+      await this.sweepExpiredTemporaryNotes();
+    } catch (error) {
+      this.logger.error(
+        'Temporary-note startup sweep failed',
+        error instanceof Error ? error.stack : undefined,
+      );
+    }
+  }
+
  // Hourly granularity: lifetimes are configured in hours, so a sub-hour
  // overshoot past the deadline is acceptable.
  @Interval('temporary-note-cleanup', 60 * 60 * 1000)
@@ -31,9 +50,11 @@ export class TemporaryNoteCleanupService {
    try {
      const now = new Date();

+      // Candidate ids (non-locking). The authoritative re-check happens per row
+      // under a row lock below, so this cheap pass just bounds the batch.
      const expired = await this.db
        .selectFrom('pages')
-        .select(['id', 'creatorId', 'workspaceId'])
+        .select(['id'])
        .where('temporaryExpiresAt', 'is not', null)
        .where('temporaryExpiresAt', '<', now)
        .where('deletedAt', 'is', null) // not already in trash
@@ -41,50 +62,53 @@ export class TemporaryNoteCleanupService {
        .execute();

      let trashed = 0;
-      for (const page of expired) {
+      for (const candidate of expired) {
        try {
-          // Re-check the deadline at deletion time. The SELECT above is not
-          // transactional, so a user may click "Make permanent"
-          // (toggleTemporary sets temporary_expires_at = null) in the window
-          // between the SELECT and this per-row removePage. removePage deletes
-          // by id with only a `deletedAt IS NULL` filter and never re-reads the
-          // deadline, so without this guard a concurrently-kept note would
-          // still be trashed. Re-read the row and skip it unless it is still
-          // armed AND still expired, so a concurrent make-permanent wins.
-          const current = await this.db
-            .selectFrom('pages')
-            .select(['temporaryExpiresAt', 'deletedAt'])
-            .where('id', '=', page.id)
-            .executeTakeFirst();
+          const didTrash = await executeTx(this.db, async (trx) => {
+            // Re-check the row UNDER A LOCK inside the transaction. `FOR UPDATE
+            // SKIP LOCKED`:
+            //  - serialises against a concurrent "Make permanent"
+            //    (toggleTemporary UPDATE takes the same row lock): if it commits
+            //    first, the deadline predicate below no longer matches and we
+            //    skip; if we lock first, it waits until this delete commits.
+            //  - SKIP LOCKED lets a second worker/instance skip a row another
+            //    sweeper already claimed instead of blocking on it (no double
+            //    processing, no thundering herd).
+            // The predicate re-asserts still-armed AND still-expired AND
+            // not-already-trashed, so a make-permanent / prior sweep drops the row.
+            const locked = await trx
+              .selectFrom('pages')
+              .select(['id', 'creatorId', 'workspaceId'])
+              .where('id', '=', candidate.id)
+              .where('temporaryExpiresAt', 'is not', null)
+              .where('temporaryExpiresAt', '<', now)
+              .where('deletedAt', 'is', null)
+              .forUpdate()
+              .skipLocked()
+              .executeTakeFirst();

-          if (
-            !current ||
-            current.deletedAt !== null ||
-            current.temporaryExpiresAt === null ||
-            new Date(current.temporaryExpiresAt) >= now
-          ) {
-            // Made permanent, already trashed, or no longer expired since the
-            // SELECT — leave it alone.
-            continue;
-          }
+            if (!locked) return false;

-          // Reuse the exact soft-delete path: recursive over children, removes
-          // shares in a transaction, and emits PAGE_SOFT_DELETED (tree
-          // invalidation + watcher notifications). Attribute the automatic
-          // deletion to the note's creator (no schema change). Both the SELECT
-          // above and removePage filter `deletedAt IS NULL`, so a double sweep
-          // is idempotent.
-          await this.pageRepo.removePage(
-            page.id,
-            // creatorId is set on every created page; a temporary note always
-            // has one. Cast to satisfy the non-null deletedById parameter.
-            page.creatorId as string,
-            page.workspaceId,
-          );
-          trashed++;
+            // Reuse the exact soft-delete path (recursive children + share
+            // removal + PAGE_SOFT_DELETED broadcast), running IN this locked
+            // transaction so the delete is atomic with the re-check and cannot
+            // deadlock on a nested independent transaction. The broadcast is
+            // deferred by removePage to this transaction's commit. Attribute the
+            // automatic deletion to the note's creator (no schema change).
+            await this.pageRepo.removePage(
+              locked.id,
+              // creatorId is set on every created page; a temporary note always
+              // has one. Cast to satisfy the non-null deletedById parameter.
+              locked.creatorId as string,
+              locked.workspaceId,
+              trx,
+            );
+            return true;
+          });
+          if (didTrash) trashed++;
        } catch (error) {
          this.logger.error(
-            `Failed to trash expired temporary note ${page.id}`,
+            `Failed to trash expired temporary note ${candidate.id}`,
            error instanceof Error ? error.stack : undefined,
          );
        }
@@ -133,6 +133,9 @@ describe('ShareAliasController authz gates', () => {
        creatorId: 'u-1',
        alias: 'promo',
        confirmReassign: true,
+        // The requesting user is forwarded so setAlias can gate the reassign
+        // 409 title disclosure on target-page view permission (#495).
+        user,
      });
      expect(result).toEqual({ id: 'alias-1' });
    });
@@ -79,6 +79,9 @@ export class ShareAliasController {
      creatorId: user.id,
      alias: dto.alias,
      confirmReassign: dto.confirmReassign,
+      // Gates whether the reassign 409 may reveal the current target's title
+      // (view-permission check on that page) — see setAlias (#495).
+      user,
    });
  }

@@ -1,4 +1,8 @@
-import { BadRequestException, ConflictException } from '@nestjs/common';
+import {
+  BadRequestException,
+  ConflictException,
+  ForbiddenException,
+} from '@nestjs/common';
 import { NoResultError } from 'kysely';
 import { ShareAliasService } from './share-alias.service';

@@ -7,6 +11,8 @@ import { ShareAliasService } from './share-alias.service';
 * 409 reassign guard, uniqueness-race handling, availability probe, and the
 * request-time readable-target resolution (which re-runs the share boundary).
 */
+const USER = { id: 'u-1' } as any;
+
 describe('ShareAliasService', () => {
  // Sentinel handed to repo calls so tests can assert they ran inside the tx.
  const trx = { __trx: true };
@@ -27,6 +33,10 @@ describe('ShareAliasService', () => {
      resolveReadableSharePage: jest.fn(),
      isSharingAllowed: jest.fn(),
    };
+    // Default: the requester CAN view the target page (validateCanView resolves),
+    // so the reassign 409 may disclose its title. Tests override to reject to
+    // assert the no-leak path.
+    const pageAccessService = { validateCanView: jest.fn().mockResolvedValue(undefined) };
    // Fake kysely db: only .transaction().execute(cb) is used by setAlias.
    const db = {
      transaction: jest.fn(() => ({
@@ -37,9 +47,10 @@ describe('ShareAliasService', () => {
      shareAliasRepo as any,
      pageRepo as any,
      shareService as any,
+      pageAccessService as any,
      db as any,
    );
-    return { service, shareAliasRepo, pageRepo, shareService, db };
+    return { service, shareAliasRepo, pageRepo, shareService, pageAccessService, db };
  }

  describe('setAlias', () => {
@@ -50,6 +61,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'A', // too short + uppercase
        }),
      ).rejects.toBeInstanceOf(BadRequestException);
@@ -66,6 +78,7 @@ describe('ShareAliasService', () => {
        workspaceId: 'ws-1',
        pageId: 'p-1',
        creatorId: 'u-1',
+        user: USER,
        alias: '  My Page ',
      });

@@ -114,6 +127,7 @@ describe('ShareAliasService', () => {
        workspaceId: 'ws-1',
        pageId: 'p-1',
        creatorId: 'u-1',
+        user: USER,
        alias: 'ted',
      });

@@ -144,6 +158,7 @@ describe('ShareAliasService', () => {
        workspaceId: 'ws-1',
        pageId: 'p-1',
        creatorId: 'u-1',
+        user: USER,
        alias: 'foo',
      });

@@ -179,6 +194,7 @@ describe('ShareAliasService', () => {
        workspaceId: 'ws-1',
        pageId: 'p-1',
        creatorId: 'u-1',
+        user: USER,
        alias: 'new',
      });

@@ -190,30 +206,77 @@ describe('ShareAliasService', () => {
      );
    });

-    it('throws 409 with current target when name is taken and not confirmed', async () => {
-      const { service, shareAliasRepo, pageRepo } = makeService();
+    it('throws 409 with the target TITLE (never its id) when the requester CAN view it', async () => {
+      const { service, shareAliasRepo, pageRepo, pageAccessService } =
+        makeService();
      shareAliasRepo.findByAliasAndWorkspace.mockResolvedValue({
        id: 'a-1',
        alias: 'foo',
        pageId: 'p-other',
      });
      pageRepo.findById.mockResolvedValue({ id: 'p-other', title: 'Other' });
+      pageAccessService.validateCanView.mockResolvedValue(undefined); // can view

      try {
        await service.setAlias({
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        });
        fail('expected ConflictException');
      } catch (err) {
        expect(err).toBeInstanceOf(ConflictException);
-        expect((err as ConflictException).getResponse()).toMatchObject({
+        const body = (err as ConflictException).getResponse();
+        expect(body).toMatchObject({
          code: 'ALIAS_REASSIGN_REQUIRED',
-          currentPageId: 'p-other',
          currentPageTitle: 'Other',
        });
+        // SECURITY (#495): the page id is NEVER disclosed, even to a viewer.
+        expect(body).not.toHaveProperty('currentPageId');
+        expect(pageAccessService.validateCanView).toHaveBeenCalledWith(
+          expect.objectContaining({ id: 'p-other' }),
+          USER,
+        );
+      }
+      expect(shareAliasRepo.updatePageId).not.toHaveBeenCalled();
+    });
+
+    it('throws 409 WITHOUT the title or id when the requester CANNOT view the target (#495)', async () => {
+      const { service, shareAliasRepo, pageRepo, pageAccessService } =
+        makeService();
+      shareAliasRepo.findByAliasAndWorkspace.mockResolvedValue({
+        id: 'a-1',
+        alias: 'foo',
+        pageId: 'p-secret',
+      });
+      pageRepo.findById.mockResolvedValue({ id: 'p-secret', title: 'Secret' });
+      // No view permission on the target page -> validateCanView throws.
+      pageAccessService.validateCanView.mockRejectedValue(
+        new ForbiddenException(),
+      );
+
+      try {
+        await service.setAlias({
+          workspaceId: 'ws-1',
+          pageId: 'p-1',
+          creatorId: 'u-1',
+          user: USER,
+          alias: 'foo',
+        });
+        fail('expected ConflictException');
+      } catch (err) {
+        expect(err).toBeInstanceOf(ConflictException);
+        const body = (err as ConflictException).getResponse() as Record<
+          string,
+          unknown
+        >;
+        expect(body).toMatchObject({ code: 'ALIAS_REASSIGN_REQUIRED' });
+        // The enumeration hole: neither the id nor the title of a page the
+        // requester cannot see may leak.
+        expect(body).not.toHaveProperty('currentPageId');
+        expect(body.currentPageTitle ?? null).toBeNull();
      }
      expect(shareAliasRepo.updatePageId).not.toHaveBeenCalled();
    });
@@ -231,6 +294,7 @@ describe('ShareAliasService', () => {
        workspaceId: 'ws-1',
        pageId: 'p-1',
        creatorId: 'u-1',
+        user: USER,
        alias: 'foo',
        confirmReassign: true,
      });
@@ -269,6 +333,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        });
        fail('expected ConflictException');
@@ -294,6 +359,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        });
        fail('expected ConflictException');
@@ -317,6 +383,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        });
        fail('expected ConflictException');
@@ -346,6 +413,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        });
        fail('expected ConflictException');
@@ -375,6 +443,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
          confirmReassign: true,
        });
@@ -406,6 +475,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'ted',
        });
        fail('expected ConflictException');
@@ -428,6 +498,7 @@ describe('ShareAliasService', () => {
          workspaceId: 'ws-1',
          pageId: 'p-1',
          creatorId: 'u-1',
+          user: USER,
          alias: 'foo',
        }),
      ).rejects.toBeInstanceOf(BadRequestException);
@@ -450,18 +521,22 @@ describe('ShareAliasService', () => {
        alias: 'free-name',
        valid: true,
        available: true,
-        currentPageId: null,
      });
+      // SECURITY (#495): the availability probe must NOT leak any page id.
+      expect(res).not.toHaveProperty('currentPageId');
    });

-    it('reports taken with the current target page', async () => {
+    it('reports taken WITHOUT leaking the current target page id (#495)', async () => {
      const { service, shareAliasRepo } = makeService();
      shareAliasRepo.findByAliasAndWorkspace.mockResolvedValue({
        id: 'a-1',
        pageId: 'p-9',
      });
      const res = await service.checkAvailability('taken', 'ws-1');
-      expect(res).toMatchObject({ available: false, currentPageId: 'p-9' });
+      expect(res).toMatchObject({ available: false });
+      // The row exists (available:false) but its pageId is never returned — an
+      // authenticated member cannot map an alias name to a page id it can't view.
+      expect(res).not.toHaveProperty('currentPageId');
    });
  });

@@ -7,7 +7,8 @@ import {
 import { ShareAliasRepo } from '@docmost/db/repos/share-alias/share-alias.repo';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
 import { ShareService } from './share.service';
-import { Page, ShareAlias } from '@docmost/db/types/entity.types';
+import { PageAccessService } from '../page/page-access/page-access.service';
+import { Page, ShareAlias, User } from '@docmost/db/types/entity.types';
 import { isValidShareAlias, normalizeShareAlias } from './share-alias.util';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB } from '@docmost/db/types/kysely.types';
@@ -43,6 +44,7 @@ export class ShareAliasService {
    private readonly shareAliasRepo: ShareAliasRepo,
    private readonly pageRepo: PageRepo,
    private readonly shareService: ShareService,
+    private readonly pageAccessService: PageAccessService,
    @InjectKysely() private readonly db: KyselyDB,
  ) {}

@@ -55,9 +57,13 @@ export class ShareAliasService {
   *     `/l/<old>` link survives
   *   - name already points at pageId    -> no-op (idempotent)
   *   - name points at ANOTHER page      -> the "swap". Without confirmReassign
-   *     we throw 409 carrying the current target so the client can confirm;
-   *     with it we UPDATE the single row's page_id (every /l/<alias> link
-   *     follows the 302 to the new page instantly — no stale cache).
+   *     we throw 409 so the client can confirm. SECURITY (#495): the 409 reveals
+   *     the current target's title ONLY when `user` may VIEW that page, and never
+   *     its id — otherwise any member with one editable+shared page could iterate
+   *     alias names with confirmReassign=false and map them to (id, title) of
+   *     pages they cannot see. With confirmReassign we UPDATE the single row's
+   *     page_id (every /l/<alias> link follows the 302 to the new page instantly
+   *     — no stale cache).
   *
   * To keep the invariant self-healing we DELETE every other alias row still
   * pointing at this page (a legacy duplicate, or the target page's own former
@@ -77,8 +83,12 @@ export class ShareAliasService {
    creatorId: string;
    alias: string;
    confirmReassign?: boolean;
+    // The requesting user — used ONLY to gate whether the reassign 409 may reveal
+    // the current target page's title (view-permission check). Not an authz gate
+    // for the write itself (the controller already validated edit on `pageId`).
+    user: User;
  }): Promise<ShareAlias> {
-    const { workspaceId, pageId, creatorId, confirmReassign } = opts;
+    const { workspaceId, pageId, creatorId, confirmReassign, user } = opts;
    const alias = normalizeShareAlias(opts.alias);
    if (!isValidShareAlias(alias)) {
      throw new BadRequestException(
@@ -97,14 +107,30 @@ export class ShareAliasService {
        // The name is occupied by a DIFFERENT (or dangling) target page.
        if (byName && byName.pageId !== pageId) {
          if (!confirmReassign) {
+            // SECURITY (#495): only disclose the current target's TITLE, and only
+            // when the requester may VIEW that page. Never disclose its id (the
+            // client's confirm-reassign UX doesn't use it, and it is an enumerable
+            // identity). A member with one editable+shared page must NOT be able to
+            // iterate alias names and map them to (id, title) of pages they cannot
+            // see. When view is denied (or the alias is dangling) the 409 is the
+            // bare "occupied" fact — the client still shows a generic confirm modal.
            const currentPage = byName.pageId
              ? await this.pageRepo.findById(byName.pageId)
              : null;
+            let currentPageTitle: string | null = null;
+            if (currentPage) {
+              try {
+                await this.pageAccessService.validateCanView(currentPage, user);
+                currentPageTitle = currentPage.title ?? null;
+              } catch {
+                // No view permission on the target -> do not reveal its title.
+                currentPageTitle = null;
+              }
+            }
            throw new ConflictException({
              message: 'Alias already in use',
              code: 'ALIAS_REASSIGN_REQUIRED',
-              currentPageId: byName.pageId,
-              currentPageTitle: currentPage?.title ?? null,
+              currentPageTitle,
            });
          }
          // Confirmed swap. ORDER MATTERS: the partial unique index on
@@ -223,21 +249,27 @@ export class ShareAliasService {
    alias: string;
    valid: boolean;
    available: boolean;
-    currentPageId: string | null;
  }> {
    const alias = normalizeShareAlias(rawAlias);
    if (!isValidShareAlias(alias)) {
-      return { alias, valid: false, available: false, currentPageId: null };
+      return { alias, valid: false, available: false };
    }
    const existing = await this.shareAliasRepo.findByAliasAndWorkspace(
      alias,
      workspaceId,
    );
+    // SECURITY (#495): return ONLY the boolean availability. The previous shape
+    // leaked `currentPageId` — the id of whatever page the alias already targets —
+    // to ANY authenticated workspace member, with no view-permission check on that
+    // page. An attacker could enumerate alias names and map them to page ids they
+    // have no access to. The taken/free bit is all the "is this address free"
+    // probe needs. The reassign flow (setAlias 409) may surface the target's
+    // TITLE, but only behind a `validateCanView` on that page (see setAlias); it
+    // never returns the page id.
    return {
      alias,
      valid: true,
      available: !existing,
-      currentPageId: existing?.pageId ?? null,
    };
  }

@@ -24,6 +24,54 @@ export const ALLOWED_RATINGS = new Set<string>([
  'poor',
 ]);

+// The ONLY route labels accepted. The endpoint is anonymous, so an un-checked
+// `route` is a free-text write surface (arbitrary high-cardinality strings /
+// injected text into the metrics table). The client only ever sends a label from
+// a finite template dictionary (`templateRoute`), so we drop anything not in it.
+//
+// PARITY: this MUST mirror `KNOWN_ROUTE_TEMPLATES` in the client's
+// `apps/client/src/lib/telemetry/route-template.ts` (the canonical source). A
+// drift means legit client routes get dropped — keep the two in lockstep; the
+// client self-consistency test asserts `templateRoute` only emits these values.
+export const ALLOWED_ROUTE_TEMPLATES = new Set<string>([
+  '/',
+  'other',
+  // Static routes.
+  '/home',
+  '/spaces',
+  '/favorites',
+  '/login',
+  '/forgot-password',
+  '/password-reset',
+  '/setup/register',
+  '/settings/account/profile',
+  '/settings/account/preferences',
+  '/settings/workspace',
+  '/settings/ai',
+  '/settings/members',
+  '/settings/groups',
+  '/settings/spaces',
+  '/settings/sharing',
+  // Dynamic templates (slugs/ids are already collapsed to `:param`).
+  '/share/:shareId/p/:slug',
+  '/share/p/:slug',
+  '/share/:shareId',
+  '/p/:slug',
+  '/s/:space/p/:slug',
+  '/s/:space/trash',
+  '/s/:space',
+  '/labels/:label',
+  '/invites/:invitationId',
+  '/settings/groups/:groupId',
+]);
+
+// `attr` is a web-vitals attribution TARGET: a CSS-selector-ish string (an
+// element path like `html>body>div#app>button.cta`), never free prose. Constrain
+// it to a conservative CSS-selector charset so the anonymous endpoint cannot be
+// used to write arbitrary text / PII / markup into the metrics table. A value
+// containing anything outside this set is DROPPED (-> null); the event is kept.
+export const ATTR_ALLOWED_CHARSET = /^[A-Za-z0-9#.\-_> :()[\]="'*+~,]+$/;
+
 // Max events accepted per batch; the rest are ignored.
 export const MAX_EVENTS_PER_BATCH = 50;

@@ -77,14 +125,20 @@ export function sanitizeVitalEvent(
      ? e.rating
      : null;

+  // route: accept ONLY a known template label (dictionary check), else drop to
+  // null. The length cap stays as a cheap pre-guard before the Set lookup.
  let route: string | null = null;
  if (typeof e.route === 'string' && e.route.length > 0) {
-    route = e.route.slice(0, MAX_ROUTE_LENGTH);
+    const candidate = e.route.slice(0, MAX_ROUTE_LENGTH);
+    route = ALLOWED_ROUTE_TEMPLATES.has(candidate) ? candidate : null;
  }

+  // attr: truncate, then accept ONLY if it is a CSS-selector-shaped string
+  // (charset whitelist); anything with characters outside the set is dropped.
  let attr: string | null = null;
  if (typeof e.attr === 'string' && e.attr.length > 0) {
-    attr = e.attr.slice(0, MAX_ATTR_LENGTH);
+    const candidate = e.attr.slice(0, MAX_ATTR_LENGTH);
+    attr = ATTR_ALLOWED_CHARSET.test(candidate) ? candidate : null;
  }

  let docSize: number | null = null;
@@ -90,6 +90,44 @@ describe('VitalsService.buildRows', () => {
    expect(rows[0].attr).toHaveLength(MAX_ATTR_LENGTH);
  });

+  it('keeps a known route template but DROPS an unknown/free-text route (#495)', () => {
+    const rows = svc.buildRows(
+      {
+        events: [
+          { name: 'INP', value: 1, route: '/s/:space/p/:slug' }, // known
+          { name: 'INP', value: 2, route: '/s/acme-corp/p/secret-slug' }, // raw path (slugs) — not a template
+          { name: 'INP', value: 3, route: 'DROP TABLE client_metrics;--' }, // injected free text
+          { name: 'INP', value: 4, route: '/home' }, // known static
+        ],
+      },
+      WS,
+    );
+    expect(rows.map((r) => r.route)).toEqual([
+      '/s/:space/p/:slug',
+      null, // raw path dropped
+      null, // free text dropped
+      '/home',
+    ]);
+  });
+
+  it('DROPS an attr that is not a CSS-selector-shaped string (#495)', () => {
+    const rows = svc.buildRows(
+      {
+        events: [
+          { name: 'INP', value: 1, attr: 'div#app>button.cta' }, // valid selector
+          { name: 'INP', value: 2, attr: 'user@example.com wrote a note' }, // free text / PII
+          { name: 'INP', value: 3, attr: '<script>alert(1)</script>' }, // markup
+        ],
+      },
+      WS,
+    );
+    expect(rows.map((r) => r.attr)).toEqual([
+      'div#app>button.cta',
+      null,
+      null,
+    ]);
+  });
+
  it('caps the batch at 50 events', () => {
    const events = Array.from({ length: 200 }, () => ({ name: 'CLS', value: 1 }));
    const rows = svc.buildRows({ events }, WS);
@@ -0,0 +1,114 @@
+import * as path from 'path';
+import { readFileSync } from 'fs';
+
+// Mock ONLY kysely's `sql.raw(...).execute()` so we can observe what
+// ensureConcurrentIndexes runs and how it handles failures, without a DB.
+const execMock = jest.fn((_db: unknown) => Promise.resolve(undefined));
+const rawMock = jest.fn((_stmt: string) => ({ execute: execMock }));
+jest.mock('kysely', () => {
+  const actual = jest.requireActual('kysely');
+  return { ...actual, sql: { ...actual.sql, raw: rawMock } };
+});
+
+import { CONCURRENT_INDEXES, ensureConcurrentIndexes } from './concurrent-indexes';
+
+describe('ensureConcurrentIndexes', () => {
+  const fakeDb = { __topLevelKysely: true } as never;
+
+  beforeEach(() => {
+    execMock.mockReset().mockResolvedValue(undefined);
+    rawMock.mockClear();
+  });
+
+  it('redefines the inlinable f_unaccent BEFORE building any index, then builds each CONCURRENTLY outside a transaction', async () => {
+    const onLog = jest.fn();
+    await ensureConcurrentIndexes(fakeDb, onLog);
+
+    // One f_unaccent redefine + one statement per index.
+    expect(rawMock).toHaveBeenCalledTimes(CONCURRENT_INDEXES.length + 1);
+    const statements = rawMock.mock.calls.map((c) => c[0] as string);
+    // ORDER: the f_unaccent redefine MUST be first — otherwise an existing
+    // tenant's old 2-arg f_unaccent makes every CONCURRENTLY build fail.
+    expect(statements[0]).toContain('CREATE OR REPLACE FUNCTION f_unaccent');
+    expect(statements[0]).toContain('SELECT public.unaccent($1)');
+    expect(statements[0]).not.toContain('CREATE INDEX');
+    // The rest are the CONCURRENTLY index builds.
+    for (const stmt of statements.slice(1)) {
+      expect(stmt).toContain('CREATE INDEX CONCURRENTLY');
+      expect(stmt).toContain('IF NOT EXISTS');
+    }
+    // Executed against the top-level db (a transaction would forbid CONCURRENTLY).
+    for (const call of execMock.mock.calls) {
+      expect(call[0]).toBe(fakeDb);
+    }
+    expect(onLog).toHaveBeenCalledTimes(CONCURRENT_INDEXES.length + 1);
+  });
+
+  it('still builds the indexes when the f_unaccent redefine fails (fresh DB, best-effort)', async () => {
+    // Redefine (first execute) throws; the index loop must still run.
+    execMock
+      .mockRejectedValueOnce(new Error('extension "unaccent" does not exist'))
+      .mockResolvedValue(undefined);
+    const onLog = jest.fn();
+
+    await expect(
+      ensureConcurrentIndexes(fakeDb, onLog),
+    ).resolves.toBeUndefined();
+
+    // 1 redefine attempt + one per index.
+    expect(execMock).toHaveBeenCalledTimes(CONCURRENT_INDEXES.length + 1);
+    const errored = onLog.mock.calls.filter((c) => c[1] !== undefined);
+    expect(errored).toHaveLength(1);
+    expect(String(errored[0][1])).toContain('unaccent');
+  });
+
+  it('is best-effort: a failing index does not abort the rest and is reported', async () => {
+    // Redefine ok; fail the FIRST index; the remaining ones must still be tried.
+    execMock
+      .mockResolvedValueOnce(undefined) // f_unaccent redefine
+      .mockRejectedValueOnce(new Error('relation "pages" does not exist'))
+      .mockResolvedValue(undefined);
+    const onLog = jest.fn();
+
+    await expect(
+      ensureConcurrentIndexes(fakeDb, onLog),
+    ).resolves.toBeUndefined();
+
+    expect(execMock).toHaveBeenCalledTimes(CONCURRENT_INDEXES.length + 1);
+    const errored = onLog.mock.calls.filter((c) => c[1] !== undefined);
+    expect(errored).toHaveLength(1);
+    expect(String(errored[0][1])).toContain('does not exist');
+  });
+});
+
+// DRIFT GUARD: each CONCURRENT_INDEXES entry pre-builds an index that a plain
+// migration ALSO creates with `IF NOT EXISTS`. If the two expressions diverge,
+// Postgres would treat them as different indexes and the pre-build would NOT
+// make the migration a no-op. Assert the migration files still contain each
+// index's functional expression.
+describe('CONCURRENT_INDEXES parity with the migrations', () => {
+  const migrationsDir = path.join(__dirname, 'migrations');
+  const files = [
+    '20260705T120000-perf-indexes.ts',
+    '20260706T120000-search-lookup-trgm.ts',
+  ].map((f) => readFileSync(path.join(migrationsDir, f), 'utf8'));
+  const allMigrationSrc = files.join('\n');
+
+  it.each(CONCURRENT_INDEXES.map((i) => [i.name, i]))(
+    'migration source still creates %s with the same expression',
+    (_name, idx) => {
+      // Extract the `USING gin ((...expr...) gin_trgm_ops)` tail from the
+      // canonical create and assert the migration source contains it verbatim.
+      const m = (idx as { create: string }).create.match(
+        /ON \w+ (USING gin .+)$/,
+      );
+      expect(m).not.toBeNull();
+      const expr = (m as RegExpMatchArray)[1];
+      expect(allMigrationSrc).toContain(expr);
+      // And the migration must build it by the SAME index name.
+      expect(allMigrationSrc).toContain(
+        `CREATE INDEX IF NOT EXISTS ${(idx as { name: string }).name}`,
+      );
+    },
+  );
+});
@@ -0,0 +1,142 @@
+import { Kysely, sql } from 'kysely';
+
+/**
+ * Indexes that MUST be built with `CREATE INDEX CONCURRENTLY` so an auto-deploy
+ * migration never takes a `SHARE` lock that blocks writes on a hot table
+ * (`pages`) for the — potentially minutes-long — GIN trigram build (#495 item 12).
+ *
+ * Kysely runs each migration INSIDE a transaction (Postgres has transactional
+ * DDL), and `CREATE INDEX CONCURRENTLY` cannot run inside a transaction block, so
+ * these cannot live in an ordinary migration. Instead {@link ensureConcurrentIndexes}
+ * builds them out-of-band (no transaction) BEFORE the migrator runs; the matching
+ * migrations keep a plain `CREATE INDEX IF NOT EXISTS` as a backstop, which then
+ * no-ops because the index already exists. So:
+ *   - existing prod DB, incremental deploy: the pre-build FIRST redefines
+ *     `f_unaccent` to its index-inlinable 1-arg form (see below), THEN builds each
+ *     index CONCURRENTLY (no write lock); the migration's IF NOT EXISTS no-ops →
+ *     the write-blocking build is gone;
+ *   - fresh DB (or a DB whose `pages` / `unaccent` extension does not exist yet):
+ *     the pre-build fails and is swallowed (best-effort), and the migration builds
+ *     the index normally on an empty/small table where the lock is irrelevant.
+ *
+ * NOT a strict "worst case = previous behaviour":
+ *   - The f_unaccent form matters. The trigram expressions use `LOWER(f_unaccent(col))`.
+ *     The INDEX-INLINABLE 1-arg `f_unaccent(text)` (`SELECT public.unaccent($1)`)
+ *     is created INSIDE migration 20260705, which runs AFTER this pre-build. On an
+ *     existing tenant the live `f_unaccent` is still the OLD 2-arg form from
+ *     20250729, which Postgres CANNOT inline into a CONCURRENTLY index expression
+ *     (`function unaccent(unknown, text) does not exist ... during inlining`) — the
+ *     build fails outright. So this pre-build redefines `f_unaccent` to the 1-arg
+ *     form FIRST (idempotent, output-identical, in lockstep with 20260705). Without
+ *     that redefine the whole feature is dead-on-arrival for the very case it
+ *     targets.
+ *   - An INTERRUPTED `CREATE INDEX CONCURRENTLY` (killed pod, cancelled query) with
+ *     a working `f_unaccent` leaves an INVALID index behind. A subsequent name-based
+ *     `IF NOT EXISTS` — in BOTH this pre-build and the migration backstop — sees the
+ *     name and skips, so the invalid index is NEVER repaired automatically and the
+ *     query keeps seq-scanning until an operator `DROP INDEX`es it. The old
+ *     in-transaction build could not leave an invalid index (a failed tx rolled the
+ *     whole index back), so this is a genuinely new failure mode, not "= previous".
+ *     (A future hardening could `DROP` an `indisvalid = false` index before
+ *     rebuilding; not done here.)
+ *
+ * The `create` text is the CANONICAL definition — it MUST match the migration's
+ * `IF NOT EXISTS` create expression exactly (same functional expression + opclass)
+ * or Postgres would treat them as two different indexes.
+ */
+export const CONCURRENT_INDEXES: ReadonlyArray<{
+  name: string;
+  create: string;
+}> = [
+  {
+    // #348 perf-indexes — pages.title trigram (coalesce-free functional expr).
+    name: 'idx_pages_title_trgm',
+    create:
+      'CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_pages_title_trgm ' +
+      'ON pages USING gin ((LOWER(f_unaccent(title))) gin_trgm_ops)',
+  },
+  {
+    // #348 perf-indexes — users.name trigram (member search-suggest).
+    name: 'idx_users_name_trgm',
+    create:
+      'CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_users_name_trgm ' +
+      'ON users USING gin ((LOWER(f_unaccent(name))) gin_trgm_ops)',
+  },
+  {
+    // #443 search-lookup-trgm — pages.text_content trigram (the slow, large one).
+    name: 'idx_pages_text_content_trgm',
+    create:
+      'CREATE INDEX CONCURRENTLY IF NOT EXISTS idx_pages_text_content_trgm ' +
+      'ON pages USING gin ((LOWER(f_unaccent(text_content))) gin_trgm_ops)',
+  },
+];
+
+/**
+ * Idempotent, output-identical redefinition of `f_unaccent` to the INDEX-INLINABLE
+ * 1-arg form. MUST stay byte-for-byte in lockstep with migration
+ * `20260705T120000-perf-indexes.ts` (same signature + body): the trigram indexes
+ * above only build CONCURRENTLY once `f_unaccent(text)` inlines, and on an existing
+ * tenant it is still the old 2-arg form until that migration runs — which is AFTER
+ * this pre-build.
+ */
+const F_UNACCENT_REDEF = `
+    CREATE OR REPLACE FUNCTION f_unaccent(text)
+      RETURNS text
+      LANGUAGE sql
+      IMMUTABLE PARALLEL SAFE STRICT
+    AS $func$
+      SELECT public.unaccent($1);
+    $func$
+`;
+
+/**
+ * Best-effort, non-transactional pre-build of {@link CONCURRENT_INDEXES}. Run
+ * BEFORE the migrator so the blocking `CREATE INDEX` in the corresponding
+ * migration becomes an `IF NOT EXISTS` no-op.
+ *
+ * `db` MUST be the top-level Kysely instance (NOT a transaction): each statement
+ * then executes on its own connection with no surrounding `BEGIN`, which is
+ * required for `CONCURRENTLY`. Every statement is independent and swallowed on
+ * error: the migration backstop still builds the index, so a failure here is
+ * never fatal. `onLog` reports progress/failures for the caller to route to its
+ * logger.
+ *
+ * ORDER MATTERS: redefine `f_unaccent` to the inlinable form BEFORE the index
+ * loop — otherwise an existing tenant's old 2-arg `f_unaccent` makes every
+ * CONCURRENTLY build fail and the feature is a no-op (see the module docstring).
+ */
+export async function ensureConcurrentIndexes(
+  db: Kysely<any>,
+  onLog?: (message: string, error?: unknown) => void,
+): Promise<void> {
+  // Make f_unaccent inlinable FIRST. On a fresh DB (no `unaccent` extension yet)
+  // this throws harmlessly and is swallowed — the migration builds everything.
+  try {
+    await sql.raw(F_UNACCENT_REDEF).execute(db);
+    onLog?.('f_unaccent redefined to the index-inlinable 1-arg form');
+  } catch (error) {
+    onLog?.(
+      'f_unaccent redefine skipped (fresh DB / no unaccent extension) — ' +
+        'the migration will define it and build the indexes',
+      error,
+    );
+  }
+
+  for (const idx of CONCURRENT_INDEXES) {
+    try {
+      await sql.raw(idx.create).execute(db);
+      onLog?.(`Concurrent index ensured: ${idx.name}`);
+    } catch (error) {
+      // Non-fatal by design — the migration's IF NOT EXISTS create is the
+      // backstop. Benign on a fresh DB (the table/extension does not exist yet).
+      // NOT benign, but still swallowed, on an existing tenant whose f_unaccent
+      // could not be redefined above (then the migration builds the index NON-
+      // concurrently under the SHARE lock this feature meant to avoid).
+      onLog?.(
+        `Concurrent index pre-build skipped for ${idx.name} ` +
+          `(will fall back to the in-migration build)`,
+        error,
+      );
+    }
+  }
+}
@@ -0,0 +1,113 @@
+import { executeTx, registerAfterCommit } from './utils';
+import { KyselyDB, KyselyTransaction } from './types/kysely.types';
+
+// Post-commit hook contract (#495 item 13): a side effect registered via
+// registerAfterCommit must run ONLY AFTER the owning transaction commits, and a
+// hook registered against a passed-through existingTrx must fire at the OUTER
+// commit boundary — never inside the inner call. We fake the Kysely transaction
+// runner so the ordering is observable without a real DB.
+
+/**
+ * A minimal db whose `.transaction().execute(cb)` records the commit ORDER: it
+ * runs `cb(trx)`, pushes 'commit' onto `log` (simulating the real commit that
+ * happens after the callback resolves), then returns the callback's result.
+ */
+function fakeDb(log: string[]): { db: KyselyDB; trx: KyselyTransaction } {
+  const trx = { __fakeTrx: true } as unknown as KyselyTransaction;
+  const db = {
+    transaction: () => ({
+      execute: async (cb: (t: KyselyTransaction) => Promise<unknown>) => {
+        const result = await cb(trx);
+        log.push('commit');
+        return result;
+      },
+    }),
+  } as unknown as KyselyDB;
+  return { db, trx };
+}
+
+describe('executeTx post-commit hooks', () => {
+  it('runs an afterCommit hook only AFTER the transaction commits', async () => {
+    const log: string[] = [];
+    const { db } = fakeDb(log);
+
+    await executeTx(db, async (trx) => {
+      log.push('body');
+      registerAfterCommit(trx, () => {
+        log.push('hook');
+      });
+      // The hook must NOT have run yet — the tx is still open.
+      expect(log).toEqual(['body']);
+    });
+
+    // Order proves post-commit: body → commit → hook (never body → hook → commit).
+    expect(log).toEqual(['body', 'commit', 'hook']);
+  });
+
+  it('drains hooks registered against a passed-through existingTrx at the OUTER commit', async () => {
+    const log: string[] = [];
+    const { db, trx: outerTrx } = fakeDb(log);
+
+    await executeTx(db, async (outer) => {
+      // Nested executeTx reuses the outer trx: it must NOT commit or drain now.
+      await executeTx(
+        db,
+        async (inner) => {
+          registerAfterCommit(inner, () => {
+            log.push('inner-hook');
+          });
+        },
+        outer,
+      );
+      // Still inside the outer tx — the inner hook has not fired.
+      expect(log).toEqual([]);
+    });
+
+    // The single (outer) commit drains the hook registered on the shared trx.
+    expect(log).toEqual(['commit', 'inner-hook']);
+    // Sanity: the trx the hooks were registered against is the outer one.
+    expect(outerTrx).toBeDefined();
+  });
+
+  it('a hook failure does not reject the already-committed executeTx', async () => {
+    const log: string[] = [];
+    const { db } = fakeDb(log);
+
+    await expect(
+      executeTx(db, async (trx) => {
+        registerAfterCommit(trx, () => {
+          throw new Error('cache del blew up');
+        });
+        registerAfterCommit(trx, () => {
+          log.push('second-hook-still-runs');
+        });
+        return 'ok';
+      }),
+    ).resolves.toBe('ok');
+
+    // The throwing hook is swallowed; a later hook still runs.
+    expect(log).toEqual(['commit', 'second-hook-still-runs']);
+  });
+
+  it('does NOT run afterCommit hooks when the transaction body throws (rollback)', async () => {
+    // The body rejects -> the fake transaction never pushes 'commit' and
+    // db.transaction().execute() rejects, mirroring a real rolled-back tx. The
+    // drain runs only AFTER the awaited (committed) transaction, so a rollback
+    // must leave every registered hook UN-run — otherwise a cache-bust / event
+    // would fire for a write that never landed.
+    const log: string[] = [];
+    const { db } = fakeDb(log);
+    const hook = jest.fn();
+
+    await expect(
+      executeTx(db, async (trx) => {
+        registerAfterCommit(trx, hook);
+        throw new Error('write failed -> rollback');
+      }),
+    ).rejects.toThrow('write failed -> rollback');
+
+    // No commit happened, and the post-commit hook never ran.
+    expect(log).toEqual([]); // no 'commit'
+    expect(hook).not.toHaveBeenCalled();
+  });
+});
@@ -33,16 +33,18 @@ import { type Kysely, sql } from 'kysely';
 *    - comments: `findPageComments` does WHERE page_id ORDER BY id ASC, but only
 *      `(page_id)` exists → extra sort.
 *
- * DEPLOY-TIME LOCK WARNING: these are plain (non-CONCURRENT) CREATE INDEX
- * statements — CONCURRENTLY is impossible because Kysely runs each migration in a
- * transaction. They take a SHARE lock that BLOCKS writes (INSERT/UPDATE/DELETE) on
- * pages/users/groups/comments/page_history for the duration of the build. The two
- * GIN trigram builds on pages.title / users.name are the slow ones and can take
- * minutes on a large tenant → a write-outage window during the deploy migration.
- * For large installations, run this migration in a maintenance window, or build
- * the trigram indexes out-of-band with CREATE INDEX CONCURRENTLY before deploying
- * (then this migration's `IF NOT EXISTS` is a no-op). Small/typical tenants are
- * unaffected.
+ * DEPLOY-TIME LOCK: these are plain (non-CONCURRENT) CREATE INDEX statements —
+ * CONCURRENTLY is impossible HERE because Kysely runs each migration in a
+ * transaction. The two GIN trigram builds on pages.title / users.name are the
+ * slow ones and would take a SHARE lock that BLOCKS writes on pages/users for
+ * minutes on a large tenant. To avoid that, `ensureConcurrentIndexes`
+ * (database/concurrent-indexes.ts) now pre-builds BOTH trigram indexes with
+ * CREATE INDEX CONCURRENTLY (no transaction) BEFORE the migrator runs, so on an
+ * existing DB the two `IF NOT EXISTS` trigram creates below no-op and no write
+ * lock is taken. On a fresh DB the pre-build is skipped and they build on an
+ * empty table. Keep the two trigram creates in lockstep with their CANONICAL
+ * definitions in CONCURRENT_INDEXES. (The plain b-tree indexes further down are
+ * fast metadata-only builds; they are not pre-built.)
 */
 export async function up(db: Kysely<any>): Promise<void> {
  // Index-compatible, output-identical redefinition of f_unaccent (see header).
@@ -26,13 +26,16 @@ import { type Kysely, sql } from 'kysely';
 *    to update than b-trees); on the small instances this fork targets that cost
 *    is acceptable and the read win on agent lookups is the priority.
 *
- * DEPLOY-TIME LOCK WARNING: plain (non-CONCURRENT) CREATE INDEX — Kysely runs
- * each migration in a transaction, so CONCURRENTLY is impossible. The build takes
- * a SHARE lock that BLOCKS writes on `pages` for its duration. The text_content
- * GIN build is the slow one and can take minutes on a large tenant. For large
- * installations, run this in a maintenance window or build the index out-of-band
- * with CREATE INDEX CONCURRENTLY before deploying (then `IF NOT EXISTS` no-ops
- * here). Small/typical tenants are unaffected.
+ * DEPLOY-TIME LOCK: this is a plain (non-CONCURRENT) CREATE INDEX — Kysely runs
+ * each migration in a transaction, so CONCURRENTLY is impossible HERE, and the
+ * build would take a SHARE lock that BLOCKS writes on `pages` for its duration
+ * (the text_content GIN build can take minutes on a large tenant). To avoid that,
+ * `ensureConcurrentIndexes` (database/concurrent-indexes.ts) now pre-builds this
+ * index with CREATE INDEX CONCURRENTLY (no transaction) BEFORE the migrator runs,
+ * so on an existing DB the `IF NOT EXISTS` below no-ops and no write lock is taken.
+ * On a fresh DB the pre-build is skipped and this builds it on an empty table
+ * where the lock is irrelevant. Keep this create in lockstep with the CANONICAL
+ * definition in CONCURRENT_INDEXES (same expression + opclass).
 */
 export async function up(db: Kysely<any>): Promise<void> {
  // The title predicate is served by #348's idx_pages_title_trgm — see header.
@@ -1,7 +1,7 @@
 import { Injectable } from '@nestjs/common';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
-import { dbOrTx, executeTx } from '../../utils';
+import { dbOrTx, executeTx, registerAfterCommit } from '../../utils';
 import {
  InsertablePage,
  Page,
@@ -349,14 +349,23 @@ export class PageRepo {
    pageId: string,
    deletedById: string,
    workspaceId: string,
+    // Optional caller transaction. When passed, the reads + soft-delete run in
+    // THAT transaction (so a caller holding a `FOR UPDATE` lock on the row — e.g.
+    // the temporary-note sweeper — can delete under the lock without deadlocking
+    // on a nested independent transaction) and the PAGE_SOFT_DELETED broadcast is
+    // deferred to the caller's COMMIT via registerAfterCommit (so a rolled-back
+    // delete never broadcasts). With no trx the behaviour is unchanged: own
+    // transaction, broadcast right after it commits.
+    existingTrx?: KyselyTransaction,
  ): Promise<void> {
    const currentDate = new Date();
+    const readDb = dbOrTx(this.db, existingTrx);

    // Read the root snapshot up front so PAGE_SOFT_DELETED can carry it without
    // a post-commit DB read (variant A). Only the root of the deleted subtree is
    // needed for the tree broadcast — the client `treeModel.remove` drops all
    // descendants, so we don't snapshot/broadcast every descendant.
-    const rootSnapshot = await this.db
+    const rootSnapshot = await readDb
      .selectFrom('pages')
      .select([
        'id',
@@ -371,7 +380,7 @@ export class PageRepo {
      .where('deletedAt', 'is', null)
      .executeTakeFirst();

-    const descendants = await this.db
+    const descendants = await readDb
      .withRecursive('page_descendants', (db) =>
        db
          .selectFrom('pages')
@@ -393,39 +402,60 @@ export class PageRepo {
    const pageIds = descendants.map((d) => d.id);

    if (pageIds.length > 0) {
-      await executeTx(this.db, async (trx) => {
-        await trx
-          .updateTable('pages')
-          .set({
-            deletedById: deletedById,
-            deletedAt: currentDate,
-          })
-          .where('id', 'in', pageIds)
-          .where('deletedAt', 'is', null)
-          .execute();
+      // Reuse the caller's transaction when given (executeTx passes it straight
+      // through), else own a fresh one.
+      await executeTx(
+        this.db,
+        async (trx) => {
+          await trx
+            .updateTable('pages')
+            .set({
+              deletedById: deletedById,
+              deletedAt: currentDate,
+            })
+            .where('id', 'in', pageIds)
+            .where('deletedAt', 'is', null)
+            .execute();

-        await trx.deleteFrom('shares').where('pageId', 'in', pageIds).execute();
-      });
+          await trx
+            .deleteFrom('shares')
+            .where('pageId', 'in', pageIds)
+            .execute();
+        },
+        existingTrx,
+      );

-      this.eventEmitter.emit(EventName.PAGE_SOFT_DELETED, {
-        pageIds: pageIds,
-        workspaceId,
-        // Root-only snapshot: one `deleteTreeNode` is enough, the client removes
-        // the whole subtree. Skip if the root vanished between the two reads.
-        pages: rootSnapshot
-          ? [
-              {
-                id: rootSnapshot.id,
-                slugId: rootSnapshot.slugId,
-                title: rootSnapshot.title,
-                icon: rootSnapshot.icon,
-                position: rootSnapshot.position,
-                spaceId: rootSnapshot.spaceId,
-                parentPageId: rootSnapshot.parentPageId,
-              },
-            ]
-          : [],
-      });
+      const emitSoftDeleted = () => {
+        this.eventEmitter.emit(EventName.PAGE_SOFT_DELETED, {
+          pageIds: pageIds,
+          workspaceId,
+          // Root-only snapshot: one `deleteTreeNode` is enough, the client
+          // removes the whole subtree. Skip if the root vanished between reads.
+          pages: rootSnapshot
+            ? [
+                {
+                  id: rootSnapshot.id,
+                  slugId: rootSnapshot.slugId,
+                  title: rootSnapshot.title,
+                  icon: rootSnapshot.icon,
+                  position: rootSnapshot.position,
+                  spaceId: rootSnapshot.spaceId,
+                  parentPageId: rootSnapshot.parentPageId,
+                },
+              ]
+            : [],
+        });
+      };
+
+      if (existingTrx) {
+        // Inside a caller transaction: the delete above is NOT committed yet.
+        // Defer the tree broadcast to the caller's commit so a rolled-back delete
+        // never broadcasts a phantom removal.
+        registerAfterCommit(existingTrx, emitSoftDeleted);
+      } else {
+        // Own transaction already committed above — broadcast now.
+        emitSoftDeleted();
+      }
    }
  }

@@ -3,7 +3,7 @@ import { CACHE_MANAGER } from '@nestjs/cache-manager';
 import { Cache } from 'cache-manager';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
-import { dbOrTx } from '../../utils';
+import { dbOrTx, registerAfterCommit } from '../../utils';
 import {
  InsertableWorkspace,
  UpdatableWorkspace,
@@ -80,16 +80,30 @@ export class WorkspaceRepo {
   */
  private async bustWorkspaceCache(
    workspace?: Pick<Workspace, 'hostname'> | undefined,
+    trx?: KyselyTransaction,
  ): Promise<void> {
-    try {
-      await this.cacheManager.del(CacheKey.WORKSPACE_SELF_HOSTED);
-      if (workspace?.hostname) {
-        await this.cacheManager.del(
-          CacheKey.WORKSPACE_BY_HOST(workspace.hostname),
-        );
+    const del = async () => {
+      try {
+        await this.cacheManager.del(CacheKey.WORKSPACE_SELF_HOSTED);
+        if (workspace?.hostname) {
+          await this.cacheManager.del(
+            CacheKey.WORKSPACE_BY_HOST(workspace.hostname),
+          );
+        }
+      } catch {
+        // cache is best-effort; TTL is the backstop
      }
-    } catch {
-      // cache is best-effort; TTL is the backstop
+    };
+    if (trx) {
+      // Inside a caller transaction the write is NOT yet committed: busting now
+      // opens a repopulation window (a concurrent reader reloads the cache with
+      // the pre-commit / stale row, which then survives until TTL). Defer the del
+      // to the transaction's commit (drained by the owning executeTx) (#495).
+      registerAfterCommit(trx, del);
+    } else {
+      // No transaction: the mutation above already auto-committed, so this del is
+      // already post-commit.
+      await del();
    }
  }

@@ -180,7 +194,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -195,7 +209,7 @@ export class WorkspaceRepo {
      .returning(this.baseFields)
      .executeTakeFirst();
    // Bust the cached "not found" so a fresh install / new tenant is seen at once.
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -249,7 +263,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -271,7 +285,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -326,7 +340,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -354,7 +368,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -376,7 +390,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -398,7 +412,7 @@ export class WorkspaceRepo {
      .where('id', '=', workspaceId)
      .returning(this.baseFields)
      .executeTakeFirst();
-    await this.bustWorkspaceCache(workspace);
+    await this.bustWorkspaceCache(workspace, trx);
    return workspace;
  }

@@ -4,6 +4,7 @@ import { promises as fs } from 'fs';
 import { Migrator, FileMigrationProvider } from 'kysely';
 import { InjectKysely } from 'nestjs-kysely';
 import { KyselyDB } from '@docmost/db/types/kysely.types';
+import { ensureConcurrentIndexes } from '@docmost/db/concurrent-indexes';

@Injectable()
 export class MigrationService {
@@ -12,6 +13,16 @@ export class MigrationService {
  constructor(@InjectKysely() private readonly db: KyselyDB) {}

  async migrateToLatest(): Promise<void> {
+    // Build write-blocking trigram indexes CONCURRENTLY (no transaction) BEFORE
+    // the migrator runs, so the corresponding in-migration `CREATE INDEX IF NOT
+    // EXISTS` no-ops instead of taking a SHARE lock on `pages` during deploy
+    // (#495). Best-effort: on a fresh DB (no `pages`/`f_unaccent` yet) this is a
+    // no-op and the migrations build the index normally.
+    await ensureConcurrentIndexes(this.db, (message, error) => {
+      if (error) this.logger.warn(`${message}: ${String(error)}`);
+      else this.logger.log(message);
+    });
+
    const migrator = new Migrator({
      db: this.db,
      provider: new FileMigrationProvider({
@@ -6,16 +6,76 @@ import { KyselyDB, KyselyTransaction } from './types/kysely.types';
 * If an existing transaction is provided, it directly executes the callback with it.
 * Otherwise, it starts a new transaction using the provided database instance and executes the callback within that transaction.
 */
+/**
+ * Post-commit side-effect hooks, keyed by the transaction they were registered
+ * against. A WeakMap so an abandoned/never-drained transaction's entry is GC'd
+ * with the trx object (no leak). Used by {@link registerAfterCommit} /
+ * {@link executeTx}.
+ */
+const afterCommitHooks = new WeakMap<
+  KyselyTransaction,
+  Array<() => Promise<void> | void>
+>();
+
+/**
+ * Register a side effect to run ONLY AFTER the transaction that owns `trx`
+ * commits. THE fix for "bust the cache inside the open transaction" bugs: a
+ * cache-invalidation (or any read-your-write-visible side effect) done while the
+ * writing transaction is still open opens a window where a concurrent reader
+ * repopulates the cache with the PRE-COMMIT (stale) row, so after commit the
+ * cache holds the old value until its TTL. Deferring the effect to post-commit
+ * closes that window.
+ *
+ * The hook is drained by the OUTERMOST {@link executeTx} that actually owns
+ * (created) this transaction — so registering against a passed-through
+ * `existingTrx` still fires at the real commit boundary, not at the inner call.
+ * NOTE: a hook registered against a transaction that was NOT created via
+ * `executeTx` (untracked) will never be drained — always create transactions
+ * through `executeTx` when you rely on post-commit hooks.
+ */
+export function registerAfterCommit(
+  trx: KyselyTransaction,
+  hook: () => Promise<void> | void,
+): void {
+  const existing = afterCommitHooks.get(trx);
+  if (existing) existing.push(hook);
+  else afterCommitHooks.set(trx, [hook]);
+}
+
 export async function executeTx<T>(
  db: KyselyDB,
  callback: (trx: KyselyTransaction) => Promise<T>,
  existingTrx?: KyselyTransaction,
 ): Promise<T> {
  if (existingTrx) {
-    return await callback(existingTrx); // Execute callback with existing transaction
-  } else {
-    return await db.transaction().execute((trx) => callback(trx)); // Start new transaction and execute callback
+    // Reuse the caller's transaction. Any post-commit hooks registered here are
+    // drained by the OUTER executeTx that created `existingTrx`, at the true
+    // commit boundary — so we must NOT drain them now.
+    return await callback(existingTrx);
  }
+  // We OWN this transaction: run the body, then (only once it has COMMITTED)
+  // drain the post-commit hooks registered against it during the body.
+  let ownTrx: KyselyTransaction | undefined;
+  const result = await db.transaction().execute((trx) => {
+    ownTrx = trx;
+    return callback(trx);
+  });
+  if (ownTrx) {
+    const hooks = afterCommitHooks.get(ownTrx);
+    if (hooks) {
+      afterCommitHooks.delete(ownTrx);
+      for (const hook of hooks) {
+        // Best-effort: a failed side effect (e.g. a cache del) must not fail the
+        // already-committed transaction.
+        try {
+          await hook();
+        } catch {
+          // swallow — the durable write already committed
+        }
+      }
+    }
+  }
+  return result;
 }

 /*
@@ -99,10 +99,12 @@ describe('AiSettingsService.getMasked reindex progress', () => {
    // actually pins the progress.total branch rather than coincidentally
    // matching the DB fallback. With fix #1 the two sources agree in practice,
    // but getMasked must still return progress.total when a record is active.
+    const startedAt = Date.now();
    reindexProgress.get.mockResolvedValue({
      total: 500,
      done: 120,
-      startedAt: Date.now(),
+      startedAt,
+      runId: 'run-abc',
    });

    const masked = await service.getMasked(WORKSPACE_ID);
@@ -110,6 +112,10 @@ describe('AiSettingsService.getMasked reindex progress', () => {
    expect(masked.indexedPages).toBe(120); // progress.done, not DB 478
    expect(masked.totalPages).toBe(500); // progress.total, not DB 478
    expect(masked.reindexing).toBe(true);
+    // The status payload must carry the run identity so the client can key its
+    // poll on it (a changed runId => a NEW run).
+    expect(masked.runId).toBe('run-abc');
+    expect(masked.reindexStartedAt).toBe(startedAt);
  });

  it('falls back to countIndexedPages when no reindex is active', async () => {
@@ -121,6 +127,10 @@ describe('AiSettingsService.getMasked reindex progress', () => {
    expect(masked.indexedPages).toBe(478);
    expect(masked.totalPages).toBe(478);
    expect(masked.reindexing).toBe(false);
+    // No active run -> no run identity surfaced (the client keeps its prior
+    // steady-state behaviour).
+    expect(masked.runId).toBeUndefined();
+    expect(masked.reindexStartedAt).toBeUndefined();
  });
 });

@@ -371,6 +371,12 @@ export class AiSettingsService {
      totalPages,
      // Optional hint for the client: a reindex run is currently in progress.
      reindexing: progress != null,
+      // Per-run identity so the client can key its poll on a stable run id and
+      // reset its per-run state when a NEW run starts. Present only while a run
+      // is active; `runId` may be '' for a legacy/degraded record (the client
+      // treats that as "no identity").
+      runId: progress?.runId,
+      reindexStartedAt: progress?.startedAt,
    };
  }

@@ -0,0 +1,86 @@
+// `.provider` alone cannot prove the gemini/ollama chat factories were built
+// with the instrumented streaming fetch — a regression dropping it (which drops
+// them back to the global undici fetch: no keep-alive recycle, no reset retries,
+// unbounded silence timeout; incident classes #140/#175/#310) would still pass.
+// So mock the factories and assert the exact fetch argument. jest.mock is
+// module-scoped, hence a dedicated file.
+
+const mockGeminiModel = { provider: 'google.generative-ai', modelId: 'm' };
+const mockOllamaModel = { provider: 'ollama.chat', modelId: 'm' };
+
+// jest allows `mock`-prefixed vars inside a jest.mock factory.
+const mockCreateGoogle = jest.fn((_settings: unknown) => () => mockGeminiModel);
+const mockCreateOllama = jest.fn((_settings: unknown) => () => mockOllamaModel);
+
+jest.mock('@ai-sdk/google', () => ({
+  createGoogleGenerativeAI: (settings: unknown) => mockCreateGoogle(settings),
+}));
+jest.mock('ai-sdk-ollama', () => ({
+  createOllama: (settings: unknown) => mockCreateOllama(settings),
+}));
+
+import { AiService } from './ai.service';
+
+describe('AiService.getChatModel provider transport fetch (gemini/ollama)', () => {
+  function serviceWith(cfg: Record<string, unknown>) {
+    const aiSettings = {
+      resolve: jest.fn().mockResolvedValue(cfg),
+    };
+    return new AiService(
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      aiSettings as any,
+      { find: jest.fn() } as never,
+      { decryptSecret: jest.fn() } as never,
+    );
+  }
+
+  beforeEach(() => {
+    mockCreateGoogle.mockClear();
+    mockCreateOllama.mockClear();
+  });
+
+  it('builds the gemini chat model with the instrumented streaming fetch', async () => {
+    await serviceWith({
+      driver: 'gemini',
+      chatModel: 'gemini-2.5-pro',
+      apiKey: 'the-key',
+    }).getChatModel('ws-1');
+    expect(mockCreateGoogle).toHaveBeenCalledTimes(1);
+    expect(mockCreateGoogle).toHaveBeenCalledWith(
+      expect.objectContaining({
+        apiKey: 'the-key',
+        fetch: expect.any(Function),
+      }),
+    );
+  });
+
+  it('builds the ollama chat model with the instrumented streaming fetch', async () => {
+    await serviceWith({
+      driver: 'ollama',
+      chatModel: 'llama3',
+      baseUrl: 'http://localhost:11434/api',
+    }).getChatModel('ws-1');
+    expect(mockCreateOllama).toHaveBeenCalledTimes(1);
+    expect(mockCreateOllama).toHaveBeenCalledWith(
+      expect.objectContaining({
+        baseURL: 'http://localhost:11434/api',
+        fetch: expect.any(Function),
+      }),
+    );
+  });
+
+  it('reuses ONE service-lifetime fetch instance across both providers', async () => {
+    const svc = serviceWith({
+      driver: 'gemini',
+      chatModel: 'gemini-2.5-pro',
+      apiKey: 'k',
+    });
+    await svc.getChatModel('ws-1');
+    const geminiFetch = mockCreateGoogle.mock.calls[0][0] as { fetch: unknown };
+    // Same instance on a second call — the fetch is held for the service
+    // lifetime to reuse the streaming dispatcher's connection pool.
+    await svc.getChatModel('ws-1');
+    const geminiFetch2 = mockCreateGoogle.mock.calls[1][0] as { fetch: unknown };
+    expect(geminiFetch.fetch).toBe(geminiFetch2.fetch);
+  });
+});
@@ -190,10 +190,22 @@ export class AiService {
        }).chat(chatModel);
      }
      case 'gemini':
-        return createGoogleGenerativeAI({ apiKey })(chatModel);
+        // Route gemini through the same instrumented streaming fetch as openai
+        // (finite silence timeouts + keep-alive recycling + pre-response
+        // connection-reset retry). Without it the provider ran on the global
+        // undici fetch — no keep-alive recycle, no reset retries, default
+        // (unbounded silence) timeout — so incident classes #140/#175/#310 were
+        // reproducible for gemini too.
+        return createGoogleGenerativeAI({
+          apiKey,
+          fetch: this.aiProviderFetch,
+        })(chatModel);
      case 'ollama':
-        // Ollama needs no API key.
-        return createOllama({ baseURL: baseUrl })(chatModel);
+        // Ollama needs no API key. Same transport hardening as above (#140/#175/#310).
+        return createOllama({
+          baseURL: baseUrl,
+          fetch: this.aiProviderFetch,
+        })(chatModel);
      default:
        throw new AiNotConfiguredException();
    }
@@ -149,4 +149,14 @@ export interface MaskedAiSettings {
  // True while a full workspace reindex is actively running (the counts above
  // then reflect the live run progress rather than the steady-state DB count).
  reindexing?: boolean;
+  // Identity of the ACTIVE reindex run (present only while `reindexing`). The
+  // client keys its poll on `runId`: a changed value means a NEW run (reset the
+  // per-run poll state it latched), the same value means the run it is already
+  // watching — removing the "same run or a fresh one?" ambiguity a stale
+  // pre-reindex snapshot otherwise causes. Absent/empty degrades gracefully.
+  runId?: string;
+  // Epoch-ms the active run started (present only while `reindexing`). Paired
+  // with `runId` so a run that restarts with the same (recycled) id is still
+  // seen as new.
+  reindexStartedAt?: number;
 }
@@ -48,19 +48,38 @@ describe('EmbeddingReindexProgressService', () => {
  }

  describe('get', () => {
-    it('maps a valid hash to a ReindexProgress object', async () => {
+    it('maps a valid hash to a ReindexProgress object (incl. the run identity)', async () => {
      const { redis, hgetall } = makeRedis();
-      hgetall.mockResolvedValue({ total: '478', done: '120', startedAt: '1000' });
+      hgetall.mockResolvedValue({
+        total: '478',
+        done: '120',
+        startedAt: '1000',
+        runId: 'run-xyz',
+      });
      const service = makeService(redis);

      await expect(service.get(WORKSPACE_ID)).resolves.toEqual({
        total: 478,
        done: 120,
        startedAt: 1000,
+        runId: 'run-xyz',
      });
      expect(hgetall).toHaveBeenCalledWith(KEY);
    });

+    it('degrades a missing runId to an empty string (legacy/partial record)', async () => {
+      const { redis, hgetall } = makeRedis();
+      // A record written before runId existed: get() must still succeed and
+      // report runId='' so the client treats it as "no identity", never breaks.
+      hgetall.mockResolvedValue({ total: '10', done: '3', startedAt: '5' });
+      await expect(makeService(redis).get(WORKSPACE_ID)).resolves.toEqual({
+        total: 10,
+        done: 3,
+        startedAt: 5,
+        runId: '',
+      });
+    });
+
    it('returns null for an empty hash (no record)', async () => {
      const { redis, hgetall } = makeRedis();
      hgetall.mockResolvedValue({});
@@ -87,11 +106,17 @@ describe('EmbeddingReindexProgressService', () => {

    it('coerces a non-finite startedAt to 0', async () => {
      const { redis, hgetall } = makeRedis();
-      hgetall.mockResolvedValue({ total: '10', done: '2', startedAt: 'nope' });
+      hgetall.mockResolvedValue({
+        total: '10',
+        done: '2',
+        startedAt: 'nope',
+        runId: 'run-1',
+      });
      await expect(makeService(redis).get(WORKSPACE_ID)).resolves.toEqual({
        total: 10,
        done: 2,
        startedAt: 0,
+        runId: 'run-1',
      });
    });

@@ -115,6 +140,21 @@ describe('EmbeddingReindexProgressService', () => {
      expect(multiObj.exec).toHaveBeenCalledTimes(1);
    });

+    it('mints a fresh non-empty runId into the record on each start', async () => {
+      const { redis, multiObj } = makeRedis();
+      const service = makeService(redis);
+      await service.start(WORKSPACE_ID, 1);
+      await service.start(WORKSPACE_ID, 1);
+
+      const firstRunId = multiObj.hset.mock.calls[0][1].runId;
+      const secondRunId = multiObj.hset.mock.calls[1][1].runId;
+      expect(typeof firstRunId).toBe('string');
+      expect(firstRunId).not.toBe('');
+      // Each run gets its OWN identity so the client can tell a re-trigger apart
+      // from the run it is already watching.
+      expect(secondRunId).not.toBe(firstRunId);
+    });
+
    it('defaults the expire TTL to the full 1h record TTL', async () => {
      const { redis, multiObj } = makeRedis();
      await makeService(redis).start(WORKSPACE_ID, 478);
@@ -1,17 +1,28 @@
 import { Injectable, Logger } from '@nestjs/common';
 import { RedisService } from '@nestjs-labs/nestjs-ioredis';
+import { randomUUID } from 'node:crypto';
 import type { Redis } from 'ioredis';

 /**
 * Live progress of an in-flight workspace embeddings reindex run.
 * `total` is the number of pages the run will process, `done` how many it has
 * already processed (success OR handled failure), `startedAt` the epoch-ms the
- * record was created.
+ * record was created, and `runId` a per-run identity minted at `start()`.
+ *
+ * `runId` gives each reindex run a stable identity so a poller can tell "same
+ * run I've been watching" from "a NEW run started" WITHOUT guessing from the
+ * progress counters (the ambiguity that a stale pre-reindex snapshot vs a fresh
+ * run otherwise causes — the bug class fixed twice under #262). It is best-
+ * effort like the rest of this record: a record written before this field
+ * existed (or a Redis hiccup) yields an empty `runId`, which the client must
+ * treat as "no identity available" and degrade to its prior behaviour, never
+ * break.
 */
 export interface ReindexProgress {
  total: number;
  done: number;
  startedAt: number;
+  runId: string;
 }

 /** Redis key namespace for the per-workspace reindex-progress record. */
@@ -86,12 +97,18 @@ export class EmbeddingReindexProgressService {
  ): Promise<void> {
    const key = this.key(workspaceId);
    try {
+      // A fresh identity per run so the client poll can key on it: a changed
+      // runId means a genuinely NEW run (reset any latched per-run poll state),
+      // the same runId means the run the client is already watching. Best-effort
+      // like the counters — never surfaced to the user, only used to disambiguate.
+      const runId = randomUUID();
      await this.redis
        .multi()
        .hset(key, {
          total: String(total),
          done: '0',
          startedAt: String(Date.now()),
+          runId,
        })
        .expire(key, ttlSeconds)
        .exec();
@@ -150,7 +167,15 @@ export class EmbeddingReindexProgressService {
      const done = Number(data.done);
      const startedAt = Number(data.startedAt);
      if (!Number.isFinite(total) || !Number.isFinite(done)) return null;
-      return { total, done, startedAt: Number.isFinite(startedAt) ? startedAt : 0 };
+      // `runId` degrades gracefully: a pre-existing record (written before this
+      // field) or a stripped value reads as '' — the client treats that as "no
+      // identity" and keeps its prior behaviour rather than breaking the poll.
+      return {
+        total,
+        done,
+        startedAt: Number.isFinite(startedAt) ? startedAt : 0,
+        runId: typeof data.runId === 'string' ? data.runId : '',
+      };
    } catch (err) {
      this.logger.warn(
        `reindex-progress read failed for workspace ${workspaceId}; ` +
@@ -33,6 +33,11 @@ describe('share_aliases one-per-page invariant [integration]', () => {
  const pageRepo = {
    findById: async (id: string) => ({ id, title: `title-${id}` }),
  };
+  // The requester can view the target page (permissive), so the reassign 409 may
+  // include its title — these tests exercise the one-per-page invariant, not the
+  // #495 disclosure gate (that is unit-tested in share-alias.service.spec.ts).
+  const pageAccessService = { validateCanView: async () => {} };
+  const USER = { id: 'u-int' } as any;

  beforeAll(async () => {
    db = getTestDb();
@@ -41,6 +46,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      repo as any,
      pageRepo as any,
      {} as any, // shareService — unused by setAlias
+      pageAccessService as any,
      db as any,
    );
    wsId = (await createWorkspace(db)).id;
@@ -188,6 +194,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId,
      creatorId,
+      user: USER,
      alias: 'te',
    });
    expect(first.alias).toBe('te');
@@ -196,6 +203,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId,
      creatorId,
+      user: USER,
      alias: 'ted',
    });
    // Same row id — a RENAME, not a new insert.
@@ -217,12 +225,14 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId,
      creatorId: null as any,
+      user: USER,
      alias: 'hello',
    });
    const again = await service.setAlias({
      workspaceId: wsId,
      pageId,
      creatorId: null as any,
+      user: USER,
      alias: 'hello',
    });
    expect(again.id).toBe(inserted.id);
@@ -244,6 +254,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      flakyRepo as any,
      pageRepo as any,
      {} as any,
+      pageAccessService as any,
      db as any,
    );

@@ -252,6 +263,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
        workspaceId: wsId,
        pageId,
        creatorId: null as any,
+        user: USER,
        alias: 'rollback-me',
      }),
    ).rejects.toBeInstanceOf(BadRequestException);
@@ -275,6 +287,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId: pageA,
      creatorId: null as any,
+      user: USER,
      alias: 'shared',
    });

@@ -283,6 +296,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
        workspaceId: wsId,
        pageId: pageB,
        creatorId: null as any,
+        user: USER,
        alias: 'shared',
      }),
    ).rejects.toBeInstanceOf(ConflictException);
@@ -291,6 +305,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId: pageB,
      creatorId: null as any,
+      user: USER,
      alias: 'shared',
      confirmReassign: true,
    });
@@ -317,12 +332,14 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId: pageA,
      creatorId: null as any,
+      user: USER,
      alias: 'shared-target',
    });
    await service.setAlias({
      workspaceId: wsId,
      pageId: pageB,
      creatorId: null as any,
+      user: USER,
      alias: 'bee',
    });

@@ -330,6 +347,7 @@ describe('share_aliases one-per-page invariant [integration]', () => {
      workspaceId: wsId,
      pageId: pageB,
      creatorId: null as any,
+      user: USER,
      alias: 'shared-target',
      confirmReassign: true,
    });
@@ -33,6 +33,12 @@ import { TransformsMixin, type ITransformsMixin } from "./client/transforms.js";
 export type { DocmostMcpConfig, SandboxPut } from "./client/context.js";
 export { formatDocmostAxiosError, assertFullUuid } from "./client/errors.js";

+// Branded canonical page-identity type (#435): the internal page UUID is a
+// distinct nominal type so an unresolved raw/slug string can't be swapped into
+// the seams that require the canonical id (see lib/page-id.ts). Re-exported on
+// the package surface for hosts that type against the resolved id.
+export type { PageId } from "./lib/page-id.js";
+
 // The full public + shared instance surface of the assembled client. Built by
 // INTERSECTING each domain mixin's public interface (each DERIVED from its class
 // and enforced by that class's `implements` clause — issue #446, no hand-mirror)
@@ -24,6 +24,7 @@ import {
  isCollabAuthFailedError,
 } from "../lib/collab-session.js";
 import { withPageLock, isUuid } from "../lib/page-lock.js";
+import type { PageId } from "../lib/page-id.js";
 import { getCollabToken, performLogin } from "../lib/auth-utils.js";
 import { formatDocmostAxiosError } from "./errors.js";
 import { GetPageConversionCache } from "./getpage-cache.js";
@@ -715,10 +716,18 @@ export abstract class DocmostClientContext {
   * once via getPageRaw and cached (both slugId->uuid and uuid->uuid), so
   * repeated edits on the same page add no extra request.
   */
-  protected async resolvePageId(pageId: string): Promise<string> {
-    if (isUuid(pageId)) return pageId;
+  protected async resolvePageId(pageId: string): Promise<PageId> {
+    // This is the ONE canonicalization seam, so it is where the `PageId` brand
+    // is minted (#435). The value is validated here — a UUID input by isUuid, a
+    // resolved id as the server's own page.id — so the downstream write path
+    // (withPageLock / mutatePageContent) can require the brand and reject any
+    // unresolved raw id at compile time. The brand is a pure compile-time marker
+    // applied by cast (no runtime guard): the guarantee is that this seam is the
+    // only place a `PageId` is produced, so every branded value went through the
+    // UUID/resolve check above.
+    if (isUuid(pageId)) return pageId as PageId;
    const cached = this.pageIdCache.get(pageId);
-    if (cached) return cached;
+    if (cached) return cached as PageId;
    const data = await this.getPageRaw(pageId);
    const uuid = data?.id;
    if (typeof uuid !== "string" || !uuid) {
@@ -727,7 +736,7 @@ export abstract class DocmostClientContext {
      );
    }
    this.pageIdCache.set(pageId, uuid);
-    return uuid;
+    return uuid as PageId;
  }


@@ -13,6 +13,7 @@ import { JSDOM } from "jsdom";
 import { markdownToProseMirror } from "@docmost/prosemirror-markdown";
 import { docmostExtensions, docmostSchema } from "./docmost-schema.js";
 import { withPageLock } from "./page-lock.js";
+import type { PageId } from "./page-id.js";
 import {
  sanitizeForYjs,
  findUnstorableAttr,
@@ -250,7 +251,10 @@ export function assertYjsEncodable(doc: any): void {
 * read->write window, and it never throws (it can NEVER break a write).
 */
 export async function mutatePageContent(
-  pageId: string,
+  // Canonical UUID only (#260/#435): the brand forces every caller to
+  // resolvePageId() BEFORE this seam so the lock + CollabSession key can never
+  // be a raw slugId.
+  pageId: PageId,
  collabToken: string,
  baseUrl: string,
  transform: (liveDoc: any) => any | null,
@@ -300,7 +304,7 @@ export async function mutatePageContent(
 * mutatePageContent.
 */
 export async function replacePageContent(
-  pageId: string,
+  pageId: PageId,
  prosemirrorDoc: any,
  collabToken: string,
  baseUrl: string,
@@ -332,7 +336,7 @@ export async function replacePageContent(
 * Tables and :::callout::: blocks survive thanks to the full schema.
 */
 export async function updatePageContentRealtime(
-  pageId: string,
+  pageId: PageId,
  markdownContent: string,
  collabToken: string,
  baseUrl: string,
@@ -0,0 +1,18 @@
+/**
+ * Branded canonical page-identity type for the MCP client (incident family #435).
+ *
+ * A Docmost page has TWO identities that are both plain strings: the internal
+ * `page.id` (a canonical UUID the server generates as UUIDv7) and the public
+ * `slugId` (a 10-char nanoid used in URLs). Because both are bare `string`s they
+ * were passed around interchangeably and silently swapped — e.g. locking/keying a
+ * collab doc by the slugId instead of the UUID (the #260 data-loss).
+ *
+ * `PageId` brands the CANONICAL id as a distinct nominal type so a raw/unresolved
+ * string cannot flow into the seams that REQUIRE the canonical id (resolvePageId's
+ * result, the per-page lock key, the collab write entrypoints) — those become a
+ * COMPILE error, catching a swap at build time. It is still a `string` at runtime
+ * (assignable INTO any `string` parameter unchanged), so branding flows outward
+ * for free; the brand is minted at the single canonicalization seam
+ * (`resolvePageId`, via `as PageId`).
+ */
+export type PageId = string & { readonly __brand: "PageId" };
@@ -8,6 +8,8 @@
 * failure) before it runs. Different pages never block each other.
 */

+import type { PageId } from "./page-id.js";
+
 const chains = new Map<string, Promise<unknown>>();

 // Canonical UUID shape (versions 1–8, matching the `uuid` package's `validate`
@@ -28,11 +30,14 @@ export function isUuid(value: string): boolean {
 // awaited/handled by the caller; only the internal chaining tail swallows
 // errors (purely to gate ordering).
 export function withPageLock<T>(
-  pageId: string,
+  pageId: PageId,
  fn: () => Promise<T>,
 ): Promise<T> {
-  // STRUCTURAL INVARIANT (issue #449, "resolve-then-lock"): the mutex key MUST
-  // be the canonical page UUID, never a raw slugId. The whole write path relies
+  // STRUCTURAL INVARIANT (issue #449/#435, "resolve-then-lock"): the mutex key
+  // MUST be the canonical page UUID, never a raw slugId. The `PageId` brand now
+  // enforces this at COMPILE time (a raw string / slugId no longer type-checks
+  // as a key); the runtime assert below stays as a backstop for untyped (JS)
+  // callers and the http/stdio transports. The whole write path relies
  // on the lock key AND the CollabSession cache key being the resolved UUID
  // (#260) — if a future write method forgot to call resolvePageId and locked
  // under a slugId, two writes to the same page would take DIFFERENT mutex keys