feat(search): semantic fusion layer (Phase B PR-1) over lexical RRF (#530)

Fuse a vector-similarity branch into #529's lexical RRF. Query embed via a
global TEI sidecar (or per-workspace provider), degrading transparently to the
byte-identical Phase-A lexical path on any failure; a kill-switch and env knobs
gate it.

- ai.service: resolveEmbeddingProvider (workspace→global TEI fallback) with a
  deterministic config fingerprint; embedQuery (query prefix + short 800ms
  timeout); extract embedWithModel core shared with embedTexts.
- page-embedding.repo: vectorCandidateArm fragment (page-level NN, dim +
  active-fingerprint filtered) + fingerprint on insertChunks.
- search.service: 3-branch RRF over lexical ∪ vector candidates; try/catch wraps
  only the embed (permission filter stays outside, fail-closed); semantic
  degrade emits search.semantic.degraded; response gains semantic{state,available,reason}.
- indexer: resolve provider, prepend doc prefix, stamp fingerprint per row.
- migration 20260712T120000: add nullable page_embeddings.fingerprint + composite index.
- infra: TEI embeddings sidecar in docker-compose + EMBEDDING_*/SEARCH_* in .env.example.
- tests: 7 semantic int cases (vector-only hit, sidecar-down, no-provider,
  permission-over-union fail-closed, hung sidecar, lexical∪vector de-dup,
  fingerprint isolation) + fingerprint/prefix/timeout unit tests.

total now = permission-filtered size of (lexical ∪ vector top-N) — a documented
change from Phase A's exact lexical count (falls back to it on degrade).

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

.env.example

@@ -198,6 +198,42 @@ MCP_DOCMOST_PASSWORD=
 # A slow/hung embeddings endpoint fails after this and the batch continues.
 # AI_EMBEDDING_TIMEOUT_MS=120000
 
+# ─── #530 Semantic search (Phase B) ──────────────────────────────────────────
+# The GLOBAL embedding provider used by search (query embed) AND the indexer
+# (document embed) when a workspace has NO embedding provider of its own. It is
+# an OpenAI-compatible endpoint — the docker-compose `embeddings` (TEI) sidecar.
+# A workspace that configures its own embedding provider OVERRIDES all of this.
+# When neither resolves, search runs lexical-only (semantic.reason=no-provider).
+EMBEDDING_ENDPOINT=http://embeddings:80/v1
+EMBEDDING_MODEL=intfloat/multilingual-e5-small
+# Dummy — the self-hosted TEI sidecar is keyless.
+EMBEDDING_API_KEY=unused
+EMBEDDING_DIMENSIONS=384
+# PLACEHOLDER — set to a PINNED intfloat/multilingual-e5-small commit sha (used
+# both as the TEI --revision and as part of the embedding fingerprint). Keep this
+# in lockstep with docker-compose; a bump changes the fingerprint (PR-2 handles
+# the generational swap/GC — PR-1 uses a single active fingerprint).
+EMBEDDING_REVISION=REPLACE_WITH_PINNED_E5_SMALL_COMMIT_SHA
+# e5 models require these input prefixes. Empty them for a non-e5 model.
+EMBEDDING_QUERY_PREFIX="query: "
+EMBEDDING_DOC_PREFIX="passage: "
+# Per-request timeout (ms) for the interactive SEARCH query embed — much shorter
+# than the batch indexer timeout: a slow/hung sidecar degrades search fast to the
+# lexical-only path instead of blocking the request. Default 800.
+# SEARCH_EMBED_TIMEOUT_MS=800
+# RRF weight of the vector leg relative to each lexical leg (1.0 = equal). Default 1.0.
+# SEARCH_VECTOR_WEIGHT=1.0
+# Vector top-N pulled into the fused candidate union per request. Default 50.
+# SEARCH_VECTOR_CANDIDATES=50
+# Per-statement timeout (ms) bounding ONLY the fused vector query (the brute-force
+# KNN seq scan — there is no ANN index). If the vector scan exceeds this it is
+# cancelled and search degrades to lexical-only (never hangs the request). Scoped
+# per-query (SET LOCAL), so it never affects other queries. Default 2000.
+# SEARCH_VECTOR_STATEMENT_TIMEOUT_MS=2000
+# Kill-switch: set to `off` to disable the semantic layer entirely (search then
+# runs the byte-identical Phase-A lexical path). Default on.
+# SEARCH_SEMANTIC=on
+
 # Silence timeout (ms) for streaming chat/agent AI calls AND external-MCP traffic.
 # Bounds time-to-first-byte and the gap BETWEEN chunks (NOT the total turn length),
 # so an arbitrarily long turn that keeps streaming is never cut. Finite so a hung

NewDesign/CommentsPanel.tsx

@@ -0,0 +1,460 @@
+/**
+ * CommentsPanel — редизайн панели комментариев с фокусом на агентские правки.
+ * Mantine v7. Узкая колонка ~340–400px, светлая/тёмная темы.
+ *
+ * Реализованные решения (см. README.md):
+ *  - Дифф-первая карточка: цитата = старая строка диффа (без тройного дублирования)
+ *  - Группировка серии прогона под одной шапкой ("Корректор · 28 правок")
+ *  - Пакетное "Accept all" на фронте (по одной под капотом) с полоской прогресса
+ *  - Метки важности/категории парсятся клиентом из тегов "[Корректура][Существенно]"
+ *  - Фильтры по важности/категории
+ *  - Безопасный Dismiss через undo-тост (удаление откладывается на клиенте)
+ *  - Состояния: pending / applied / dismissed / conflict (потерян якорь)
+ *  - Крайние диффы: вставка (␣), удаление, одна буква, дефис→тире, длинный абзац
+ *  - Человеческий тред не деградирует (та же система, другое наполнение)
+ *  - Вкладки Open/Resolved сохранены
+ */
+import { useMemo, useState, useCallback, useRef } from 'react';
+import {
+  Box, Group, Stack, Text, Badge, Button, ActionIcon, Tabs, Progress,
+  Avatar, Tooltip, ScrollArea, useMantineColorScheme, useMantineTheme,
+} from '@mantine/core';
+import { notifications } from '@mantine/notifications';
+
+/* ─────────────────────────── Типы данных ─────────────────────────── */
+
+export type Severity = 'critical' | 'major' | 'minor';
+export type Category = string; // "Корректура" | "Факт" | "Стиль" | …
+
+/** Один сегмент интра-диффа: изменённый фрагмент подсвечивается. */
+export interface DiffSegment {
+  text: string;
+  changed: boolean;
+}
+
+/** Правка: сервер уже отдаёт посегментную разметку обеих строк. */
+export interface SuggestedEdit {
+  before: DiffSegment[]; // "было" (пустой массив ⇒ чистая вставка)
+  after: DiffSegment[];  // "стало" (пустой массив ⇒ чистое удаление)
+}
+
+export type CommentStatus = 'pending' | 'applied' | 'dismissed' | 'conflict';
+
+export interface Comment {
+  id: string;
+  runId?: string;          // id прогона агента — для группировки серии
+  authorName: string;      // "Корректор" | "Нарратор" | имя человека
+  authorKind: 'agent' | 'human';
+  triggeredBy?: string;    // кто запустил агента ("vvzvlad")
+  createdAtLabel: string;  // "15 ч", "2 дн" — форматируется вызывающим кодом
+  /** Сырой текст комментария от агента, метки в квадратных скобках. */
+  bodyRaw: string;
+  edit?: SuggestedEdit;    // есть ⇒ агентская правка; нет ⇒ обычный тред
+  status: CommentStatus;
+  replyCount?: number;
+  anchorLost?: boolean;    // сервер ответил конфликтом якоря
+}
+
+export interface CommentsPanelProps {
+  comments: Comment[];
+  /** Применить одну правку. Резолвит тред на сервере. */
+  onApply: (id: string) => Promise<void>;
+  /** Жёсткое удаление на сервере (для треда без ответов). Вызывается ПОСЛЕ окна undo. */
+  onDismiss: (id: string) => Promise<void>;
+  /** Резолв обычного треда. */
+  onResolve?: (id: string) => Promise<void>;
+  /** Клик по карточке/цитате — скролл документа к якорю и подсветка. */
+  onNavigateToAnchor: (id: string) => void;
+  onClose?: () => void;
+  /** Мс до фактического удаления после Dismiss (окно undo). По умолчанию 5000. */
+  dismissUndoMs?: number;
+}
+
+/* ───────────────────── Клиентский парсинг меток ───────────────────── */
+
+const SEVERITY_WORDS: Record<string, Severity> = {
+  'существенно': 'major', 'критично': 'critical', 'критическая': 'critical',
+  'незначительно': 'minor', 'мелко': 'minor', 'major': 'major',
+  'minor': 'minor', 'critical': 'critical',
+};
+
+interface ParsedBody {
+  category?: Category;
+  severity: Severity;
+  text: string; // тело без скобочных тегов
+}
+
+/** "[Корректура] [Существенно] Пробел…" → {category, severity, text}. */
+export function parseBody(raw: string): ParsedBody {
+  const tags: string[] = [];
+  const text = raw.replace(/\[([^\]]+)\]/g, (_, t) => { tags.push(t.trim()); return ''; }).trim();
+  let severity: Severity = 'minor';
+  let category: Category | undefined;
+  for (const t of tags) {
+    const sv = SEVERITY_WORDS[t.toLowerCase()];
+    if (sv) severity = sv; else if (!category) category = t;
+  }
+  return { category, severity, text };
+}
+
+const SEV_COLOR: Record<Severity, string> = {
+  critical: 'red', major: 'orange', minor: 'gray',
+};
+
+/* ──────────────────────────── Дифф ──────────────────────────── */
+
+/** Делает невидимые символы видимыми в подсветке (пробел, таб). */
+function visibleWhitespace(s: string): string {
+  return s.replace(/ /g, '␣').replace(/\t/g, '⇥');
+}
+
+function DiffLine({ segments, kind }: { segments: DiffSegment[]; kind: 'del' | 'ins' }) {
+  const theme = useMantineTheme();
+  const { colorScheme } = useMantineColorScheme();
+  const dark = colorScheme === 'dark';
+  const isDel = kind === 'del';
+  const sign = isDel ? '−' : '+';
+  const signColor = isDel ? theme.colors.red[dark ? 5 : 6] : theme.colors.green[dark ? 5 : 6];
+  const baseColor = isDel
+    ? (dark ? theme.colors.gray[5] : theme.colors.gray[6])
+    : (dark ? theme.colors.gray[1] : theme.colors.dark[9]);
+  const markBg = isDel
+    ? (dark ? 'rgba(224,49,49,.22)' : '#ffe3e3')
+    : (dark ? 'rgba(47,158,68,.22)' : '#d3f9d8');
+  const markFg = isDel
+    ? (dark ? theme.colors.red[3] : theme.colors.red[8])
+    : (dark ? theme.colors.green[3] : theme.colors.green[9]);
+
+  return (
+    <Group gap={7} wrap="nowrap" align="flex-start">
+      <Text ff="monospace" fw={600} fz={12} c={signColor} w={11} ta="center" style={{ flex: 'none', lineHeight: 1.5 }}>{sign}</Text>
+      <Text fz={13.5} c={baseColor} style={{ lineHeight: 1.45 }}>
+        {segments.map((seg, i) =>
+          seg.changed ? (
+            <Box key={i} component="mark" px={3} fw={600}
+              style={{ background: markBg, color: markFg, borderRadius: 3,
+                textDecoration: isDel ? 'line-through' : 'none' }}>
+              {visibleWhitespace(seg.text)}
+            </Box>
+          ) : (
+            <Box key={i} component="span" style={{ textDecoration: isDel ? 'line-through' : 'none' }}>{seg.text}</Box>
+          )
+        )}
+      </Text>
+    </Group>
+  );
+}
+
+function DiffBlock({ edit }: { edit: SuggestedEdit }) {
+  const pureInsert = edit.before.length === 0;
+  const pureDelete = edit.after.length === 0;
+  return (
+    <Stack gap={1}>
+      {!pureInsert && <DiffLine segments={edit.before} kind="del" />}
+      {!pureDelete && <DiffLine segments={edit.after} kind="ins" />}
+    </Stack>
+  );
+}
+
+/* ──────────────────────── Карточка правки ──────────────────────── */
+
+function EditCard({ c, onApply, onDismiss, onNavigateToAnchor, dismissUndoMs = 5000 }: {
+  c: Comment;
+  onApply: CommentsPanelProps['onApply'];
+  onDismiss: CommentsPanelProps['onDismiss'];
+  onNavigateToAnchor: CommentsPanelProps['onNavigateToAnchor'];
+  dismissUndoMs?: number;
+}) {
+  const parsed = useMemo(() => parseBody(c.bodyRaw), [c.bodyRaw]);
+  const [busy, setBusy] = useState(false);
+  const timer = useRef<number>();
+
+  const apply = useCallback(async () => {
+    setBusy(true);
+    try { await onApply(c.id); } finally { setBusy(false); }
+  }, [c.id, onApply]);
+
+  // Безопасный Dismiss: прячем сразу, удаляем на сервере после окна undo.
+  const dismiss = useCallback(() => {
+    let undone = false;
+    const nid = notifications.show({
+      message: 'Edit dismissed',
+      color: 'gray',
+      autoClose: dismissUndoMs,
+      withCloseButton: false,
+      // Кнопка "Вернуть" — см. README про кастомный рендер экшена.
+    });
+    timer.current = window.setTimeout(() => {
+      if (!undone) onDismiss(c.id);
+    }, dismissUndoMs);
+    // undo вызывается из UI: clearTimeout(timer.current); undone = true;
+    return { nid, cancel: () => { undone = true; clearTimeout(timer.current); } };
+  }, [c.id, onDismiss, dismissUndoMs]);
+
+  const conflict = c.status === 'conflict' || c.anchorLost;
+
+  return (
+    <Box
+      p="10px 12px"
+      onClick={() => onNavigateToAnchor(c.id)}
+      style={(t) => ({
+        cursor: 'pointer',
+        borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}`,
+      })}
+    >
+      <Stack gap={8}>
+        {c.edit && <DiffBlock edit={c.edit} />}
+
+        {conflict && (
+          <Group gap={7} wrap="nowrap" p="6px 8px"
+            style={(t) => ({ background: t.colorScheme === 'dark' ? 'rgba(230,180,20,.12)' : '#fff9db', borderRadius: 7 })}>
+            <Text fz={12}>⚠</Text>
+            <Text fz={12} c="yellow.8" style={{ lineHeight: 1.4 }}>Text changed — this edit can’t be applied</Text>
+          </Group>
+        )}
+
+        {parsed.text && !conflict && (
+          <Text fz={12.5} c="dimmed" style={{ lineHeight: 1.45 }}>{parsed.text}</Text>
+        )}
+
+        <Group gap={8} wrap="nowrap" onClick={(e) => e.stopPropagation()}>
+          <Box w={8} h={8} style={(t) => ({ flex: 'none', borderRadius: '50%', background: t.colors[SEV_COLOR[parsed.severity]][6] })} />
+          <Text fz={10} fw={600} tt="uppercase" c="dimmed" style={{ letterSpacing: '.06em', flex: 1, whiteSpace: 'nowrap', overflow: 'hidden', textOverflow: 'ellipsis' }}>
+            {[parsed.category, parsed.severity === 'major' ? 'Major' : parsed.severity === 'critical' ? 'Critical' : null].filter(Boolean).join(' · ')}
+          </Text>
+
+          {c.status === 'applied' ? (
+            <Badge color="green" variant="light" radius="xl" size="sm">✓ Applied</Badge>
+          ) : c.status === 'dismissed' ? (
+            <Badge color="gray" variant="light" radius="xl" size="sm">Dismissed</Badge>
+          ) : conflict ? (
+            <Button size="compact-sm" variant="default" onClick={() => onNavigateToAnchor(c.id)}>Go to text</Button>
+          ) : (
+            <>
+              <Button size="compact-sm" variant="default" color="gray" onClick={dismiss}>Dismiss</Button>
+              <Button size="compact-sm" color="green" loading={busy} onClick={apply}>Apply</Button>
+            </>
+          )}
+        </Group>
+      </Stack>
+    </Box>
+  );
+}
+
+/* ──────────────────────── Человеческий тред ──────────────────────── */
+
+function HumanThread({ c, onResolve, onNavigateToAnchor }: {
+  c: Comment; onResolve?: CommentsPanelProps['onResolve']; onNavigateToAnchor: CommentsPanelProps['onNavigateToAnchor'];
+}) {
+  const parsed = useMemo(() => parseBody(c.bodyRaw), [c.bodyRaw]);
+  return (
+    <Box p="12px 14px" style={(t) => ({ borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+      <Stack gap={9}>
+        <Group gap={8} wrap="nowrap">
+          <Avatar size={26} radius="xl" color="orange">{c.authorName[0]}</Avatar>
+          <Text fz={12.5} fw={600}>{c.authorName}
+            <Text span c="dimmed" fw={400}> · {c.triggeredBy ? `${c.triggeredBy} · ` : ''}{c.createdAtLabel}</Text>
+          </Text>
+          <Box style={{ flex: 1 }} />
+          <Tooltip label="Resolve"><ActionIcon variant="default" size="md" onClick={() => onResolve?.(c.id)}>✓</ActionIcon></Tooltip>
+          <ActionIcon variant="default" size="md">⋯</ActionIcon>
+        </Group>
+        <Text fz={13.5} style={{ lineHeight: 1.5 }}>{parsed.text}</Text>
+        <Group gap={8}>
+          {c.replyCount ? <Text fz={12} c="dimmed">{c.replyCount} replies</Text> : null}
+          <Box style={{ flex: 1 }} />
+          <Button variant="subtle" size="compact-sm">Reply</Button>
+        </Group>
+      </Stack>
+    </Box>
+  );
+}
+
+/* ──────────────────── Шапка серии + прогресс ──────────────────── */
+
+function RunHeader({ runComments, onAcceptAll, progress }: {
+  runComments: Comment[];
+  onAcceptAll: () => void;
+  progress?: { done: number; total: number } | null;
+}) {
+  const head = runComments[0];
+  const majors = runComments.filter((c) => parseBody(c.bodyRaw).severity !== 'minor').length;
+  const applied = runComments.filter((c) => c.status === 'applied').length;
+  return (
+    <Box>
+      <Group gap={10} wrap="nowrap" p="11px 13px" style={(t) => ({ borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+        <Box style={{ position: 'relative', width: 30, height: 30, flex: 'none' }}>
+          <Avatar size={30} radius={8} color="teal">{head.authorName[0]}</Avatar>
+          {head.triggeredBy && (
+            <Avatar size={14} radius="xl" color="gray"
+              style={{ position: 'absolute', right: -4, bottom: -4, border: '2px solid var(--mantine-color-body)' }} />
+          )}
+        </Box>
+        <Stack gap={1} style={{ minWidth: 0 }}>
+          <Text fz={13} fw={600}>{head.authorName}
+            <Text span c="dimmed" fw={400}> · {head.triggeredBy} · {head.createdAtLabel}</Text>
+          </Text>
+          <Text fz={11.5} c="dimmed">
+            {runComments.length} edits · <Text span c="orange.7" fw={600}>{majors} major</Text> · {applied} applied
+          </Text>
+        </Stack>
+        <Box style={{ flex: 1 }} />
+        {!progress && <Button size="compact-sm" color="green" onClick={onAcceptAll}>Accept all</Button>}
+      </Group>
+      {progress && (
+        <Box p="9px 13px 11px" style={(t) => ({ background: t.colorScheme === 'dark' ? 'rgba(47,158,68,.08)' : '#f8fbf9', borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[2]}` })}>
+          <Group gap={8} mb={6}>
+            <Text fz={12} fw={600} c="green.7">Applying {progress.done} of {progress.total}…</Text>
+            <Box style={{ flex: 1 }} />
+            <Button variant="subtle" color="gray" size="compact-xs">Stop</Button>
+          </Group>
+          <Progress value={(progress.done / progress.total) * 100} color="green" size="sm" radius="xl" />
+        </Box>
+      )}
+    </Box>
+  );
+}
+
+/* ──────────────────────────── Панель ──────────────────────────── */
+
+/** Роль-автор для фильтра: у агентов — имя роли, у людей — «Пользователь». */
+function roleOf(c: Comment): string {
+  return c.authorKind === 'human' ? 'Пользователь' : c.authorName;
+}
+
+export function CommentsPanel(props: CommentsPanelProps) {
+  const { comments, onApply, onDismiss, onResolve, onNavigateToAnchor, onClose, dismissUndoMs } = props;
+  const [tab, setTab] = useState<'open' | 'resolved'>('open');
+  const [roleFilter, setRoleFilter] = useState<string | null>(null);
+  const [progress, setProgress] = useState<Record<string, { done: number; total: number }>>({});
+
+  const open = comments.filter((c) => c.status === 'pending' || c.status === 'conflict');
+  const resolved = comments.filter((c) => c.status === 'applied' || c.status === 'dismissed');
+  const list = tab === 'open' ? open : resolved;
+
+  const filtered = list.filter((c) => !roleFilter || roleOf(c) === roleFilter);
+
+  // Роли и счётчики для чипов-фильтров (только роли: Корректор / Фактчекер / Пользователь …).
+  const roles = useMemo(() => {
+    const order: string[] = [];
+    const count: Record<string, number> = {};
+    for (const c of list) {
+      const r = roleOf(c);
+      if (!(r in count)) { count[r] = 0; order.push(r); }
+      count[r]++;
+    }
+    return order.map((r) => ({ role: r, count: count[r] }));
+  }, [list]);
+
+  // Группировка агентских правок по runId; человеческие треды — по одному.
+  const groups = useMemo(() => {
+    const map = new Map<string, Comment[]>();
+    const singles: Comment[] = [];
+    for (const c of filtered) {
+      if (c.authorKind === 'agent' && c.runId && c.edit) {
+        if (!map.has(c.runId)) map.set(c.runId, []);
+        map.get(c.runId)!.push(c);
+      } else singles.push(c);
+    }
+    return { runs: [...map.entries()], singles };
+  }, [filtered]);
+
+  // Пакетное "Accept all" — по одной на фронте, обновляем прогресс.
+  const acceptAll = useCallback(async (runId: string, items: Comment[]) => {
+    const minor = items.filter((c) => parseBody(c.bodyRaw).severity === 'minor' && c.status === 'pending');
+    for (let i = 0; i < minor.length; i++) {
+      setProgress((p) => ({ ...p, [runId]: { done: i, total: minor.length } }));
+      try { await onApply(minor[i].id); } catch { /* конфликт — пропускаем, копим для сводки */ }
+    }
+    setProgress((p) => { const n = { ...p }; delete n[runId]; return n; });
+    notifications.show({ color: 'green', message: `${minor.length} applied` });
+  }, [onApply]);
+
+  return (
+    <Stack gap={0} h="100%" style={{ width: 380, maxWidth: '100%', borderLeft: '1px solid var(--mantine-color-default-border)' }}>
+      {/* Вкладки — статусная ось. Open/Resolved сохранены. */}
+      <Group gap={4} p="10px 14px" style={(t) => ({ borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[2]}` })}>
+        <Tabs value={tab} onChange={(v) => setTab(v as 'open' | 'resolved')} variant="pills">
+          <Tabs.List>
+            <Tabs.Tab value="open" rightSection={<Badge size="sm" variant="light" color="blue">{open.length}</Badge>}>Open</Tabs.Tab>
+            <Tabs.Tab value="resolved" rightSection={<Badge size="sm" variant="light" color="gray">{resolved.length}</Badge>}>Resolved</Tabs.Tab>
+          </Tabs.List>
+        </Tabs>
+        <Box style={{ flex: 1 }} />
+        {onClose && <ActionIcon variant="subtle" color="gray" onClick={onClose}>✕</ActionIcon>}
+      </Group>
+
+      {/* Фильтры-чипы: только по ролям (Корректор / Фактчекер / Пользователь). */}
+      {tab === 'open' && roles.length > 1 && (
+        <ScrollArea type="never" p="8px 14px" style={(t) => ({ borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+          <Group gap={6} wrap="nowrap">
+            <FilterChip active={!roleFilter} onClick={() => setRoleFilter(null)} label={`All ${list.length}`} solid />
+            {roles.map(({ role, count }) => (
+              <FilterChip key={role} active={roleFilter === role} onClick={() => setRoleFilter(roleFilter === role ? null : role)} label={`${role} ${count}`} />
+            ))}
+          </Group>
+        </ScrollArea>
+      )}
+
+      {/* Лента */}
+      <ScrollArea style={{ flex: 1 }} bg="var(--mantine-color-default-hover)">
+        {filtered.length === 0 ? (
+          <EmptyState tab={tab} />
+        ) : (
+          <>
+            {groups.runs.map(([runId, items]) => (
+              <Box key={runId} m="6px 10px" bg="var(--mantine-color-body)" style={{ border: '1px solid var(--mantine-color-default-border)', borderRadius: 10, overflow: 'hidden' }}>
+                <RunHeader runComments={items} progress={progress[runId] ?? null} onAcceptAll={() => acceptAll(runId, items)} />
+                {items.map((c) => (
+                  <EditCard key={c.id} c={c} onApply={onApply} onDismiss={onDismiss} onNavigateToAnchor={onNavigateToAnchor} dismissUndoMs={dismissUndoMs} />
+                ))}
+              </Box>
+            ))}
+            {groups.singles.map((c) => (
+              <Box key={c.id} m="6px 10px" bg="var(--mantine-color-body)" style={{ border: '1px solid var(--mantine-color-default-border)', borderRadius: 10, overflow: 'hidden' }}>
+                {c.edit
+                  ? <EditCard c={c} onApply={onApply} onDismiss={onDismiss} onNavigateToAnchor={onNavigateToAnchor} dismissUndoMs={dismissUndoMs} />
+                  : <HumanThread c={c} onResolve={onResolve} onNavigateToAnchor={onNavigateToAnchor} />}
+              </Box>
+            ))}
+          </>
+        )}
+      </ScrollArea>
+    </Stack>
+  );
+}
+
+/* ─────────────────────── Мелкие компоненты ─────────────────────── */
+
+function FilterChip({ label, active, onClick, dot, solid }: {
+  label: string; active: boolean; onClick: () => void; dot?: string; solid?: boolean;
+}) {
+  return (
+    <Button
+      onClick={onClick}
+      size="compact-sm"
+      radius="xl"
+      variant={active ? 'filled' : 'default'}
+      color={active ? (solid ? 'dark' : 'blue') : 'gray'}
+      leftSection={dot ? <Box w={7} h={7} style={(t) => ({ borderRadius: '50%', background: t.colors[dot][6] })} /> : undefined}
+      styles={{ root: { flex: 'none' }, label: { fontWeight: 600, fontSize: 12 } }}
+    >
+      {label}
+    </Button>
+  );
+}
+
+function EmptyState({ tab }: { tab: 'open' | 'resolved' }) {
+  const isOpen = tab === 'open';
+  return (
+    <Stack align="center" gap={6} p="40px 20px">
+      <Avatar size={40} radius="xl" color={isOpen ? 'green' : 'gray'}>{isOpen ? '✓' : '◌'}</Avatar>
+      <Text fw={600} fz={13}>{isOpen ? 'All caught up' : 'Nothing here yet'}</Text>
+      <Text fz={12} c="dimmed" ta="center" style={{ lineHeight: 1.45 }}>
+        {isOpen ? 'No edits waiting on you.' : 'Applied and closed items will appear here.'}
+      </Text>
+    </Stack>
+  );
+}
+
+export default CommentsPanel;

NewDesign/PageHistoryModal.tsx

@@ -0,0 +1,362 @@
+/**
+ * PageHistoryModal — редизайн окна «Page history».
+ * Mantine v7. Light/dark. Полноразмерное модальное окно.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * ЧТО ЭТО
+ * ─────────────────────────────────────────────────────────────────────────
+ * Окно истории версий страницы. Слева — панель навигации: мини-календарь
+ * (heatmap: яркость дня = число ревизий, обводка = выбранный день) + плотный
+ * список ревизий (одна ревизия = одна строка). Справа — реально отрендеренная
+ * версия страницы с подсветкой изменений. Все контролы — в одну строку шапки.
+ *
+ * Ключевые решения:
+ *  - Плотный список: одна ревизия на строку (аватар · время · автор · [SAVED]).
+ *    Бейдж показывается ТОЛЬКО у сохранённых версий; всё остальное — autosave.
+ *  - Агентские ревизии визуально отличаются: квадратный глиф роли (К/Ф) вместо
+ *    круглого аватара пользователя + «via <кто запустил>».
+ *  - Календарь объединён со списком в одной панели (мини-календарь = date jumper).
+ *  - Highlight changes + навигация по изменениям (N of M, ↑/↓) вынесены в шапку.
+ *  - Текущая версия помечена; Restore для неё недоступен.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * УСТАНОВКА / ИСПОЛЬЗОВАНИЕ
+ * ─────────────────────────────────────────────────────────────────────────
+ *   import { PageHistoryModal, Revision } from './PageHistoryModal';
+ *
+ *   <PageHistoryModal
+ *     opened={open}
+ *     onClose={() => setOpen(false)}
+ *     revisions={revisions}                 // Revision[]
+ *     selectedId={selId}
+ *     onSelect={setSelId}                    // клик по ревизии → рендер справа
+ *     renderVersion={(rev) => <ArticleView versionId={rev.id} />}  // ваш рендер страницы
+ *     highlightChanges={hl}
+ *     onToggleHighlight={setHl}
+ *     changeNav={{ index: 1, total: 3, onPrev, onNext }}  // навигация по диффу
+ *     onlySaved={onlySaved}
+ *     onToggleOnlySaved={setOnlySaved}
+ *     onRestore={(rev) => api.restore(rev.id)}   // async; текущая версия → disabled
+ *   />
+ *
+ * Требует MantineProvider на корне приложения (defaultColorScheme="auto" для тем).
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * ФОРМАТ ДАННЫХ
+ * ─────────────────────────────────────────────────────────────────────────
+ *   interface Revision {
+ *     id: string;
+ *     at: Date | string;        // время ревизии; форматируется вызывающим/util-ом
+ *     dayGroup: string;         // 'Today' | 'Yesterday' | 'Mon 12 Jul' — заголовок группы
+ *     saved: boolean;           // true → бейдж SAVED; false → autosave (без бейджа)
+ *     author: { name: string } & (
+ *       | { kind: 'human' }
+ *       | { kind: 'agent'; role: string; triggeredBy: string }  // роль + кто запустил
+ *     );
+ *     isCurrent?: boolean;      // текущая (последняя) версия — Restore недоступен
+ *   }
+ *
+ * Ревизии приходят плоским массивом; группировка по dayGroup — на клиенте.
+ * Никаких изменений API не требуется: агентское авторство берётся из author.kind.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * СОСТОЯНИЯ (нарисованы/поддержаны)
+ * ─────────────────────────────────────────────────────────────────────────
+ *  - Ревизия: обычная / hover / выбранная / текущая / агентская / человеческая.
+ *  - Restore: default / disabled (выбрана текущая) / loading (спиннер после клика).
+ *  - Highlight changes: on/off; при off навигация по изменениям неактивна.
+ *  - Пустая история: одна версия / нет ревизий → EmptyState.
+ *  - Тёмная тема: через токены Mantine (useMantineColorScheme, var(--mantine-*)).
+ */
+import { useMemo, useState, useCallback } from 'react';
+import {
+  Modal, Box, Group, Stack, Text, Switch, Avatar, Badge, Button, ActionIcon,
+  ScrollArea, useMantineTheme, useMantineColorScheme,
+} from '@mantine/core';
+
+/* ─────────────────────────── Типы ─────────────────────────── */
+
+export type Author =
+  | { name: string; kind: 'human' }
+  | { name: string; kind: 'agent'; role: string; triggeredBy: string };
+
+export interface Revision {
+  id: string;
+  at: Date | string;
+  atLabel: string;      // предформатированное «5:35AM»
+  dayGroup: string;     // «Today» | «Yesterday» | «Mon 12 Jul»
+  saved: boolean;
+  author: Author;
+  isCurrent?: boolean;
+}
+
+export interface CalendarDay {
+  label: string;        // «12»
+  inMonth: boolean;
+  count: number;        // число ревизий за день (для heatmap)
+  selected?: boolean;
+  date: Date | string;
+}
+
+export interface PageHistoryModalProps {
+  opened: boolean;
+  onClose: () => void;
+  revisions: Revision[];
+  selectedId: string | null;
+  onSelect: (id: string) => void;
+  renderVersion: (rev: Revision | null) => React.ReactNode;
+  highlightChanges: boolean;
+  onToggleHighlight: (v: boolean) => void;
+  changeNav?: { index: number; total: number; onPrev: () => void; onNext: () => void };
+  onlySaved: boolean;
+  onToggleOnlySaved: (v: boolean) => void;
+  onRestore: (rev: Revision) => Promise<void>;
+  /** Ячейки текущего месяца календаря + управление. Необязательно — без него панель = только список. */
+  calendar?: {
+    monthLabel: string;      // «July 2025»
+    weekdays: string[];      // ['Mon',…,'Sun']
+    days: CalendarDay[];     // обычно 35/42 ячейки
+    onPrevMonth: () => void;
+    onNextMonth: () => void;
+    onToday: () => void;
+    onPickDay: (d: CalendarDay) => void;
+  };
+}
+
+/* ─────────────────────────── Утилиты представления ─────────────────────────── */
+
+const USER_PALETTE = ['#495057', '#e8590c', '#1c7ed6', '#7048e8', '#0ca678'];
+function userColor(name: string) {
+  let h = 0;
+  for (let i = 0; i < name.length; i++) h = (h * 31 + name.charCodeAt(i)) >>> 0;
+  return USER_PALETTE[h % USER_PALETTE.length];
+}
+const AGENT_GRAD: Record<string, string> = {
+  'Корректор': 'linear-gradient(135deg,#20c997,#12b886)',
+  'Фактчекер': 'linear-gradient(135deg,#4c6ef5,#7048e8)',
+};
+function agentGrad(role: string) { return AGENT_GRAD[role] ?? 'linear-gradient(135deg,#868e96,#495057)'; }
+
+/** heatmap: число ревизий → фон/текст ячейки календаря. */
+function heat(n: number, dark: boolean) {
+  if (n === 0) return { bg: 'transparent', fg: dark ? '#5c5f66' : '#adb5bd' };
+  if (n <= 2) return { bg: dark ? 'rgba(34,139,230,.20)' : '#e7f0ff', fg: dark ? '#74c0fc' : '#1971c2' };
+  if (n <= 4) return { bg: dark ? 'rgba(34,139,230,.45)' : '#a5c8ff', fg: dark ? '#dbeafe' : '#0b3d91' };
+  return { bg: '#4c8dff', fg: '#ffffff' };
+}
+
+/* ─────────────────────────── Строка ревизии ─────────────────────────── */
+
+function RevisionRow({ rev, selected, onSelect }: { rev: Revision; selected: boolean; onSelect: () => void }) {
+  const theme = useMantineTheme();
+  const { colorScheme } = useMantineColorScheme();
+  const dark = colorScheme === 'dark';
+  const a = rev.author;
+  const isAgent = a.kind === 'agent';
+
+  return (
+    <Group
+      gap={8} wrap="nowrap" h={28} px="12px 14px" m="1px 6px"
+      onClick={onSelect}
+      style={{
+        cursor: 'pointer', borderRadius: 7,
+        background: selected ? (dark ? 'rgba(34,139,230,.14)' : '#eef4ff')
+          : isAgent ? (dark ? 'rgba(112,72,232,.06)' : '#fbfaff') : undefined,
+      }}
+    >
+      {/* аватар/глиф */}
+      {isAgent ? (
+        <Box style={{ flex: 'none', width: 16, height: 16, borderRadius: 4, background: agentGrad(a.role), display: 'flex', alignItems: 'center', justifyContent: 'center', color: '#fff', font: '700 8px system-ui' }}>
+          {a.role[0]}
+        </Box>
+      ) : (
+        <Box style={{ flex: 'none', width: 16, height: 16, borderRadius: '50%', background: userColor(a.name), display: 'flex', alignItems: 'center', justifyContent: 'center', color: '#fff', font: '700 8px system-ui' }}>
+          {a.name[0].toUpperCase()}
+        </Box>
+      )}
+      {/* время */}
+      <Text fz={12.5} fw={rev.isCurrent ? 600 : 500} c={rev.isCurrent ? undefined : 'dimmed'} style={{ flex: 'none', minWidth: 58 }}>
+        {rev.atLabel}
+      </Text>
+      {/* автор + via */}
+      <Group gap={4} wrap="nowrap" style={{ flex: 1, minWidth: 0, alignItems: 'baseline', overflow: 'hidden' }}>
+        <Text fz={12} fw={isAgent ? 600 : 400} c={isAgent ? undefined : 'dimmed'} truncate style={{ flex: 'none', maxWidth: 100 }}>
+          {isAgent ? a.role : a.name}
+        </Text>
+        {isAgent && <Text fz={11} c="dimmed" truncate>· via {a.triggeredBy}</Text>}
+      </Group>
+      {/* бейдж — только SAVED */}
+      {rev.saved && <Badge size="sm" radius="sm" variant="light" color="blue">SAVED</Badge>}
+    </Group>
+  );
+}
+
+/* ─────────────────────────── Мини-календарь ─────────────────────────── */
+
+function MiniCalendar({ cal }: { cal: NonNullable<PageHistoryModalProps['calendar']> }) {
+  const { colorScheme } = useMantineColorScheme();
+  const dark = colorScheme === 'dark';
+  return (
+    <Box p="10px 12px 8px" style={(t) => ({ borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+      <Group gap={6} mb={6}>
+        <ActionIcon variant="subtle" color="gray" size="sm" onClick={cal.onPrevMonth}>‹</ActionIcon>
+        <Text fz={12} fw={600}>{cal.monthLabel}</Text>
+        <ActionIcon variant="subtle" color="gray" size="sm" onClick={cal.onNextMonth}>›</ActionIcon>
+        <Box style={{ flex: 1 }} />
+        <Button variant="subtle" size="compact-xs" onClick={cal.onToday}>Today</Button>
+      </Group>
+      <Box style={{ display: 'grid', gridTemplateColumns: 'repeat(7,1fr)', gap: 2, marginBottom: 2 }}>
+        {cal.weekdays.map((w) => (
+          <Text key={w} ta="center" fz={8.5} fw={600} c="dimmed">{w}</Text>
+        ))}
+      </Box>
+      <Box style={{ display: 'grid', gridTemplateColumns: 'repeat(7,1fr)', gap: 2 }}>
+        {cal.days.map((d, i) => {
+          const h = heat(d.count, dark);
+          return (
+            <Box
+              key={i} onClick={() => cal.onPickDay(d)}
+              style={{
+                position: 'relative', height: 26, borderRadius: 6, cursor: 'pointer',
+                display: 'flex', alignItems: 'center', justifyContent: 'center',
+                background: d.inMonth ? h.bg : 'transparent',
+                boxShadow: d.selected ? `inset 0 0 0 2px ${dark ? '#f1f3f5' : '#1a1b1e'}` : undefined,
+              }}
+            >
+              <Text fz={10.5} fw={d.selected ? 700 : 500} style={{ color: d.inMonth ? h.fg : (dark ? '#3a3d42' : '#dee2e6') }}>
+                {d.label}
+              </Text>
+            </Box>
+          );
+        })}
+      </Box>
+      {/* легенда heatmap */}
+      <Group gap={6} mt={8} align="center">
+        <Text fz={9.5} c="dimmed">fewer</Text>
+        {['#e7f0ff', '#a5c8ff', '#4c8dff'].map((c) => (
+          <Box key={c} style={{ width: 14, height: 10, borderRadius: 2, background: c }} />
+        ))}
+        <Text fz={9.5} c="dimmed">more revisions</Text>
+      </Group>
+    </Box>
+  );
+}
+
+/* ─────────────────────────── Окно ─────────────────────────── */
+
+export function PageHistoryModal(props: PageHistoryModalProps) {
+  const {
+    opened, onClose, revisions, selectedId, onSelect, renderVersion,
+    highlightChanges, onToggleHighlight, changeNav, onlySaved, onToggleOnlySaved,
+    onRestore, calendar,
+  } = props;
+  const [restoring, setRestoring] = useState(false);
+
+  const selected = revisions.find((r) => r.id === selectedId) ?? null;
+  const isEmpty = revisions.length <= 1;
+
+  // группировка по dayGroup, с фильтром Only saved
+  const groups = useMemo(() => {
+    const list = onlySaved ? revisions.filter((r) => r.saved) : revisions;
+    const map: { head: string; items: Revision[] }[] = [];
+    for (const r of list) {
+      let g = map.find((x) => x.head === r.dayGroup);
+      if (!g) { g = { head: r.dayGroup, items: [] }; map.push(g); }
+      g.items.push(r);
+    }
+    return map;
+  }, [revisions, onlySaved]);
+
+  const restore = useCallback(async () => {
+    if (!selected || selected.isCurrent) return;
+    setRestoring(true);
+    try { await onRestore(selected); } finally { setRestoring(false); }
+  }, [selected, onRestore]);
+
+  return (
+    <Modal
+      opened={opened} onClose={onClose} withCloseButton={false}
+      size="80rem" radius="lg" padding={0}
+      styles={{ body: { height: '80vh', maxHeight: 760, display: 'flex', flexDirection: 'column' } }}
+      overlayProps={{ backgroundOpacity: 0.5, blur: 1 }}
+    >
+      {/* ── single-row toolbar ── */}
+      <Group gap={14} h={60} px={16} wrap="nowrap" style={(t) => ({ flex: 'none', borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+        <Text fz={16} fw={600}>Page history</Text>
+        <Box style={(t) => ({ width: 1, height: 22, background: t.colorScheme === 'dark' ? t.colors.dark[4] : t.colors.gray[2] })} />
+        <Stack gap={1} style={{ minWidth: 0 }}>
+          <Text fz={12} fw={600} truncate>{selected ? `${selected.dayGroup} · ${selected.atLabel}` : '—'}</Text>
+          <Text fz={10.5} c="dimmed">selected version</Text>
+        </Stack>
+
+        <Box style={{ flex: 1 }} />
+
+        {/* diff cluster */}
+        <Group gap={2} p={3} wrap="nowrap" style={(t) => ({ background: t.colorScheme === 'dark' ? t.colors.dark[6] : '#f4f6f8', borderRadius: 10 })}>
+          <Button variant="white" size="compact-sm" onClick={() => onToggleHighlight(!highlightChanges)}
+            leftSection={<Switch checked={highlightChanges} onChange={() => {}} size="xs" tabIndex={-1} styles={{ track: { cursor: 'pointer' } }} />}
+            styles={{ root: { boxShadow: '0 1px 2px rgba(0,0,0,.08)' } }}>
+            Highlight changes
+          </Button>
+          {changeNav && (
+            <>
+              <Text fz={12} fw={600} c="dimmed" px={6}>{changeNav.index} / {changeNav.total}</Text>
+              <Box style={{ display: 'flex' }}>
+                <ActionIcon variant="subtle" color="gray" size="lg" disabled={!highlightChanges} onClick={changeNav.onPrev} style={{ width: 26 }}>↑</ActionIcon>
+                <ActionIcon variant="subtle" color="gray" size="lg" disabled={!highlightChanges} onClick={changeNav.onNext} style={{ width: 26 }}>↓</ActionIcon>
+              </Box>
+            </>
+          )}
+        </Group>
+
+        <Box style={(t) => ({ width: 1, height: 22, background: t.colorScheme === 'dark' ? t.colors.dark[4] : t.colors.gray[2] })} />
+
+        <Button color="blue" onClick={restore} loading={restoring} disabled={!selected || selected.isCurrent}>
+          Restore this version
+        </Button>
+        <ActionIcon variant="subtle" color="gray" size="lg" onClick={onClose}>✕</ActionIcon>
+      </Group>
+
+      {/* ── body ── */}
+      <Box style={{ flex: 1, display: 'flex', minHeight: 0 }}>
+        {/* left: nav panel */}
+        <Stack gap={0} style={(t) => ({ flex: 'none', width: 280, minHeight: 0, borderRight: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}`, background: t.colorScheme === 'dark' ? t.colors.dark[7] : '#fcfcfd' })}>
+          <Group h={44} px={16} style={(t) => ({ flex: 'none', borderBottom: `1px solid ${t.colorScheme === 'dark' ? t.colors.dark[5] : t.colors.gray[1]}` })}>
+            <Switch checked={onlySaved} onChange={(e) => onToggleOnlySaved(e.currentTarget.checked)} size="sm" label="Only saved" labelPosition="right" styles={{ label: { fontSize: 13, fontWeight: 500 } }} />
+          </Group>
+
+          {calendar && <MiniCalendar cal={calendar} />}
+
+          <ScrollArea style={{ flex: 1 }}>
+            {groups.map((g) => (
+              <Box key={g.head}>
+                <Text px={16} pt={9} pb={5} fz={10.5} fw={600} tt="uppercase" c="dimmed" style={{ letterSpacing: '.05em', position: 'sticky', top: 0, zIndex: 2, background: 'var(--mantine-color-body)' }}>
+                  {g.head}
+                </Text>
+                {g.items.map((r) => (
+                  <RevisionRow key={r.id} rev={r} selected={r.id === selectedId} onSelect={() => onSelect(r.id)} />
+                ))}
+              </Box>
+            ))}
+          </ScrollArea>
+        </Stack>
+
+        {/* right: rendered version */}
+        <ScrollArea style={{ flex: 1, minWidth: 0 }} bg="var(--mantine-color-default-hover)">
+          {isEmpty ? (
+            <Stack align="center" justify="center" h="100%" gap={6} p={40}>
+              <Text fw={600} fz={14}>No earlier versions</Text>
+              <Text fz={12.5} c="dimmed" ta="center">У страницы пока одна версия — сравнивать не с чем.</Text>
+            </Stack>
+          ) : (
+            <Box p="26px 0" maw={660} mx="auto">
+              {renderVersion(selected)}
+            </Box>
+          )}
+        </ScrollArea>
+      </Box>
+    </Modal>
+  );
+}
+
+export default PageHistoryModal;

NewDesign/TimeWorkedModal.tsx

@@ -0,0 +1,234 @@
+/**
+ * TimeWorkedModal — редизайн окна «Time worked on this article».
+ * Mantine v7. Light/dark.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * ЧТО ЭТО
+ * ─────────────────────────────────────────────────────────────────────────
+ * Сводка трудозатрат по дням в виде суточных таймлайнов. Главное отличие от
+ * старой версии: на барах ВИДНО время суток. Реализовано двумя способами
+ * (проп `axis`):
+ *   - 'grid'   — ось часов 00–06–12–18–24 сверху + вертикальные деления,
+ *                ночные часы (0–6, 21–24) слегка затемнены. По умолчанию.
+ *   - 'phases' — цветные полосы фаз дня за блоками (Ночь/Утро/День/Вечер),
+ *                период суток читается сразу, без счёта делений.
+ *
+ * Блоки позиционируются по времени: left = start/24, width = dur/24.
+ * Интенсивность (opacity) блока = длительность/плотность работы, чтобы
+ * очень короткие сессии не терялись (минимальная видимая ширина задана).
+ * Hover-тултип на блоке: начало–конец · длительность.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * УСТАНОВКА / ИСПОЛЬЗОВАНИЕ
+ * ─────────────────────────────────────────────────────────────────────────
+ *   import { TimeWorkedModal, DaySummary } from './TimeWorkedModal';
+ *
+ *   <TimeWorkedModal
+ *     opened={open}
+ *     onClose={() => setOpen(false)}
+ *     totalLabel="≈ 34h"
+ *     agentLabel="≈ 1h 20m"          // undefined → строку agent не показываем
+ *     days={days}                     // DaySummary[]
+ *     axis="grid"                     // 'grid' | 'phases'
+ *     tz="Europe/Moscow"
+ *     inactivityGapMin={15}
+ *   />
+ *
+ * Требует MantineProvider на корне.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * ФОРМАТ ДАННЫХ
+ * ─────────────────────────────────────────────────────────────────────────
+ *   interface Block {
+ *     start: number;   // час начала в сутках, 0..24 (напр. 13.7 = 13:42)
+ *     end: number;     // час конца
+ *     kind: 'work' | 'agent';
+ *   }
+ *   interface DaySummary {
+ *     label: string;       // «Mon 29 Jun»
+ *     totalLabel: string;  // «1h 24m» | «—» для пустого дня
+ *     blocks: Block[];     // [] → пустой день (полупрозрачная дорожка, итог «—»)
+ *     isToday?: boolean;   // сегодня — неполные сутки (рисуем границу «сейчас»)
+ *     nowFraction?: number;// 0..1 позиция «сейчас» для isToday
+ *   }
+ *
+ * Данные уже есть в бэкенде трудозатрат: сессии с началом/концом + тип
+ * (work/agent). Часы = локальные к tz. Изменений API не требуется.
+ *
+ * ─────────────────────────────────────────────────────────────────────────
+ * СОСТОЯНИЯ (нарисованы/поддержаны)
+ * ─────────────────────────────────────────────────────────────────────────
+ *  - День: с активностью (work+agent) / только work / только agent / пустой (—).
+ *  - Сегодня: граница «сейчас» на дорожке (nowFraction).
+ *  - Короткий блок: минимальная ширина, не исчезает.
+ *  - Пустая панель целиком: нет трудозатрат → EmptyState.
+ *  - Длинный период: вертикальный скролл списка дней, шапка/легенда липкие.
+ *  - Тёмная тема: токены Mantine.
+ */
+import { Modal, Box, Group, Stack, Text, ActionIcon, ScrollArea, Tooltip, useMantineColorScheme } from '@mantine/core';
+
+/* ─────────────────────────── Типы ─────────────────────────── */
+
+export interface Block { start: number; end: number; kind: 'work' | 'agent'; }
+export interface DaySummary {
+  label: string;
+  totalLabel: string;
+  blocks: Block[];
+  isToday?: boolean;
+  nowFraction?: number;
+}
+export interface TimeWorkedModalProps {
+  opened: boolean;
+  onClose: () => void;
+  totalLabel: string;
+  agentLabel?: string;
+  days: DaySummary[];
+  axis?: 'grid' | 'phases';
+  tz?: string;
+  inactivityGapMin?: number;
+}
+
+const WORK = '#3b82f6';
+const AGENT = '#c026d3';
+
+const PHASES = [
+  { name: 'Ночь', s: 0, e: 6, bg: 'rgba(99,102,241,.10)', lg: '#e8eaff', fg: '#5b60c9' },
+  { name: 'Утро', s: 6, e: 12, bg: 'rgba(245,159,0,.10)', lg: '#fff2dc', fg: '#b5820e' },
+  { name: 'День', s: 12, e: 18, bg: 'rgba(56,178,172,.10)', lg: '#dcf5f2', fg: '#1a857d' },
+  { name: 'Вечер', s: 18, e: 24, bg: 'rgba(139,92,246,.11)', lg: '#efe7ff', fg: '#6d43c0' },
+];
+
+/* ─────────────────────────── Блок активности ─────────────────────────── */
+
+function fmtHour(h: number) {
+  const hh = Math.floor(h);
+  const mm = Math.round((h - hh) * 60);
+  return `${String(hh).padStart(2, '0')}:${String(mm).padStart(2, '0')}`;
+}
+function fmtDur(h: number) {
+  const total = Math.round(h * 60);
+  const hh = Math.floor(total / 60), mm = total % 60;
+  return hh ? `${hh}h ${mm}m` : `${mm}m`;
+}
+
+function ActivityBlock({ b }: { b: Block }) {
+  const left = (b.start / 24) * 100;
+  const width = Math.max(((b.end - b.start) / 24) * 100, 0.6);
+  const dur = b.end - b.start;
+  const opacity = dur < 0.16 ? 0.5 : dur < 0.35 ? 0.8 : 1;
+  return (
+    <Tooltip label={`${fmtHour(b.start)} – ${fmtHour(b.end)} · ${fmtDur(dur)}`} withArrow openDelay={120} fz={11}>
+      <Box style={{
+        position: 'absolute', top: 5, height: 14, left: `${left}%`, width: `${width}%`,
+        borderRadius: 2, background: b.kind === 'agent' ? AGENT : WORK, opacity, cursor: 'default',
+      }} />
+    </Tooltip>
+  );
+}
+
+/* ─────────────────────────── Дорожка дня ─────────────────────────── */
+
+function DayTrack({ d, axis, dark }: { d: DaySummary; axis: 'grid' | 'phases'; dark: boolean }) {
+  const trackBg = axis === 'grid'
+    ? 'linear-gradient(90deg, rgba(90,100,130,.10) 0 25%, rgba(90,100,130,.02) 25% 87.5%, rgba(90,100,130,.10) 87.5% 100%)'
+    : (dark ? 'rgba(255,255,255,.04)' : '#f6f8fa');
+
+  return (
+    <Group gap={0} h={30} wrap="nowrap">
+      <Text style={{ flex: 'none', width: 92 }} fz={13} c="dimmed">{d.label}</Text>
+      <Box style={{ position: 'relative', flex: 1, height: 24, margin: '0 4px', borderRadius: 5, overflow: 'hidden', background: trackBg }}>
+        {/* фон: полосы фаз или деления сетки */}
+        {axis === 'phases'
+          ? PHASES.map((ph) => (
+              <Box key={ph.name} style={{ position: 'absolute', top: 0, bottom: 0, left: `${(ph.s / 24) * 100}%`, width: `${((ph.e - ph.s) / 24) * 100}%`, background: ph.bg }} />
+            ))
+          : [25, 50, 75].map((p) => (
+              <Box key={p} style={{ position: 'absolute', top: 0, bottom: 0, left: `${p}%`, width: 1, background: 'rgba(120,130,150,.16)' }} />
+            ))}
+        {/* блоки */}
+        {d.blocks.map((b, i) => <ActivityBlock key={i} b={b} />)}
+        {/* граница «сейчас» для сегодняшнего дня */}
+        {d.isToday && d.nowFraction != null && (
+          <Box style={{ position: 'absolute', top: 0, bottom: 0, left: `${d.nowFraction * 100}%`, width: 2, background: '#fa5252' }} />
+        )}
+      </Box>
+      <Text style={{ flex: 'none', width: 64, textAlign: 'right' }} fz={12.5} fw={500} c={d.totalLabel === '—' ? 'dimmed' : undefined}>
+        {d.totalLabel}
+      </Text>
+    </Group>
+  );
+}
+
+/* ─────────────────────────── Окно ─────────────────────────── */
+
+export function TimeWorkedModal(props: TimeWorkedModalProps) {
+  const { opened, onClose, totalLabel, agentLabel, days, axis = 'grid', tz = 'Europe/Moscow', inactivityGapMin = 15 } = props;
+  const { colorScheme } = useMantineColorScheme();
+  const dark = colorScheme === 'dark';
+  const isEmpty = days.length === 0 || days.every((d) => d.blocks.length === 0);
+
+  return (
+    <Modal opened={opened} onClose={onClose} withCloseButton={false} size="46rem" radius="lg" padding={0}
+      overlayProps={{ backgroundOpacity: 0.5, blur: 1 }}>
+      <Box p="22px 24px 20px">
+        {/* header */}
+        <Group mb={14}>
+          <Text fz={17} fw={600}>Time worked on this article</Text>
+          <Box style={{ flex: 1 }} />
+          <ActionIcon variant="subtle" color="gray" size="lg" onClick={onClose}>✕</ActionIcon>
+        </Group>
+
+        {/* summary */}
+        <Group align="baseline" gap={16} mb={axis === 'grid' ? 8 : 16}>
+          <Text fz={22} fw={700}>{totalLabel}</Text>
+          {agentLabel && <Text fz={13} c="dimmed">agent: {agentLabel}</Text>}
+        </Group>
+
+        {/* legend (grid only) */}
+        {axis === 'grid' && (
+          <Group gap={16} mb={16}>
+            <Group gap={6}><Box w={12} h={12} style={{ borderRadius: 3, background: WORK }} /><Text fz={12} fw={500} c="dimmed">Work</Text></Group>
+            <Group gap={6}><Box w={12} h={12} style={{ borderRadius: 3, background: AGENT }} /><Text fz={12} fw={500} c="dimmed">Agent</Text></Group>
+          </Group>
+        )}
+
+        {isEmpty ? (
+          <Stack align="center" gap={6} p="48px 20px">
+            <Text fw={600} fz={14}>No time tracked yet</Text>
+            <Text fz={12.5} c="dimmed" ta="center">По этой статье ещё нет трудозатрат.</Text>
+          </Stack>
+        ) : (
+          <>
+            {/* axis header (sticky) */}
+            <Group gap={0} mb={5} wrap="nowrap" style={{ position: 'sticky', top: 0, zIndex: 2, background: 'var(--mantine-color-body)' }}>
+              <Box style={{ flex: 'none', width: 92 }} />
+              {axis === 'grid' ? (
+                <Box style={{ flex: 1, position: 'relative', height: 14, margin: '0 4px' }}>
+                  {[['0%', '00', 'flex-start'], ['25%', '06', 'center'], ['50%', '12', 'center'], ['75%', '18', 'center'], ['100%', '24', 'flex-end']].map(([l, t, al]) => (
+                    <Text key={t as string} fz={10} fw={500} c="dimmed" style={{ position: 'absolute', left: l as string, transform: al === 'center' ? 'translateX(-50%)' : al === 'flex-end' ? 'translateX(-100%)' : undefined }}>{t}</Text>
+                  ))}
+                </Box>
+              ) : (
+                <Box style={{ flex: 1, display: 'flex', height: 16, margin: '0 4px', borderRadius: 4, overflow: 'hidden' }}>
+                  {PHASES.map((ph) => (
+                    <Box key={ph.name} style={{ flex: ph.e - ph.s, display: 'flex', alignItems: 'center', justifyContent: 'center', background: ph.lg, color: ph.fg, font: '600 9.5px system-ui' }}>{ph.name}</Box>
+                  ))}
+                </Box>
+              )}
+              <Box style={{ flex: 'none', width: 64 }} />
+            </Group>
+
+            {/* day rows */}
+            <ScrollArea.Autosize mah="60vh" type="hover">
+              {days.map((d, i) => <DayTrack key={i} d={d} axis={axis} dark={dark} />)}
+            </ScrollArea.Autosize>
+          </>
+        )}
+
+        <Text mt={16} fz={11.5} c="dimmed">Estimate · timezone {tz} · inactivity gap {inactivityGapMin} min</Text>
+      </Box>
+    </Modal>
+  );
+}
+
+export default TimeWorkedModal;

agent-roles-catalog/bundles/assistants/en.yaml

@@ -346,6 +346,9 @@ roles:
         □ Working sections ("Log", "Open Questions", "Revision") are moved to an
           appendix at the end of the document or clearly separated from the report
           body.
+        □ Once the report is complete, call save_page_version to pin the finished
+          document as a restorable named version (a checkpoint the reader can
+          return to). A repeat call after no further edits is a harmless no-op.
       Be honest about gaps. If you couldn't find something, say so — don't disguise
       a guess as a fact.
     autoStart: false
@@ -442,6 +445,7 @@ roles:
       - **The language of the notes = the main language of the call.** Technical terms — in their canonical spelling (usually Latin).
       - **Don't evaluate the participants** and don't comment on the quality of the discussion.
       - The output is the notes only, with no preambles or meta-comments, apart from targeted uncertainty marks.
+      - **Pin the result.** Once the notes are complete on the page, call save_page_version to save them as a restorable named version (a checkpoint). Calling it again after no further edits is a harmless no-op.
 
       ## Style example (excerpt)
 

agent-roles-catalog/bundles/assistants/ru.yaml

@@ -345,6 +345,9 @@ roles:
         □ Working sections («Журнал», «Открытые вопросы», «Ревизия») are moved to
           an appendix at the end of the document or clearly separated from the
           report body.
+        □ Once the report is complete, call save_page_version to pin the finished
+          document as a restorable named version (a checkpoint the reader can
+          return to). A repeat call after no further edits is a harmless no-op.
       Be honest about gaps. If you couldn't find something, say so — don't disguise
       a guess as a fact.
     autoStart: false
@@ -441,6 +444,7 @@ roles:
       - **Язык конспекта = основной язык созвона.** Технические термины — в каноническом написании (обычно латиницей).
       - **Не оценивай участников** и не комментируй качество обсуждения.
       - На выходе — только конспект, без преамбул и мета-комментариев, кроме точечных пометок неуверенности.
+      - **Зафиксируй результат.** Когда конспект готов на странице, вызови save_page_version, чтобы сохранить его как именованную версию (точку восстановления). Повторный вызов без новых правок — безвредный no-op.
 
       ## Пример стиля (фрагмент)
 

agent-roles-catalog/index.yaml

@@ -33,6 +33,6 @@ bundles:
       - en
     roles:
       - slug: researcher
-        version: 9
+        version: 10
       - slug: call-summarizer
-        version: 1
+        version: 2

apps/client/public/locales/en-US/translation.json

@@ -1,4 +1,7 @@
 {
+  "1 year": "1 year",
+  "30 days": "30 days",
+  "90 days": "90 days",
   "A new version is available": "A new version is available",
   "Account": "Account",
   "Active": "Active",
@@ -10,11 +13,16 @@
   "Add space members": "Add space members",
   "Add to favorites": "Add to favorites",
   "Admin": "Admin",
+  "API key created": "API key created",
+  "API key revoked": "API key revoked",
+  "API keys across the workspace.": "API keys across the workspace.",
   "Are you sure you want to delete this group? Members will lose access to resources this group has access to.": "Are you sure you want to delete this group? Members will lose access to resources this group has access to.",
   "Are you sure you want to delete this page?": "Are you sure you want to delete this page?",
   "Are you sure you want to remove this user from the group? The user will lose access to resources this group has access to.": "Are you sure you want to remove this user from the group? The user will lose access to resources this group has access to.",
   "Are you sure you want to remove this user from the space? The user will lose all access to this space.": "Are you sure you want to remove this user from the space? The user will lose all access to this space.",
   "Are you sure you want to restore this version? Any changes not versioned will be lost.": "Are you sure you want to restore this version? Any changes not versioned will be lost.",
+  "Are you sure you want to revoke \"{{name}}\"? Any client using this key will immediately lose access. This cannot be undone.": "Are you sure you want to revoke \"{{name}}\"? Any client using this key will immediately lose access. This cannot be undone.",
+  "Author": "Author",
   "Can become members of groups and spaces in workspace": "Can become members of groups and spaces in workspace",
   "Can create and edit pages in space.": "Can create and edit pages in space.",
   "Can edit": "Can edit",
@@ -33,7 +41,9 @@
   "Confirm": "Confirm",
   "Copy as Markdown": "Copy as Markdown",
   "Copy link": "Copy link",
+  "Copy your API key now and store it somewhere safe. For security reasons it will not be shown again.": "Copy your API key now and store it somewhere safe. For security reasons it will not be shown again.",
   "Create": "Create",
+  "Create API key": "Create API key",
   "Create group": "Create group",
   "Create page": "Create page",
   "Create space": "Create space",
@@ -55,7 +65,14 @@
   "e.g Sales": "e.g Sales",
   "e.g Space for product team": "e.g Space for product team",
   "e.g Space for sales team to collaborate": "e.g Space for sales team to collaborate",
+  "e.g. CI deploy token": "e.g. CI deploy token",
   "Edit": "Edit",
+  "Expiring soon": "Expiring soon",
+  "Failed to create API key": "Failed to create API key",
+  "Failed to load API keys.": "Failed to load API keys.",
+  "Failed to revoke API key": "Failed to revoke API key",
+  "Never used": "Never used",
+  "No API keys yet": "No API keys yet",
   "Read": "Read",
   "Edit group": "Edit group",
   "Email": "Email",
@@ -134,6 +151,8 @@
   "page": "page",
   "Page deleted successfully": "Page deleted successfully",
   "Page history": "Page history",
+  "Revoke API key": "Revoke API key",
+  "Revoke {{name}}": "Revoke {{name}}",
   "Select version": "Select version",
   "Highlight changes": "Highlight changes",
   "Page import is in progress. Please do not close this tab.": "Page import is in progress. Please do not close this tab.",
@@ -189,6 +208,9 @@
   "Template": "Template",
   "Templates": "Templates",
   "Theme": "Theme",
+  "This key expires within 30 days": "This key expires within 30 days",
+  "This key has expired": "This key has expired",
+  "This key never expires": "This key never expires",
   "To change your email, you have to enter your password and new email.": "To change your email, you have to enter your password and new email.",
   "Toggle full page width": "Toggle full page width",
   "Unable to import pages. Please try again.": "Unable to import pages. Please try again.",
@@ -196,6 +218,7 @@
   "Untitled": "Untitled",
   "Updated successfully": "Updated successfully",
   "User": "User",
+  "Within the last hour": "Within the last hour",
   "Workspace": "Workspace",
   "Workspace Name": "Workspace Name",
   "Workspace settings": "Workspace settings",
@@ -240,6 +263,8 @@
   "Comment re-opened successfully": "Comment re-opened successfully",
   "Comment unresolved successfully": "Comment unresolved successfully",
   "Failed to resolve comment": "Failed to resolve comment",
+  "Failed to re-open comment": "Failed to re-open comment",
+  "Comment no longer exists": "Comment no longer exists",
   "Resolve comment": "Resolve comment",
   "Unresolve comment": "Unresolve comment",
   "Resolve Comment Thread": "Resolve Comment Thread",
@@ -424,6 +449,7 @@
   "Write anything. Enter \"/\" for commands": "Write anything. Enter \"/\" for commands",
   "Write...": "Write...",
   "Column count": "Column count",
+  "Your personal API keys.": "Your personal API keys.",
   "{{count}} Columns": "{{count}} Columns",
   "{{count}} command available_one": "1 command available",
   "{{count}} command available_other": "{{count}} commands available",
@@ -1428,5 +1454,20 @@
   "Boundary": "Boundary",
   "Autosave": "Autosave",
   "Only versions": "Only versions",
-  "No saved versions yet.": "No saved versions yet."
+  "No saved versions yet.": "No saved versions yet.",
+  "Time worked on this article": "Time worked on this article",
+  "Show time worked on this page": "Show time worked on this page",
+  "Estimated time worked (inactivity gap {{gap}} min)": "Estimated time worked (inactivity gap {{gap}} min)",
+  "Estimate · timezone {{tz}} · inactivity gap {{gap}} min": "Estimate · timezone {{tz}} · inactivity gap {{gap}} min",
+  "No editing activity recorded yet.": "No editing activity recorded yet.",
+  "× {{count}} days without edits": "× {{count}} days without edits",
+  "agent: {{value}}": "agent: {{value}}",
+  "Work": "Work",
+  "Agent": "Agent",
+  "≈ {{hours}}h {{minutes}}m": "≈ {{hours}}h {{minutes}}m",
+  "≈ {{hours}}h": "≈ {{hours}}h",
+  "≈ {{minutes}}m": "≈ {{minutes}}m",
+  "{{hours}}h {{minutes}}m": "{{hours}}h {{minutes}}m",
+  "{{hours}}h": "{{hours}}h",
+  "{{minutes}}m": "{{minutes}}m"
 }

apps/client/public/locales/ru-RU/translation.json

@@ -1,4 +1,7 @@
 {
+  "1 year": "1 год",
+  "30 days": "30 дней",
+  "90 days": "90 дней",
   "A new version is available": "Доступна новая версия",
   "Account": "Аккаунт",
   "Active": "Активный",
@@ -10,11 +13,16 @@
   "Add space members": "Добавить участников пространства",
   "Add to favorites": "Добавить в избранное",
   "Admin": "Администратор",
+  "API key created": "API ключ создан",
+  "API key revoked": "API ключ отозван",
+  "API keys across the workspace.": "API ключи всего рабочего пространства.",
   "Are you sure you want to delete this group? Members will lose access to resources this group has access to.": "Вы уверены, что хотите удалить эту группу? Участники потеряют доступ к материалам, к которым у этой группы есть доступ.",
   "Are you sure you want to delete this page?": "Вы уверены, что хотите удалить эту страницу?",
   "Are you sure you want to remove this user from the group? The user will lose access to resources this group has access to.": "Вы уверены, что хотите удалить этого пользователя из группы? Пользователь потеряет доступ к материалам, к которым есть доступ у этой группы.",
   "Are you sure you want to remove this user from the space? The user will lose all access to this space.": "Вы уверены, что хотите удалить этого пользователя из пространства? Пользователь потеряет весь доступ к этому пространству.",
   "Are you sure you want to restore this version? Any changes not versioned will be lost.": "Вы уверены, что хотите восстановить эту версию? Все не зафиксированные изменения будут потеряны.",
+  "Are you sure you want to revoke \"{{name}}\"? Any client using this key will immediately lose access. This cannot be undone.": "Вы уверены, что хотите отозвать «{{name}}»? Любой клиент, использующий этот ключ, немедленно потеряет доступ. Это действие нельзя отменить.",
+  "Author": "Автор",
   "Can become members of groups and spaces in workspace": "Могут становиться участниками групп и пространств в рабочей области",
   "Can create and edit pages in space.": "Может создавать и редактировать страницы в пространстве.",
   "Can edit": "Может изменять",
@@ -33,7 +41,9 @@
   "Confirm": "Подтвердить",
   "Copy as Markdown": "Копировать как Markdown",
   "Copy link": "Копировать ссылку",
+  "Copy your API key now and store it somewhere safe. For security reasons it will not be shown again.": "Скопируйте API ключ сейчас и сохраните его в надёжном месте. В целях безопасности он больше не будет показан.",
   "Create": "Создать",
+  "Create API key": "Создать API ключ",
   "Create group": "Создать группу",
   "Create page": "Создать страницу",
   "Create space": "Создать пространство",
@@ -46,6 +56,7 @@
   "Are you sure you want to delete this page? This will delete its children and page history. This action is irreversible.": "Вы уверены, что хотите удалить эту страницу? Это удалит её дочерние страницы, а также историю страницы. Это действие необратимо.",
   "Description": "Описание",
   "Details": "Подробности",
+  "Done": "Готово",
   "e.g ACME": "например, ACME",
   "e.g ACME Inc": "например, ACME Inc",
   "e.g Developers": "например, Developers",
@@ -55,7 +66,15 @@
   "e.g Sales": "например, Sales",
   "e.g Space for product team": "например, Пространство для команды продукта",
   "e.g Space for sales team to collaborate": "например, Пространство для совместной работы команды продаж",
+  "e.g. CI deploy token": "например, токен деплоя CI",
   "Edit": "Редактировать",
+  "Expiring soon": "Скоро истекает",
+  "Failed to create API key": "Не удалось создать API ключ",
+  "Failed to load API keys.": "Не удалось загрузить API ключи.",
+  "Failed to revoke API key": "Не удалось отозвать API ключ",
+  "Name is required": "Имя обязательно",
+  "Never used": "Не использовался",
+  "No API keys yet": "Пока нет API ключей",
   "Read": "Чтение",
   "Edit group": "Редактировать группу",
   "Email": "Электронная почта",
@@ -134,6 +153,8 @@
   "page": "страница",
   "Page deleted successfully": "Страница успешно удалена",
   "Page history": "История страницы",
+  "Revoke API key": "Отозвать API ключ",
+  "Revoke {{name}}": "Отозвать {{name}}",
   "Select version": "Выбрать версию",
   "Highlight changes": "Выделить изменения",
   "Page import is in progress. Please do not close this tab.": "Импорт страницы в процессе. Пожалуйста, не закрывайте эту вкладку.",
@@ -189,6 +210,9 @@
   "Template": "Шаблон",
   "Templates": "Шаблоны",
   "Theme": "Тема",
+  "This key expires within 30 days": "Этот ключ истекает в течение 30 дней",
+  "This key has expired": "Этот ключ истек",
+  "This key never expires": "Этот ключ никогда не истекает",
   "To change your email, you have to enter your password and new email.": "Чтобы изменить электронную почту, вам нужно ввести пароль и новый адрес.",
   "Toggle full page width": "Переключить полную ширину страницы",
   "Unable to import pages. Please try again.": "Не удалось импортировать страницы. Пожалуйста, попробуйте ещё раз.",
@@ -196,6 +220,7 @@
   "Untitled": "Без названия",
   "Updated successfully": "Успешно обновлено",
   "User": "Пользователь",
+  "Within the last hour": "За последний час",
   "Workspace": "Рабочее пространство",
   "Workspace Name": "Название рабочего пространства",
   "Workspace settings": "Настройки рабочего пространства",
@@ -240,6 +265,8 @@
   "Comment re-opened successfully": "Комментарий успешно открыт повторно",
   "Comment unresolved successfully": "Комментарий успешно переведён в нерешённые",
   "Failed to resolve comment": "Не удалось разрешить комментарий",
+  "Failed to re-open comment": "Не удалось переоткрыть комментарий",
+  "Comment no longer exists": "Комментарий больше не существует",
   "Resolve comment": "Решить комментарий",
   "Unresolve comment": "Снять статус решённого с комментария",
   "Resolve Comment Thread": "Решить ветку комментариев",
@@ -430,6 +457,7 @@
   "Write anything. Enter \"/\" for commands": "Пишите что угодно. Введите \"/\" для команд",
   "Write...": "Напишите...",
   "Column count": "Количество столбцов",
+  "Your personal API keys.": "Ваши личные API ключи.",
   "{{count}} Columns": "{count, plural, one{# столбец} few{# столбца} many{# столбцов} other{# столбца}}",
   "{{count}} command available_one": "Доступна 1 команда",
   "{{count}} command available_other": "Доступно {{count}} команд",
@@ -1443,5 +1471,20 @@
   "Boundary": "Граница",
   "Autosave": "Автосейв",
   "Only versions": "Только версии",
-  "No saved versions yet.": "Пока нет сохранённых версий."
+  "No saved versions yet.": "Пока нет сохранённых версий.",
+  "Time worked on this article": "Время работы над статьёй",
+  "Show time worked on this page": "Показать время работы над страницей",
+  "Estimated time worked (inactivity gap {{gap}} min)": "Оценка времени работы (порог паузы {{gap}} мин)",
+  "Estimate · timezone {{tz}} · inactivity gap {{gap}} min": "Оценка · таймзона {{tz}} · порог паузы {{gap}} мин",
+  "No editing activity recorded yet.": "Правок пока нет.",
+  "× {{count}} days without edits": "× {{count}} дн. без правок",
+  "agent: {{value}}": "агент: {{value}}",
+  "Work": "Работа",
+  "Agent": "Агент",
+  "≈ {{hours}}h {{minutes}}m": "≈ {{hours}} ч {{minutes}} мин",
+  "≈ {{hours}}h": "≈ {{hours}} ч",
+  "≈ {{minutes}}m": "≈ {{minutes}} мин",
+  "{{hours}}h {{minutes}}m": "{{hours}} ч {{minutes}} м",
+  "{{hours}}h": "{{hours}} ч",
+  "{{minutes}}m": "{{minutes}} м"
 }

apps/client/src/App.tsx

@@ -44,6 +44,12 @@ const AccountSettings = lazy(
 const AccountPreferences = lazy(
   () => import("@/pages/settings/account/account-preferences.tsx"),
 );
+// #506 — lazy leaf (own chunk): the API-keys management page is route-split so
+// its code (Mantine table/modals + the create/revoke flow) stays out of the
+// entry bundle (post-#342 bundle discipline).
+const AccountApiKeys = lazy(
+  () => import("@/pages/settings/account/account-api-keys.tsx"),
+);
 const WorkspaceSettings = lazy(
   () => import("@/pages/settings/workspace/workspace-settings"),
 );
@@ -105,6 +111,7 @@ export default function App() {
               path={"account/preferences"}
               element={<AccountPreferences />}
             />
+            <Route path={"account/api-keys"} element={<AccountApiKeys />} />
             <Route path={"workspace"} element={<WorkspaceSettings />} />
             <Route path={"ai"} element={<AiSettings />} />
             <Route path={"members"} element={<WorkspaceMembers />} />

apps/client/src/components/settings/settings-sidebar.tsx

@@ -10,6 +10,7 @@ import {
   IconBrush,
   IconWorld,
   IconSparkles,
+  IconKey,
 } from "@tabler/icons-react";
 import { Link, useLocation } from "react-router-dom";
 import classes from "./settings.module.css";
@@ -46,6 +47,11 @@ const groupedData: DataGroup[] = [
         icon: IconBrush,
         path: "/settings/account/preferences",
       },
+      {
+        label: "API keys",
+        icon: IconKey,
+        path: "/settings/account/api-keys",
+      },
     ],
   },
   {

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

@@ -1254,4 +1254,88 @@ describe("ChatThread — live reconnect + stalled", () => {
     });
     expect(onResumeFallback).toHaveBeenCalledWith(false); // POLL_IDLE_CAP -> idle -> disarm
   });
+
+  it("#541: getRun HANGS on a live disconnect — the timeout race still enters reconnect (no silent freeze in `streaming`)", async () => {
+    // MUTATION-VERIFY: revert the race to a bare `getRun(cid).then/.catch` and this
+    // reddens — a HUNG (never-settling, NOT rejected) getRun leaves the FSM stuck in
+    // `streaming` with no reconnect banner and no poll (the axios client sets no
+    // request timeout, and the stalled-idle cap only arms AFTER reconnecting/polling).
+    renderLive();
+    h.state.getRun.mockReset();
+    h.state.getRun.mockReturnValue(new Promise(() => {})); // getRun HANGS forever
+    await disconnect(); // live partial = liveMsg (id "a2")
+    expect(h.state.getRun).toHaveBeenCalledWith("c1");
+    // BEFORE the bound fires the FSM is still in the live turn — the very freeze bug.
+    expect(screen.queryByText(/reconnecting/i)).toBeNull();
+    // The recovery-start bound fires -> the SAME fallback as the reject path.
+    act(() => {
+      vi.advanceTimersByTime(4_000);
+    });
+    expect(screen.getByText(/reconnecting/i)).toBeTruthy();
+    // The live partial (a2) was DROPPED from the store (replay-from-start, no stale
+    // tail-apply base). Mirrors the getRun-REJECT test's inspection.
+    const removedLivePartial = (
+      h.state.setMessages as unknown as {
+        mock: { calls: [unknown][] };
+      }
+    ).mock.calls.some(([updater]) => {
+      if (typeof updater !== "function") return false;
+      const out = (updater as (p: { id: string }[]) => { id: string }[])([
+        { id: "a2" },
+        { id: "u1" },
+      ]);
+      return !out.some((m) => m.id === "a2");
+    });
+    expect(removedLivePartial).toBe(true);
+    // ...and the anchor was NULLED -> replay-from-start (no ?anchor=&n= over the partial).
+    advanceToAttempt(1);
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream",
+    );
+  });
+
+  it("#541: a getRun resolve AFTER the timeout already fired is IGNORED (no double reconnect, no stale re-seed)", async () => {
+    // The timeout wins first and enters the ladder via replay-from-start. When the
+    // hung getRun FINALLY answers, its late `.then` must be a full no-op: it must not
+    // re-seed the store from the (now stale) persisted row, must not re-set the
+    // anchor, and must not re-enter reconnect. The local `settled` flag makes the
+    // resolve/reject/timeout branches mutually exclusive.
+    // MUTATION-VERIFY: drop the `settled` guard (let the late `.then` run) and the
+    // late re-seed re-sets the anchor (URL regains ?anchor=a2&n=3) + calls setMessages.
+    renderLive();
+    let resolveGetRun!: (v: unknown) => void;
+    h.state.getRun.mockReset();
+    h.state.getRun.mockReturnValue(
+      new Promise((r) => {
+        resolveGetRun = r;
+      }),
+    );
+    await disconnect();
+    act(() => {
+      vi.advanceTimersByTime(4_000); // bound fires -> replay-from-start, reconnecting
+    });
+    expect(screen.getByText(/reconnecting/i)).toBeTruthy();
+    advanceToAttempt(1);
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+    // Anchor is null -> replay-from-start URL (the pre-condition the late resolve must
+    // not undo).
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream",
+    );
+    const setMessagesCallsBefore = h.state.setMessages.mock.calls.length;
+    // NOW the hung getRun finally resolves with a persisted anchor (id a2, steps 3).
+    await act(async () => {
+      resolveGetRun(persistedAnchor());
+      await Promise.resolve();
+    });
+    // The late resolve did NOT re-seed the store...
+    expect(h.state.setMessages.mock.calls.length).toBe(setMessagesCallsBefore);
+    // ...did NOT re-set the anchor (URL stays replay-from-start, no ?anchor=&n=)...
+    expect(h.state.transport!.prepareReconnectToStreamRequest!().api).toBe(
+      "/api/ai-chat/runs/c1/stream",
+    );
+    // ...and did NOT trigger a fresh reconnect attach.
+    expect(h.state.resumeStream).toHaveBeenCalledTimes(1);
+  });
 });

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

@@ -81,6 +81,17 @@ const STREAM_THROTTLE_MS = 50;
 // THREAD now (the FSM owns polling->stalled); the window just polls while armed.
 const DEGRADED_POLL_IDLE_MAX_MS = 10 * 60_000;
 
+// #541: the bound on the persist re-seed (getRun) wait when ENTERING the reconnect
+// ladder after a live SSE drop. The axios client (lib/api-client.ts) sets NO request
+// timeout, and the stalled-idle cap only arms AFTER the FSM enters reconnecting/
+// polling — which never happens if getRun HANGS (connection established, no response).
+// Without this bound a live drop + a hung getRun sticks the FSM in `streaming` with
+// no banner and no poll until the browser socket timeout (minutes) — the silent-freeze
+// class #497 eliminates. This is a deliberate RECOVERY-START bound (drop the live
+// partial + enter the ladder so the poll/idle-cap arms), intentionally much shorter
+// than any network socket timeout — not a network read timeout.
+const RECONNECT_RESEED_TIMEOUT_MS = 4_000;
+
 /** The #487 active (non-terminal) run statuses — mirrors the server's
  *  ACTIVE_RUN_STATUSES. A run-fact is "active" only for these. */
 function isActiveRunStatus(status: string | null | undefined): boolean {
@@ -286,6 +297,9 @@ export default function ChatThread({
   // stalled inactivity cap. Cleared by the cancelReconnect effect / the cap effect.
   const reconnectTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
   const idleCapTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  // #541: the persist re-seed (getRun) timeout race on the disconnect->reconnect
+  // entry. Held in a ref so unmount (DISPOSE cleanup) clears it — no dangling timer.
+  const reseedTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   // #491 tail-only: seed EVERY persisted row unchanged (no strip). The streaming
   // tail holds steps 0..N-1; the run-stream registry's tail (steps >= N) is APPENDED
@@ -748,37 +762,68 @@ export default function ChatThread({
           anchorRef.current = null;
         };
         if (cid) {
-          void getRun(cid)
-            .then((res) => {
-              if (!mountedRef.current) return;
-              const persisted = res.message;
-              if (persisted && persisted.role === "assistant") {
-                anchorRef.current = {
-                  id: persisted.id,
-                  stepsPersisted: stepsPersistedOf(persisted),
-                };
-                // Replace the live partial with the persisted row IN PLACE by id —
-                // the re-seed from persist. The attach's tail (steps >= N) then
-                // appends to a store holding EXACTLY steps 0..N-1: no duplication.
-                setMessages((prev) => mergeById(prev, rowToUiMessage(persisted)));
-              } else {
-                // No persisted assistant row (pre-first-frame break): drop the live
-                // partial + replay from start (no anchor/n) so nothing is duplicated.
-                dropLivePartialAndReplayFromStart();
-              }
-              enterReconnect(res.run?.id ?? runId);
-            })
-            .catch(() => {
-              if (!mountedRef.current) return;
-              // Persist read FAILED: we cannot re-seed from fresh persist, and a
-              // stale mount-time anchor over the live partial would tail-apply
-              // already-present steps -> duplication (a flaky-network blip:
-              // SSE + getRun both fail, network recovers in ~1s, the registry still
-              // covers from the mount frontier). Restore the removed-filter guarantee
-              // instead: drop the live partial + replay from start / 204 -> poll.
+          // #541: bound the persist re-seed wait with a timeout race. getRun goes
+          // through the axios client, which has NO request timeout; a HUNG getRun
+          // (connection open, no response) — distinct from a REJECT, which the
+          // `.catch` already handles — would otherwise never let us enter the ladder,
+          // leaving the FSM stuck in `streaming` with no banner and no poll until the
+          // browser socket timeout. `settled` makes the three branches (resolve /
+          // reject / timeout) mutually exclusive: whichever fires FIRST wins and the
+          // others become no-ops. So a LATE getRun resolve AFTER the timeout is fully
+          // ignored — it cannot (i) re-enter reconnect a second time, (ii) overwrite
+          // the timeout's replay-from-start with a stale re-seed, or (iii) re-arm any
+          // timer. On timeout we take the SAME fallback as the reject path (drop the
+          // live partial + enter the ladder, so the poll and the stalled-idle cap arm).
+          let settled = false;
+          const finishReseed = (apply: () => void): void => {
+            if (settled) return;
+            settled = true;
+            if (reseedTimerRef.current) {
+              clearTimeout(reseedTimerRef.current);
+              reseedTimerRef.current = null;
+            }
+            if (!mountedRef.current) return;
+            apply();
+          };
+          reseedTimerRef.current = setTimeout(() => {
+            finishReseed(() => {
               dropLivePartialAndReplayFromStart();
               enterReconnect(runId);
             });
+          }, RECONNECT_RESEED_TIMEOUT_MS);
+          void getRun(cid)
+            .then((res) => {
+              finishReseed(() => {
+                const persisted = res.message;
+                if (persisted && persisted.role === "assistant") {
+                  anchorRef.current = {
+                    id: persisted.id,
+                    stepsPersisted: stepsPersistedOf(persisted),
+                  };
+                  // Replace the live partial with the persisted row IN PLACE by id —
+                  // the re-seed from persist. The attach's tail (steps >= N) then
+                  // appends to a store holding EXACTLY steps 0..N-1: no duplication.
+                  setMessages((prev) => mergeById(prev, rowToUiMessage(persisted)));
+                } else {
+                  // No persisted assistant row (pre-first-frame break): drop the live
+                  // partial + replay from start (no anchor/n) so nothing is duplicated.
+                  dropLivePartialAndReplayFromStart();
+                }
+                enterReconnect(res.run?.id ?? runId);
+              });
+            })
+            .catch(() => {
+              finishReseed(() => {
+                // Persist read FAILED: we cannot re-seed from fresh persist, and a
+                // stale mount-time anchor over the live partial would tail-apply
+                // already-present steps -> duplication (a flaky-network blip:
+                // SSE + getRun both fail, network recovers in ~1s, the registry still
+                // covers from the mount frontier). Restore the removed-filter guarantee
+                // instead: drop the live partial + replay from start / 204 -> poll.
+                dropLivePartialAndReplayFromStart();
+                enterReconnect(runId);
+              });
+            });
         } else {
           dropLivePartialAndReplayFromStart();
           enterReconnect(runId);
@@ -903,6 +948,12 @@ export default function ChatThread({
     }
     return () => {
       mountedRef.current = false;
+      // #541: clear the in-flight persist re-seed timeout (not an FSM effect timer,
+      // so DISPOSE does not touch it) — no dangling setTimeout after unmount.
+      if (reseedTimerRef.current) {
+        clearTimeout(reseedTimerRef.current);
+        reseedTimerRef.current = null;
+      }
       dispatch({ type: "DISPOSE" }); // aborts attach + timers, bumps epoch (I5)
     };
     // Mount-only by design; the parent remounts per chat via `key`.

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

@@ -27,6 +27,7 @@ vi.mock("@/features/ai-chat/utils/markdown.ts", async () => {
 
 import MessageItem from "./message-item";
 import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
+import { splitPlainChunks } from "./streaming-plain-text";
 
 // matchMedia (read by MantineProvider) is stubbed globally in vitest.setup.ts.
 
@@ -114,3 +115,89 @@ describe("MessageItem markdown memoization", () => {
     expect(queryByText("streamed answer")).not.toBeNull();
   });
 });
+
+// PERF SMOKE (#492): the whole point of the incremental streaming render is that
+// the ANSWER path costs O(number of markdown blocks), NOT O(number of throttled
+// ~20Hz ticks). Pre-#492 the finalized MarkdownPart re-parsed the WHOLE growing
+// answer on every delta — a synthetic ~100 KB stream measured 394 renderChatMarkdown
+// calls (one per tick). With the incremental render each STABILIZED block is parsed
+// exactly once (memoized in MarkdownChunk) and the live tail is cheap plain text, so
+// the call count collapses to ~= the block count regardless of tick granularity.
+describe("MessageItem streaming answer render is O(blocks), not O(ticks)", () => {
+  // ~100 KB answer. Each section is a heading + a paragraph — TWO blank-line
+  // delimited markdown blocks — so the safe-cut block count is ~2× the section
+  // count. The perf claim is about the BLOCK count (the memoization granularity),
+  // measured directly with splitPlainChunks below, not the section count.
+  const buildAnswer = () => {
+    const SECTIONS = 100;
+    const paragraphs: string[] = [];
+    for (let i = 0; i < SECTIONS; i++) {
+      paragraphs.push(`## Section ${i}\n\n` + "lorem ipsum dolor ".repeat(55));
+    }
+    const full = paragraphs.join("\n\n");
+    // The number of memoized markdown blocks the incremental render splits into
+    // (all but the live tail are parsed once each).
+    return { full, blocks: splitPlainChunks(full).length };
+  };
+
+  const streamMsg = (text: string, state: "streaming" | "done"): UIMessage =>
+    ({
+      id: "m1",
+      role: "assistant",
+      parts: [{ type: "text", text, state }],
+    }) as UIMessage;
+
+  it("parses each block ~once over a 100KB stream (≈blocks, ≪ ticks)", () => {
+    renderChatMarkdownSpy.mockClear();
+    const { full, blocks } = buildAnswer();
+    const CHUNK = 128; // a realistic ~20Hz throttled delta size
+    const ticks = Math.ceil(full.length / CHUNK);
+
+    let msg = streamMsg(full.slice(0, CHUNK), "streaming");
+    const { rerender } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg}
+          signature={messageSignature(msg)}
+          turnStreaming
+        />
+      </MantineProvider>,
+    );
+    for (let end = 2 * CHUNK; end < full.length; end += CHUNK) {
+      msg = streamMsg(full.slice(0, end), "streaming");
+      rerender(
+        <MantineProvider>
+          <MessageItem
+            message={msg}
+            signature={messageSignature(msg)}
+            turnStreaming
+          />
+        </MantineProvider>,
+      );
+    }
+    // Finalize: the streaming→done flip renders the whole answer through ONE
+    // canonical pass (visual parity), so the finished DOM matches the pre-#492
+    // output. This is the single extra parse on top of the per-block ones.
+    const done = streamMsg(full, "done");
+    rerender(
+      <MantineProvider>
+        <MessageItem message={done} signature={messageSignature(done)} />
+      </MantineProvider>,
+    );
+
+    const calls = renderChatMarkdownSpy.mock.calls.length;
+    // Sanity: the stream really had far more ticks than blocks (else the test is
+    // vacuous — the point is that calls scale with blocks, not ticks).
+    expect(ticks).toBeGreaterThan(blocks * 3);
+    // O(blocks): each stabilized block parsed once + the single final whole-text
+    // parse. A small constant absorbs the finalize render and the live-tail block;
+    // the load-bearing claim is the bound below.
+    expect(calls).toBeLessThanOrEqual(blocks + 2);
+    // ≪ ticks — and, non-vacuously, the blocks WERE parsed (not skipped entirely).
+    expect(calls).toBeLessThan(ticks / 3);
+    expect(calls).toBeGreaterThan(blocks / 2);
+    // MUTATION-VERIFY (documented, not run here): dropping the `memo()` wrapper on
+    // MarkdownChunk (so every stable block re-parses each tick) drives `calls`
+    // toward `ticks` (~394), reddening both upper-bound assertions above.
+  });
+});

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

@@ -0,0 +1,112 @@
+import { describe, expect, it, vi } from "vitest";
+import { render } from "@testing-library/react";
+import { MantineProvider } from "@mantine/core";
+import type { UIMessage } from "@ai-sdk/react";
+
+// Stub react-i18next (the component reads `useTranslation`). Mirrors the other
+// message-item specs.
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (key: string) => key }),
+}));
+
+import MessageItem from "./message-item";
+import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
+// The REAL canonical renderer (NOT the spy the memo test installs): this file
+// exercises the actual markdown output so the visual-regression assertions below
+// compare against genuine HTML (incl. the schema's `<li><p>` wrappers).
+import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
+import classes from "./ai-chat.module.css";
+
+const msg = (
+  parts: UIMessage["parts"],
+  extra?: Partial<UIMessage>,
+): UIMessage =>
+  ({ id: "m1", role: "assistant", parts, ...extra }) as UIMessage;
+
+const renderRow = (message: UIMessage, turnStreaming = false) =>
+  render(
+    <MantineProvider>
+      <MessageItem
+        message={message}
+        signature={messageSignature(message)}
+        turnStreaming={turnStreaming}
+      />
+    </MantineProvider>,
+  );
+
+// A rich multi-block answer that exercises headings, a list (the `<li><p>` case
+// the scoped CSS tightens), inline emphasis, and multiple paragraphs.
+const ANSWER = [
+  "# Заголовок",
+  "",
+  "Первый абзац с **жирным** и `кодом`.",
+  "",
+  "- пункт один",
+  "- пункт два",
+  "",
+  "Второй абзац.",
+].join("\n");
+
+describe("MessageItem final render — visual parity with the canonical pipeline", () => {
+  it("a finalized text part renders exactly renderChatMarkdown(text)", () => {
+    const { container } = renderRow(
+      msg([{ type: "text", text: ANSWER, state: "done" }]),
+    );
+    const block = container.querySelector(`.${classes.markdown}`);
+    expect(block).not.toBeNull();
+    // Byte-for-byte the canonical output (the SAME whole-text pass the pre-#492
+    // MarkdownPart produced), including `<li><p>…</p></li>` wrappers.
+    expect(block!.innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
+    // The list wrapper is really present (guards against a vacuous empty render).
+    expect(container.querySelectorAll("li p").length).toBe(2);
+  });
+
+  it("the streaming incremental view CONVERGES to the canonical render on finish", () => {
+    // Mount mid-stream (live tail) — the DOM here is the incremental view.
+    const { container, rerender } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg([{ type: "text", text: ANSWER, state: "streaming" }])}
+          signature={messageSignature(
+            msg([{ type: "text", text: ANSWER, state: "streaming" }]),
+          )}
+          turnStreaming
+        />
+      </MantineProvider>,
+    );
+    // Finish the turn: state flips to done AND the turn is no longer streaming.
+    const done = msg([{ type: "text", text: ANSWER, state: "done" }]);
+    rerender(
+      <MantineProvider>
+        <MessageItem message={done} signature={messageSignature(done)} />
+      </MantineProvider>,
+    );
+    // After finish there is exactly ONE canonical markdown container whose HTML is
+    // the whole-text render — identical to the non-streaming path above.
+    const blocks = container.querySelectorAll(`.${classes.markdown}`);
+    expect(blocks.length).toBe(1);
+    expect(blocks[0].innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
+  });
+
+  it("neutralizeInternalLinks is honored on the finalized render", () => {
+    const linkAnswer = "См. [страницу](/p/abc).";
+    const { container } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg([{ type: "text", text: linkAnswer, state: "done" }])}
+          signature={messageSignature(
+            msg([{ type: "text", text: linkAnswer, state: "done" }]),
+          )}
+          neutralizeInternalLinks
+        />
+      </MantineProvider>,
+    );
+    const block = container.querySelector(`.${classes.markdown}`);
+    expect(block!.innerHTML).toBe(
+      renderChatMarkdown(linkAnswer, { neutralizeInternalLinks: true }),
+    );
+    // The internal link was made inert (no href) by the neutralization flag.
+    const a = container.querySelector("a");
+    expect(a?.hasAttribute("href")).toBe(false);
+  });
+});

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

@@ -4,6 +4,7 @@ import { useTranslation } from "react-i18next";
 import type { UIMessage } from "@ai-sdk/react";
 import ToolCallCard from "@/features/ai-chat/components/tool-call-card.tsx";
 import ReasoningBlock from "@/features/ai-chat/components/reasoning-block.tsx";
+import { StreamingMarkdownText } from "@/features/ai-chat/components/streaming-markdown-text.tsx";
 import ChatErrorAlert from "@/features/ai-chat/components/chat-error-alert.tsx";
 import ChatStoppedNotice from "@/features/ai-chat/components/chat-stopped-notice.tsx";
 import { ToolUiPart, isToolPart } from "@/features/ai-chat/utils/tool-parts.tsx";
@@ -86,17 +87,39 @@ interface MessageItemProps {
  * One assistant text part rendered as sanitized markdown. Memoized on its inputs
  * so a finalized text part is NOT re-parsed on every streamed delta: during a
  * turn only the actively-growing tail part changes its `text`, so every earlier
- * part hits the memo and skips the expensive marked + DOMPurify pass. Props are
- * primitives, so React.memo's default shallow compare is exactly right (the
- * `text` string is compared by value).
+ * part hits the memo and skips the expensive canonical parse + DOMPurify pass.
+ * Props are primitives, so React.memo's default shallow compare is exactly right
+ * (the `text` string is compared by value).
+ *
+ * Streaming gate (#492) — mirrors ReasoningBlock:
+ *  - `streaming` (this is the live, actively-growing tail part of an in-flight
+ *    turn): render incrementally via StreamingMarkdownText — the stabilized blocks
+ *    go through the canonical pipeline (each parsed ONCE, memoized) and only the
+ *    live tail is cheap plain text. This makes the per-tick cost O(new blocks),
+ *    not the pre-#492 O(ticks) whole-answer re-parse on every ~20Hz delta.
+ *  - finalized (the common case, and the turn-end flip): render the WHOLE text
+ *    through ONE canonical pass — byte-identical to the pre-#492 output (visual
+ *    parity). The row re-renders on the streaming→done flip because
+ *    `messageSignature` tracks each part's `state` (and `turnStreaming` flips at
+ *    turn end), so the incremental view always converges to this single render.
  */
 const MarkdownPart = memo(function MarkdownPart({
   text,
   neutralizeInternalLinks,
+  streaming,
 }: {
   text: string;
   neutralizeInternalLinks: boolean;
+  streaming: boolean;
 }) {
+  if (streaming) {
+    return (
+      <StreamingMarkdownText
+        text={text}
+        neutralizeInternalLinks={neutralizeInternalLinks}
+      />
+    );
+  }
   const html = renderChatMarkdown(text, { neutralizeInternalLinks });
   if (html) {
     return (
@@ -179,47 +202,10 @@ function MessageItem({
         {resolveAssistantName(assistantName) ?? t("AI agent")}
       </Text>
       {message.parts.map((part, index) => {
-        if (part.type === "reasoning") {
-          // Reasoning ("thinking") -> a collapsible block with its own token
-          // count. Empty/whitespace reasoning with no authoritative count carries
-          // nothing to show, so skip it (avoids an empty 0-token block).
-          const text = (part as { text?: string }).text ?? "";
-          if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
-            return null;
-          // Absent state (persisted rows) and "done" both mean finalized.
-          // `messageSignature` already includes each part's `state`, so the
-          // streaming→done flip changes the row signature and re-renders this
-          // row — which is what lets ReasoningBlock switch from chunked plain
-          // text to its one-time markdown parse (see reasoning-block.tsx).
-          // ALSO require the turn to be live: a part stranded at
-          // `state:"streaming"` after the turn ended (no `reasoning-end` — see
-          // the `turnStreaming` prop doc) must still finalize and parse.
-          const streaming =
-            turnStreaming && (part as { state?: string }).state === "streaming";
-          return (
-            <ReasoningBlock
-              key={index}
-              text={text}
-              tokens={reasoningTokens}
-              streaming={streaming}
-            />
-          );
-        }
-
-        if (part.type === "text") {
-          // Skip empty/whitespace-only text parts (a streaming message often
-          // starts with an empty text part before the first token arrives); the
-          // typing indicator covers that gap until real content streams in.
-          if (!part.text.trim()) return null;
-          return (
-            <MarkdownPart
-              key={index}
-              text={part.text}
-              neutralizeInternalLinks={neutralizeInternalLinks}
-            />
-          );
-        }
-
+        // Tool parts (`tool-*` / `dynamic-tool`) are template-literal kinds, so
+        // they cannot be a `switch` case; the runtime guard handles them, and the
+        // switch below covers every CLOSED (literal-typed) part kind with a
+        // compile-time exhaustiveness check in its default.
         if (isToolPart(part.type)) {
           return (
             <ToolCallCard
@@ -232,7 +218,76 @@ function MessageItem({
           );
         }
 
-        return null;
+        switch (part.type) {
+          case "reasoning": {
+            // Reasoning ("thinking") -> a collapsible block with its own token
+            // count. Empty/whitespace reasoning with no authoritative count
+            // carries nothing to show, so skip it (avoids an empty 0-token block).
+            const text = part.text ?? "";
+            if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
+              return null;
+            // Absent state (persisted rows) and "done" both mean finalized.
+            // `messageSignature` already includes each part's `state`, so the
+            // streaming→done flip changes the row signature and re-renders this
+            // row — which is what lets ReasoningBlock switch from chunked plain
+            // text to its one-time markdown parse (see reasoning-block.tsx).
+            // ALSO require the turn to be live: a part stranded at
+            // `state:"streaming"` after the turn ended (no `reasoning-end` — see
+            // the `turnStreaming` prop doc) must still finalize and parse.
+            const streaming = turnStreaming && part.state === "streaming";
+            return (
+              <ReasoningBlock
+                key={index}
+                text={text}
+                tokens={reasoningTokens}
+                streaming={streaming}
+              />
+            );
+          }
+
+          case "text": {
+            // Skip empty/whitespace-only text parts (a streaming message often
+            // starts with an empty text part before the first token arrives); the
+            // typing indicator covers that gap until real content streams in.
+            if (!part.text.trim()) return null;
+            // The live, actively-growing tail part of the in-flight turn renders
+            // incrementally (see MarkdownPart); a finalized part (persisted, or
+            // the turn-end flip) renders the whole text through one canonical
+            // pass. Same liveness rule as the reasoning branch above.
+            const streaming = turnStreaming && part.state === "streaming";
+            return (
+              <MarkdownPart
+                key={index}
+                text={part.text}
+                neutralizeInternalLinks={neutralizeInternalLinks}
+                streaming={streaming}
+              />
+            );
+          }
+
+          case "source-url":
+          case "source-document":
+          case "file":
+          case "step-start":
+            // Not surfaced in the chat bubble (v1) — same as the pre-#492 default.
+            return null;
+
+          default: {
+            // Compile-time exhaustiveness over the CLOSED union members: every
+            // literal-typed part kind is handled above, so the only kinds that
+            // can reach here are the OPEN template-literal ones (`tool-*` — caught
+            // by the guard at runtime — and `data-*`) plus `dynamic-tool`. Adding
+            // a NEW closed part kind to UIMessagePart makes this assignment fail
+            // to compile, forcing it to be handled instead of silently ignored
+            // (this replaces the pre-#492 fall-through `return null` + WARNING).
+            const _exhaustive:
+              | `tool-${string}`
+              | "dynamic-tool"
+              | `data-${string}` = part.type;
+            void _exhaustive;
+            return null;
+          }
+        }
       })}
       {/* A persisted turn error (server stored it in metadata.error). Rendered
           here so it survives a thread remount and shows in reopened history. */}

apps/client/src/features/ai-chat/components/streaming-markdown-text.tsx

@@ -0,0 +1,96 @@
+import { memo, useMemo } from "react";
+import { splitPlainChunks } from "@/features/ai-chat/components/streaming-plain-text.tsx";
+import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
+import classes from "@/features/ai-chat/components/ai-chat.module.css";
+
+/**
+ * One STABILIZED markdown block, rendered through the canonical pipeline and
+ * memoized on its string prop. During streaming only the TAIL chunk grows (the
+ * `splitPlainChunks` append-only invariant guarantees every earlier chunk is
+ * byte-identical across deltas), so React skips every stable block and each one
+ * is parsed by `renderChatMarkdown` EXACTLY ONCE — turning the pre-#492
+ * "re-parse the whole accumulated answer on every ~20Hz tick" (O(ticks)) into
+ * O(number of blocks). The markup is DOMPurify-sanitized inside renderChatMarkdown
+ * before it reaches `dangerouslySetInnerHTML`.
+ *
+ * NOTE (transient streaming-only artifact): a safe cut is a blank-line boundary,
+ * so a construct that legitimately contains a blank line (e.g. a fenced code block
+ * with an empty line) can be split across chunks and render oddly WHILE it is still
+ * streaming. This is cosmetic and self-heals: the moment the part finalizes,
+ * MarkdownPart renders the WHOLE text through one canonical pass (visual parity
+ * with the pre-#492 output). The reasoning path makes the same trade (plain text
+ * while streaming, one markdown parse at the end).
+ */
+const MarkdownChunk = memo(function MarkdownChunk({
+  text,
+  neutralizeInternalLinks,
+}: {
+  text: string;
+  neutralizeInternalLinks: boolean;
+}) {
+  const html = renderChatMarkdown(text, { neutralizeInternalLinks });
+  if (html) {
+    return (
+      <div
+        className={classes.markdown}
+        // Sanitized by renderChatMarkdown (DOMPurify) before insertion.
+        dangerouslySetInnerHTML={{ __html: html }}
+      />
+    );
+  }
+  // Malformed/unsupported markdown could not render synchronously: raw text.
+  return (
+    <div className={classes.markdown} style={{ whiteSpace: "pre-wrap" }}>
+      {text}
+    </div>
+  );
+});
+
+/**
+ * The cheap streaming-time stand-in for the finalized answer's one-time markdown
+ * parse (see MarkdownPart in message-item.tsx). Mirrors StreamingPlainText's
+ * chunked-memo pattern but renders the STABILIZED prefix as real markdown (each
+ * block parsed once, memoized) and only the LIVE tail as flat plain text — so the
+ * user sees formatted output for everything up to the last safe cut, and the not-
+ * yet-stable tail (which markdown-parsing every tick would make O(ticks)) stays a
+ * single cheap escaped text node until it stabilizes into a new block.
+ *
+ * `splitPlainChunks` yields chunks where, under append-only growth, every chunk
+ * except the LAST is immutable; the last chunk is the live tail. Index keys are
+ * therefore stable (a given index never changes to a different chunk's content).
+ */
+export function StreamingMarkdownText({
+  text,
+  neutralizeInternalLinks,
+}: {
+  text: string;
+  neutralizeInternalLinks: boolean;
+}) {
+  const chunks = useMemo(() => splitPlainChunks(text), [text]);
+  return (
+    <>
+      {chunks.map((chunk, index) =>
+        index < chunks.length - 1 ? (
+          <MarkdownChunk
+            key={index}
+            text={chunk}
+            neutralizeInternalLinks={neutralizeInternalLinks}
+          />
+        ) : (
+          // The live tail: flat, React-escaped plain text (no markdown parse, no
+          // sanitizer, no innerHTML). `pre-wrap` preserves its newlines; trailing
+          // separator newlines are dropped at display time so the block gap comes
+          // from the markdown margins, not a doubled empty line (mirrors
+          // PlainChunk in streaming-plain-text.tsx).
+          <div
+            key={index}
+            className={classes.markdown}
+            style={{ whiteSpace: "pre-wrap" }}
+          >
+            {chunk.replace(/\n+$/, "")}
+          </div>
+        ),
+      )}
+    </>
+  );
+}

apps/client/src/features/api-key/components/api-keys-manager.test.tsx

@@ -0,0 +1,234 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import {
+  render,
+  screen,
+  fireEvent,
+  waitFor,
+  within,
+} from "@testing-library/react";
+import { MantineProvider } from "@mantine/core";
+import { ModalsProvider } from "@mantine/modals";
+import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
+import { Provider, createStore } from "jotai";
+import { UserRole } from "@/lib/types";
+import { currentUserAtom } from "@/features/user/atoms/current-user-atom";
+import { IApiKey } from "@/features/api-key/types/api-key.types";
+
+// Mock the service layer so no real HTTP is attempted; every test drives the
+// component through these three functions.
+vi.mock("@/features/api-key/services/api-key-service", () => ({
+  getApiKeys: vi.fn(),
+  createApiKey: vi.fn(),
+  revokeApiKey: vi.fn(),
+}));
+
+import {
+  getApiKeys,
+  createApiKey,
+  revokeApiKey,
+} from "@/features/api-key/services/api-key-service";
+import ApiKeysManager from "./api-keys-manager";
+
+// matchMedia (read by MantineProvider) is stubbed globally in vitest.setup.ts.
+
+const ISO_SOON = new Date(Date.now() + 10 * 864e5).toISOString();
+const ISO_FAR = new Date(Date.now() + 200 * 864e5).toISOString();
+const ISO_EXPIRED = new Date(Date.now() - 3 * 864e5).toISOString();
+
+// Dump the storage stub via the Web Storage API — its data lives in a closure
+// (see vitest.setup.ts), so JSON.stringify(localStorage) would be vacuous.
+function storageDump(): string {
+  let out = "";
+  for (let i = 0; i < localStorage.length; i++) {
+    const k = localStorage.key(i) as string;
+    out += `${k}=${localStorage.getItem(k)};`;
+  }
+  return out;
+}
+
+function makeKey(overrides: Partial<IApiKey> = {}): IApiKey {
+  return {
+    id: "key-1",
+    name: "CI token",
+    expiresAt: ISO_FAR,
+    lastUsedAt: null,
+    createdAt: "2026-01-01T00:00:00.000Z",
+    creator: { id: "u1", name: "Alice", email: "a@x.io", avatarUrl: null },
+    ...overrides,
+  };
+}
+
+function renderManager(role: UserRole) {
+  const store = createStore();
+  store.set(currentUserAtom, {
+    user: { id: "me", role } as never,
+    workspace: {} as never,
+  });
+  const queryClient = new QueryClient({
+    defaultOptions: { queries: { retry: false } },
+  });
+  const utils = render(
+    <Provider store={store}>
+      <QueryClientProvider client={queryClient}>
+        <MantineProvider>
+          <ModalsProvider>
+            <ApiKeysManager />
+          </ModalsProvider>
+        </MantineProvider>
+      </QueryClientProvider>
+    </Provider>,
+  );
+  return { store, queryClient, ...utils };
+}
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  localStorage.clear();
+});
+
+describe("ApiKeysManager — list rendering", () => {
+  it("renders an explicit expiry date and highlights a <30-day key (acceptance #3)", async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([
+      makeKey({ id: "k-soon", name: "Soon key", expiresAt: ISO_SOON }),
+      makeKey({ id: "k-far", name: "Far key", expiresAt: ISO_FAR }),
+    ]);
+
+    renderManager(UserRole.MEMBER);
+
+    await screen.findByText("Soon key");
+    // Explicit dates, not "in N days": the year is rendered verbatim.
+    const soonYear = new Date(ISO_SOON).getFullYear().toString();
+    expect(screen.getAllByText(new RegExp(soonYear)).length).toBeGreaterThan(0);
+    // Exactly one key is inside the 30-day warning window.
+    expect(screen.getAllByText("Expiring soon")).toHaveLength(1);
+  });
+
+  it('shows "Expired" (not "Expiring soon") for an already-expired key', async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([
+      makeKey({ id: "k-dead", name: "Dead key", expiresAt: ISO_EXPIRED }),
+    ]);
+    renderManager(UserRole.MEMBER);
+    await screen.findByText("Dead key");
+    // A past expiry is labelled "Expired", never the forward-looking badge.
+    expect(screen.getByText("Expired")).toBeDefined();
+    expect(screen.queryByText("Expiring soon")).toBeNull();
+  });
+
+  it('shows "Never" for an unlimited key and no highlight', async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([
+      makeKey({ id: "k-forever", name: "Forever", expiresAt: null }),
+    ]);
+    renderManager(UserRole.MEMBER);
+    await screen.findByText("Forever");
+    expect(screen.getByText("Never")).toBeDefined();
+    expect(screen.queryByText("Expiring soon")).toBeNull();
+  });
+
+  it("empty list shows the empty state", async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([]);
+    renderManager(UserRole.MEMBER);
+    expect(await screen.findByText("No API keys yet")).toBeDefined();
+  });
+});
+
+describe("ApiKeysManager — admin vs member view (acceptance #6)", () => {
+  it("a member does NOT see the author column", async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([
+      makeKey({ creator: { id: "me", name: "Me", email: "m@x.io", avatarUrl: null } }),
+    ]);
+    renderManager(UserRole.MEMBER);
+    await screen.findByText("CI token");
+    // Author header absent + creator name not rendered (no author column).
+    expect(screen.queryByText("Author")).toBeNull();
+    expect(screen.queryByText("Me")).toBeNull();
+  });
+
+  it("an admin sees the author column with the creator's name", async () => {
+    vi.mocked(getApiKeys).mockResolvedValue([
+      makeKey({ creator: { id: "u1", name: "Alice", email: "a@x.io", avatarUrl: null } }),
+    ]);
+    renderManager(UserRole.ADMIN);
+    await screen.findByText("CI token");
+    expect(screen.getByText("Author")).toBeDefined();
+    expect(screen.getByText("Alice")).toBeDefined();
+  });
+});
+
+describe("ApiKeysManager — revoke (acceptance #4)", () => {
+  it("revoke removes the row from the list", async () => {
+    vi.mocked(getApiKeys)
+      .mockResolvedValueOnce([
+        makeKey({ id: "k1", name: "Doomed" }),
+        makeKey({ id: "k2", name: "Survivor" }),
+      ])
+      // After revoke, invalidation refetches the reduced list.
+      .mockResolvedValue([makeKey({ id: "k2", name: "Survivor" })]);
+    vi.mocked(revokeApiKey).mockResolvedValue();
+
+    renderManager(UserRole.MEMBER);
+    await screen.findByText("Doomed");
+
+    fireEvent.click(screen.getByLabelText("Revoke Doomed"));
+    // Confirm modal → click the destructive confirm button.
+    const confirm = await screen.findByRole("button", { name: "Revoke" });
+    fireEvent.click(confirm);
+
+    await waitFor(() =>
+      expect(screen.queryByText("Doomed")).toBeNull(),
+    );
+    expect(screen.getByText("Survivor")).toBeDefined();
+    expect(revokeApiKey).toHaveBeenCalledWith("k1");
+  });
+});
+
+describe("ApiKeysManager — show-once token (acceptance #1 & #2)", () => {
+  it("shows the token once, then discards it from the UI, localStorage and query cache", async () => {
+    const SECRET = "gm_secret-token-value-xyz";
+    vi.mocked(getApiKeys).mockResolvedValue([]);
+    vi.mocked(createApiKey).mockResolvedValue({
+      token: SECRET,
+      apiKey: {
+        id: "new-1",
+        name: "My key",
+        expiresAt: ISO_FAR,
+        createdAt: new Date().toISOString(),
+      },
+    });
+
+    const { queryClient } = renderManager(UserRole.MEMBER);
+    await screen.findByText("No API keys yet");
+
+    // Open create modal (use the header CTA), fill the name, submit.
+    fireEvent.click(
+      screen.getAllByRole("button", { name: "Create API key" })[0],
+    );
+    const nameInput = await screen.findByLabelText(/Name/);
+    fireEvent.change(nameInput, { target: { value: "My key" } });
+    fireEvent.click(screen.getByRole("button", { name: "Create" }));
+
+    // Token is shown exactly once in the show-once modal.
+    const tokenEl = await screen.findByTestId("api-key-token");
+    expect(tokenEl.textContent).toBe(SECRET);
+
+    // Close the modal → token discarded from the DOM.
+    fireEvent.click(screen.getByRole("button", { name: "Done" }));
+    await waitFor(() =>
+      expect(screen.queryByTestId("api-key-token")).toBeNull(),
+    );
+
+    // Acceptance #2: the secret is nowhere in localStorage or the react-query
+    // caches (query cache never carried it; the mutation copy was reset()).
+    // Non-vacuous: currentUser IS in storage, so the dump is exercised.
+    const dump = storageDump();
+    expect(dump).toContain("currentUser");
+    expect(dump).not.toContain(SECRET);
+    const cacheDump = JSON.stringify(
+      queryClient.getQueryCache().getAll().map((q) => q.state.data),
+    );
+    expect(cacheDump).not.toContain(SECRET);
+    const mutationDump = JSON.stringify(
+      queryClient.getMutationCache().getAll().map((m) => m.state.data),
+    );
+    expect(mutationDump).not.toContain(SECRET);
+  });
+});

apps/client/src/features/api-key/components/api-keys-manager.tsx

@@ -0,0 +1,264 @@
+import { useState } from "react";
+import {
+  ActionIcon,
+  Badge,
+  Button,
+  Center,
+  Group,
+  Loader,
+  Stack,
+  Table,
+  Text,
+  Tooltip,
+} from "@mantine/core";
+import { modals } from "@mantine/modals";
+import { notifications } from "@mantine/notifications";
+import { IconAlertTriangle, IconKey, IconTrash } from "@tabler/icons-react";
+import { useTranslation } from "react-i18next";
+import useUserRole from "@/hooks/use-user-role.tsx";
+import { formatLocalized, useDateFnsLocale } from "@/lib/date-locale";
+import { timeAgo } from "@/lib/time";
+import {
+  useApiKeysQuery,
+  useCreateApiKeyMutation,
+  useRevokeApiKeyMutation,
+} from "@/features/api-key/queries/api-key-query";
+import {
+  IApiKey,
+  ICreateApiKeyResponse,
+} from "@/features/api-key/types/api-key.types";
+import {
+  isExpired,
+  isExpiringSoon,
+  lastUsedBucket,
+} from "@/features/api-key/utils";
+import { CreateApiKeyModal } from "./create-api-key-modal";
+import { ShowTokenModal } from "./show-token-modal";
+
+export default function ApiKeysManager() {
+  const { t } = useTranslation();
+  const locale = useDateFnsLocale();
+  // Manage-on-API === Owner/Admin server-side, so the admin (workspace-wide)
+  // list + "author" column mirror exactly the roles the server serves the
+  // whole-workspace response to.
+  const { isAdmin } = useUserRole();
+
+  const { data: keys, isLoading, isError } = useApiKeysQuery();
+  const createMutation = useCreateApiKeyMutation();
+  const revokeMutation = useRevokeApiKeyMutation();
+
+  const [createOpened, setCreateOpened] = useState(false);
+  // SECURITY: the show-once token lives ONLY here, in this component's local
+  // state. It is never written to localStorage, the query cache or a log. Closing
+  // the modal sets it back to null (handleCloseToken) — discarded forever.
+  const [createdKey, setCreatedKey] = useState<ICreateApiKeyResponse | null>(
+    null,
+  );
+
+  const handleCreate = async (values: {
+    name: string;
+    expiresAt: string | null;
+  }): Promise<boolean> => {
+    try {
+      const res = await createMutation.mutateAsync(values);
+      setCreateOpened(false);
+      // Move the token into local state, then immediately purge react-query's
+      // own copy of the mutation result so the secret does not linger in the
+      // mutation cache.
+      setCreatedKey(res);
+      createMutation.reset();
+      return true;
+    } catch {
+      notifications.show({
+        message: t("Failed to create API key"),
+        color: "red",
+      });
+      return false;
+    }
+  };
+
+  const handleCloseToken = () => {
+    // Token discarded — the only copy the UI ever held is dropped here.
+    setCreatedKey(null);
+  };
+
+  const openRevokeModal = (key: IApiKey) =>
+    modals.openConfirmModal({
+      title: t("Revoke API key"),
+      centered: true,
+      children: (
+        <Text size="sm">
+          {t(
+            'Are you sure you want to revoke "{{name}}"? Any client using this key will immediately lose access. This cannot be undone.',
+            { name: key.name },
+          )}
+        </Text>
+      ),
+      labels: { confirm: t("Revoke"), cancel: t("Cancel") },
+      confirmProps: { color: "red" },
+      onConfirm: () => revokeMutation.mutate(key.id),
+    });
+
+  const formatDate = (iso: string) =>
+    formatLocalized(new Date(iso), "MMM dd, yyyy", "PP", locale);
+
+  const renderLastUsed = (lastUsedAt: string | null) => {
+    switch (lastUsedBucket(lastUsedAt)) {
+      case "never":
+        return t("Never used");
+      case "recent":
+        return t("Within the last hour");
+      case "stale":
+        // Coarse relative time — last_used_at is throttled to ~1h server-side,
+        // so we don't promise finer precision.
+        return timeAgo(new Date(lastUsedAt as string));
+    }
+  };
+
+  if (isLoading) {
+    return (
+      <Center py="xl">
+        <Loader size="sm" />
+      </Center>
+    );
+  }
+
+  const rows = (keys ?? []).map((key) => {
+    // Mutually exclusive: an already-expired key is labelled "Expired" (a past
+    // expiry) rather than the forward-looking "Expiring soon". isExpiringSoon
+    // also matches past expiries, so gate "soon" on !expired.
+    const expired = isExpired(key.expiresAt);
+    const soon = !expired && isExpiringSoon(key.expiresAt);
+    return (
+      <Table.Tr key={key.id}>
+        <Table.Td>
+          <Text fw={500}>{key.name}</Text>
+        </Table.Td>
+        <Table.Td>{formatDate(key.createdAt)}</Table.Td>
+        <Table.Td>
+          {key.expiresAt ? (
+            <Group gap={6} wrap="nowrap">
+              <Text size="sm">{formatDate(key.expiresAt)}</Text>
+              {expired && (
+                <Tooltip label={t("This key has expired")} withArrow>
+                  <Badge
+                    color="red"
+                    variant="light"
+                    size="sm"
+                    leftSection={<IconAlertTriangle size={12} />}
+                  >
+                    {t("Expired")}
+                  </Badge>
+                </Tooltip>
+              )}
+              {soon && (
+                <Tooltip
+                  label={t("This key expires within 30 days")}
+                  withArrow
+                >
+                  <Badge
+                    color="orange"
+                    variant="light"
+                    size="sm"
+                    leftSection={<IconAlertTriangle size={12} />}
+                  >
+                    {t("Expiring soon")}
+                  </Badge>
+                </Tooltip>
+              )}
+            </Group>
+          ) : (
+            <Text size="sm" c="dimmed">
+              {t("Never")}
+            </Text>
+          )}
+        </Table.Td>
+        <Table.Td>
+          <Text size="sm" c="dimmed">
+            {renderLastUsed(key.lastUsedAt)}
+          </Text>
+        </Table.Td>
+        {isAdmin && (
+          <Table.Td>
+            <Text size="sm">
+              {key.creator?.name ?? key.creator?.email ?? t("Unknown")}
+            </Text>
+          </Table.Td>
+        )}
+        <Table.Td style={{ textAlign: "right" }}>
+          <Tooltip label={t("Revoke")} withArrow>
+            <ActionIcon
+              variant="subtle"
+              color="red"
+              aria-label={t("Revoke {{name}}", { name: key.name })}
+              onClick={() => openRevokeModal(key)}
+            >
+              <IconTrash size={16} />
+            </ActionIcon>
+          </Tooltip>
+        </Table.Td>
+      </Table.Tr>
+    );
+  });
+
+  return (
+    <>
+      <Group justify="space-between" mb="md">
+        <Text size="sm" c="dimmed">
+          {isAdmin
+            ? t("API keys across the workspace.")
+            : t("Your personal API keys.")}
+        </Text>
+        <Button
+          leftSection={<IconKey size={16} />}
+          onClick={() => setCreateOpened(true)}
+        >
+          {t("Create API key")}
+        </Button>
+      </Group>
+
+      {isError ? (
+        <Text c="red" size="sm">
+          {t("Failed to load API keys.")}
+        </Text>
+      ) : rows.length === 0 ? (
+        <Stack align="center" gap="xs" py="xl">
+          <IconKey size={32} opacity={0.4} />
+          <Text c="dimmed">{t("No API keys yet")}</Text>
+          <Button
+            variant="light"
+            leftSection={<IconKey size={16} />}
+            onClick={() => setCreateOpened(true)}
+          >
+            {t("Create API key")}
+          </Button>
+        </Stack>
+      ) : (
+        <Table.ScrollContainer minWidth={600}>
+          <Table verticalSpacing="sm" highlightOnHover>
+            <Table.Thead>
+              <Table.Tr>
+                <Table.Th>{t("Name")}</Table.Th>
+                <Table.Th>{t("Created")}</Table.Th>
+                <Table.Th>{t("Expires")}</Table.Th>
+                <Table.Th>{t("Last used")}</Table.Th>
+                {isAdmin && <Table.Th>{t("Author")}</Table.Th>}
+                <Table.Th />
+              </Table.Tr>
+            </Table.Thead>
+            <Table.Tbody>{rows}</Table.Tbody>
+          </Table>
+        </Table.ScrollContainer>
+      )}
+
+      <CreateApiKeyModal
+        opened={createOpened}
+        onClose={() => setCreateOpened(false)}
+        onSubmit={handleCreate}
+        loading={createMutation.isPending}
+      />
+
+      <ShowTokenModal created={createdKey} onClose={handleCloseToken} />
+    </>
+  );
+}

apps/client/src/features/api-key/components/create-api-key-modal.tsx

@@ -0,0 +1,104 @@
+import { Button, Group, Modal, Select, Stack, TextInput } from "@mantine/core";
+import { useForm } from "@mantine/form";
+import { useTranslation } from "react-i18next";
+import {
+  ApiKeyLifetime,
+  DEFAULT_LIFETIME,
+  lifetimeToExpiresAt,
+} from "@/features/api-key/utils";
+
+interface Props {
+  opened: boolean;
+  onClose: () => void;
+  // Resolves the create request, returning true on success. The parent owns the
+  // mutation (and the token it returns); this modal only collects the name +
+  // lifetime. On failure (false) the form state is kept so the user can retry.
+  onSubmit: (values: {
+    name: string;
+    expiresAt: string | null;
+  }) => Promise<boolean>;
+  loading?: boolean;
+}
+
+interface FormValues {
+  name: string;
+  lifetime: ApiKeyLifetime;
+}
+
+export function CreateApiKeyModal({
+  opened,
+  onClose,
+  onSubmit,
+  loading,
+}: Props) {
+  const { t } = useTranslation();
+
+  const form = useForm<FormValues>({
+    initialValues: {
+      name: "",
+      lifetime: DEFAULT_LIFETIME,
+    },
+    validate: {
+      name: (value) =>
+        value.trim().length === 0 ? t("Name is required") : null,
+    },
+  });
+
+  const lifetimeOptions: { value: ApiKeyLifetime; label: string }[] = [
+    { value: "30d", label: t("30 days") },
+    { value: "90d", label: t("90 days") },
+    { value: "1y", label: t("1 year") },
+    { value: "never", label: t("No expiration") },
+  ];
+
+  const handleSubmit = form.onSubmit(async (values) => {
+    const ok = await onSubmit({
+      name: values.name.trim(),
+      expiresAt: lifetimeToExpiresAt(values.lifetime),
+    });
+    // Reset only after a successful submit so a failed create keeps the form
+    // state (the parent surfaces the error via a notification).
+    if (ok) form.reset();
+  });
+
+  const handleClose = () => {
+    form.reset();
+    onClose();
+  };
+
+  return (
+    <Modal
+      opened={opened}
+      onClose={handleClose}
+      title={t("Create API key")}
+      centered
+    >
+      <form onSubmit={handleSubmit}>
+        <Stack gap="sm">
+          <TextInput
+            label={t("Name")}
+            placeholder={t("e.g. CI deploy token")}
+            data-autofocus
+            withAsterisk
+            {...form.getInputProps("name")}
+          />
+          <Select
+            label={t("Expiration")}
+            data={lifetimeOptions}
+            allowDeselect={false}
+            checkIconPosition="right"
+            {...form.getInputProps("lifetime")}
+          />
+          <Group justify="flex-end" mt="xs">
+            <Button variant="default" onClick={handleClose} type="button">
+              {t("Cancel")}
+            </Button>
+            <Button type="submit" loading={loading}>
+              {t("Create")}
+            </Button>
+          </Group>
+        </Stack>
+      </form>
+    </Modal>
+  );
+}

apps/client/src/features/api-key/components/show-token-modal.tsx

@@ -0,0 +1,107 @@
+import {
+  Alert,
+  Button,
+  Code,
+  Group,
+  Modal,
+  Stack,
+  Text,
+} from "@mantine/core";
+import { IconAlertTriangle, IconCheck, IconCopy } from "@tabler/icons-react";
+import { useTranslation } from "react-i18next";
+import { CopyButton } from "@/components/common/copy-button";
+import { formatLocalized, useDateFnsLocale } from "@/lib/date-locale";
+import { ICreateApiKeyResponse } from "@/features/api-key/types/api-key.types";
+
+interface Props {
+  // The freshly-created key incl. its token. Owned by the parent; this modal
+  // only renders it and never copies it into its own persistent state.
+  created: ICreateApiKeyResponse | null;
+  // Closing MUST discard the token in the parent (set the `created` prop back to
+  // null) — the token is shown exactly once.
+  onClose: () => void;
+}
+
+export function ShowTokenModal({ created, onClose }: Props) {
+  const { t } = useTranslation();
+  const locale = useDateFnsLocale();
+
+  const expiresAt = created?.apiKey.expiresAt ?? null;
+
+  return (
+    <Modal
+      opened={created !== null}
+      onClose={onClose}
+      title={t("API key created")}
+      centered
+      // No dismiss-on-outside-click: the token is irretrievable, so closing is a
+      // deliberate act (the user confirms they have saved it).
+      closeOnClickOutside={false}
+    >
+      {created && (
+        <Stack gap="sm">
+          <Alert
+            color="orange"
+            icon={<IconAlertTriangle size={18} />}
+            variant="light"
+          >
+            {t(
+              "Copy your API key now and store it somewhere safe. For security reasons it will not be shown again.",
+            )}
+          </Alert>
+
+          <div>
+            <Text size="sm" c="dimmed" mb={4}>
+              {t("Token")}
+            </Text>
+            <Group gap="xs" wrap="nowrap" align="flex-start">
+              <Code
+                block
+                data-testid="api-key-token"
+                style={{ flex: 1, wordBreak: "break-all" }}
+              >
+                {created.token}
+              </Code>
+              <CopyButton value={created.token}>
+                {({ copied, copy }) => (
+                  <Button
+                    variant="light"
+                    size="xs"
+                    color={copied ? "teal" : "blue"}
+                    leftSection={
+                      copied ? (
+                        <IconCheck size={16} />
+                      ) : (
+                        <IconCopy size={16} />
+                      )
+                    }
+                    onClick={copy}
+                  >
+                    {copied ? t("Copied") : t("Copy")}
+                  </Button>
+                )}
+              </CopyButton>
+            </Group>
+          </div>
+
+          <Text size="sm" c="dimmed">
+            {expiresAt
+              ? t("Expires {{date}}", {
+                  date: formatLocalized(
+                    new Date(expiresAt),
+                    "MMM dd, yyyy",
+                    "PP",
+                    locale,
+                  ),
+                })
+              : t("This key never expires")}
+          </Text>
+
+          <Group justify="flex-end" mt="xs">
+            <Button onClick={onClose}>{t("Done")}</Button>
+          </Group>
+        </Stack>
+      )}
+    </Modal>
+  );
+}

apps/client/src/features/api-key/queries/api-key-query.ts

@@ -0,0 +1,66 @@
+import {
+  useMutation,
+  useQuery,
+  useQueryClient,
+  UseQueryResult,
+} from "@tanstack/react-query";
+import { notifications } from "@mantine/notifications";
+import { useTranslation } from "react-i18next";
+import {
+  createApiKey,
+  getApiKeys,
+  revokeApiKey,
+} from "@/features/api-key/services/api-key-service";
+import {
+  IApiKey,
+  ICreateApiKey,
+  ICreateApiKeyResponse,
+} from "@/features/api-key/types/api-key.types";
+
+export const API_KEYS_QUERY_KEY = ["api-keys"];
+
+export function useApiKeysQuery(): UseQueryResult<IApiKey[], Error> {
+  return useQuery({
+    queryKey: API_KEYS_QUERY_KEY,
+    queryFn: () => getApiKeys(),
+  });
+}
+
+/**
+ * Create mutation.
+ *
+ * SECURITY: the response contains the token exactly once. This hook deliberately
+ * does NOT stash it anywhere — the caller reads it from `mutateAsync`'s resolved
+ * value, moves it into the show-once modal's local state, then calls
+ * `mutation.reset()` to purge react-query's own copy immediately. `gcTime: 0`
+ * is a second belt so nothing lingers in the mutation cache after the observer
+ * unmounts. The list is invalidated here (the list carries no token).
+ */
+export function useCreateApiKeyMutation() {
+  const queryClient = useQueryClient();
+  return useMutation<ICreateApiKeyResponse, Error, ICreateApiKey>({
+    mutationFn: (data) => createApiKey(data),
+    gcTime: 0,
+    onSuccess: () => {
+      queryClient.invalidateQueries({ queryKey: API_KEYS_QUERY_KEY });
+    },
+  });
+}
+
+export function useRevokeApiKeyMutation() {
+  const { t } = useTranslation();
+  const queryClient = useQueryClient();
+  return useMutation<void, Error, string>({
+    mutationFn: (id) => revokeApiKey(id),
+    onSuccess: () => {
+      queryClient.invalidateQueries({ queryKey: API_KEYS_QUERY_KEY });
+      notifications.show({ message: t("API key revoked") });
+    },
+    onError: () => {
+      notifications.show({
+        message: t("Failed to revoke API key"),
+        color: "red",
+      });
+    },
+  });
+}

apps/client/src/features/api-key/services/api-key-service.test.ts

@@ -0,0 +1,78 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+
+// Mock the api-client (axios instance). vi.mock replaces the whole module, so
+// the real response interceptor — which returns `response.data`, i.e. the
+// server envelope { data, success, status } — is bypassed. Our mocked post/get
+// therefore resolve to that post-interceptor envelope shape directly, and the
+// service under test reads `.data` off it to unwrap the inner payload. This is
+// the one bug-prone line the component tests (which fully mock the service)
+// never exercise.
+const { post, get } = vi.hoisted(() => ({ post: vi.fn(), get: vi.fn() }));
+vi.mock("@/lib/api-client", () => ({
+  default: { post, get },
+}));
+
+import {
+  createApiKey,
+  getApiKeys,
+} from "@/features/api-key/services/api-key-service";
+
+beforeEach(() => {
+  vi.clearAllMocks();
+});
+
+describe("api-key-service response-contract unwrap", () => {
+  it("createApiKey unwraps the envelope to { token, apiKey }", async () => {
+    const payload = {
+      token: "gm_secret-abc",
+      apiKey: {
+        id: "key-1",
+        name: "CI token",
+        expiresAt: "2027-07-11T12:00:00.000Z",
+        createdAt: "2026-07-11T12:00:00.000Z",
+      },
+    };
+    // The server envelope as the interceptor hands it to the service.
+    post.mockResolvedValue({ data: payload, success: true, status: 200 });
+
+    const result = await createApiKey({
+      name: "CI token",
+      expiresAt: "2027-07-11T12:00:00.000Z",
+    });
+
+    // The inner payload only — not the { data, success, status } wrapper.
+    expect(result).toEqual(payload);
+    expect(result.token).toBe("gm_secret-abc");
+    expect(result.apiKey).toEqual(payload.apiKey);
+    expect(post).toHaveBeenCalledWith("/api-keys/create", {
+      name: "CI token",
+      expiresAt: "2027-07-11T12:00:00.000Z",
+    });
+  });
+
+  it("getApiKeys unwraps the envelope to the array of rows", async () => {
+    const rows = [
+      {
+        id: "key-1",
+        name: "CI token",
+        expiresAt: null,
+        lastUsedAt: null,
+        createdAt: "2026-01-01T00:00:00.000Z",
+      },
+      {
+        id: "key-2",
+        name: "Deploy token",
+        expiresAt: "2027-01-01T00:00:00.000Z",
+        lastUsedAt: null,
+        createdAt: "2026-02-01T00:00:00.000Z",
+      },
+    ];
+    post.mockResolvedValue({ data: rows, success: true, status: 200 });
+
+    const result = await getApiKeys();
+
+    expect(result).toEqual(rows);
+    expect(result).toHaveLength(2);
+    expect(post).toHaveBeenCalledWith("/api-keys/list");
+  });
+});

apps/client/src/features/api-key/services/api-key-service.ts

@@ -0,0 +1,30 @@
+import api from "@/lib/api-client";
+import {
+  IApiKey,
+  ICreateApiKey,
+  ICreateApiKeyResponse,
+} from "@/features/api-key/types/api-key.types";
+
+// Mint a new key. The response carries the token ONCE — the caller must move it
+// straight into the show-once modal's local state and never cache it. See
+// queries/api-key-query.ts (gcTime: 0 + query invalidation) and
+// components/api-keys-manager.tsx `handleCreate` (createMutation.reset() right
+// after reading the token) for the reset()-after-read pattern.
+export async function createApiKey(
+  data: ICreateApiKey,
+): Promise<ICreateApiKeyResponse> {
+  const res = await api.post<ICreateApiKeyResponse>("/api-keys/create", data);
+  return res.data as ICreateApiKeyResponse;
+}
+
+// List the caller's keys (or, for an admin, every key in the workspace with
+// creator attribution). Never returns token material.
+export async function getApiKeys(): Promise<IApiKey[]> {
+  const res = await api.post<IApiKey[]>("/api-keys/list");
+  return res.data as IApiKey[];
+}
+
+// Revocation is server-side immediate; the caller drops the row on success.
+export async function revokeApiKey(id: string): Promise<void> {
+  await api.post("/api-keys/revoke", { id });
+}

apps/client/src/features/api-key/types/api-key.types.ts

@@ -0,0 +1,48 @@
+// Compact creator attribution embedded in the admin (workspace-wide) list. A
+// normal member's list only ever contains their own keys, so the field is
+// present but redundant; the author column is only rendered for admins.
+export interface IApiKeyCreator {
+  id: string;
+  name: string;
+  email: string;
+  avatarUrl: string | null;
+}
+
+// A single api-key row as returned by `POST /api/api-keys/list`. Note: the list
+// NEVER carries token material — only metadata.
+export interface IApiKey {
+  id: string;
+  name: string;
+  // ISO string, or null for an unlimited ("never expires") key.
+  expiresAt: string | null;
+  // ISO string, or null if the key was never used. Throttled to ~1h server-side
+  // (#501), so the UI must not promise sub-hour precision.
+  lastUsedAt: string | null;
+  createdAt: string;
+  creator?: IApiKeyCreator | null;
+}
+
+// Payload for `POST /api/api-keys/create`. `expiresAt`: an ISO date string for a
+// bounded lifetime, or null for an unlimited key. (undefined would let the
+// server apply its 1-year default, but the form always sends an explicit value.)
+export interface ICreateApiKey {
+  name: string;
+  expiresAt: string | null;
+}
+
+// The metadata half of the create response. The token itself is carried
+// separately (see ICreateApiKeyResponse) and is shown exactly once.
+export interface ICreatedApiKey {
+  id: string;
+  name: string;
+  expiresAt: string | null;
+  createdAt: string;
+}
+
+// Response of `POST /api/api-keys/create`. `token` is the ONLY time the secret
+// is ever returned — it must live only in the show-once modal's local state and
+// must never be cached, persisted or logged.
+export interface ICreateApiKeyResponse {
+  token: string;
+  apiKey: ICreatedApiKey;
+}

apps/client/src/features/api-key/utils.test.ts

@@ -0,0 +1,100 @@
+import { describe, it, expect } from "vitest";
+import {
+  DEFAULT_LIFETIME,
+  EXPIRY_WARNING_DAYS,
+  isExpired,
+  isExpiringSoon,
+  lastUsedBucket,
+  lifetimeToExpiresAt,
+} from "./utils";
+
+const NOW = new Date("2026-07-11T12:00:00.000Z");
+const daysFromNow = (n: number) =>
+  new Date(NOW.getTime() + n * 24 * 60 * 60 * 1000).toISOString();
+
+describe("lifetimeToExpiresAt", () => {
+  it("default lifetime is 1 year (acceptance #7)", () => {
+    expect(DEFAULT_LIFETIME).toBe("1y");
+    const iso = lifetimeToExpiresAt("1y", NOW);
+    expect(iso).toBe("2027-07-11T12:00:00.000Z");
+  });
+
+  it('"never" sends null (acceptance #7)', () => {
+    expect(lifetimeToExpiresAt("never", NOW)).toBeNull();
+  });
+
+  it("30d / 90d map to the exact future instant", () => {
+    expect(lifetimeToExpiresAt("30d", NOW)).toBe(daysFromNow(30));
+    expect(lifetimeToExpiresAt("90d", NOW)).toBe(daysFromNow(90));
+  });
+});
+
+describe("isExpiringSoon (acceptance #3 highlight)", () => {
+  it("an unlimited key is never 'soon'", () => {
+    expect(isExpiringSoon(null, NOW)).toBe(false);
+  });
+
+  it("highlights a key expiring within the 30-day window", () => {
+    expect(isExpiringSoon(daysFromNow(EXPIRY_WARNING_DAYS - 1), NOW)).toBe(true);
+    expect(isExpiringSoon(daysFromNow(10), NOW)).toBe(true);
+  });
+
+  it("does not highlight a key well outside the window", () => {
+    expect(isExpiringSoon(daysFromNow(EXPIRY_WARNING_DAYS + 1), NOW)).toBe(
+      false,
+    );
+    expect(isExpiringSoon(daysFromNow(200), NOW)).toBe(false);
+  });
+
+  it("an already-expired key is highlighted", () => {
+    expect(isExpiringSoon(daysFromNow(-3), NOW)).toBe(true);
+  });
+});
+
+describe("isExpired (past vs future expiry)", () => {
+  it("an unlimited key is never expired", () => {
+    expect(isExpired(null, NOW)).toBe(false);
+  });
+
+  it("a past expiry is expired", () => {
+    expect(isExpired(daysFromNow(-3), NOW)).toBe(true);
+    expect(isExpired(daysFromNow(-1), NOW)).toBe(true);
+  });
+
+  it("a future expiry is not expired (even within the warning window)", () => {
+    expect(isExpired(daysFromNow(1), NOW)).toBe(false);
+    expect(isExpired(daysFromNow(EXPIRY_WARNING_DAYS - 1), NOW)).toBe(false);
+    expect(isExpired(daysFromNow(200), NOW)).toBe(false);
+  });
+
+  it("an expiry exactly at 'now' counts as expired (boundary is inclusive)", () => {
+    expect(isExpired(NOW.toISOString(), NOW)).toBe(true);
+  });
+
+  it("is mutually distinguishable from isExpiringSoon: expired vs soon-but-future", () => {
+    // A key 3 days in the past: expired, and (by design) also matches
+    // isExpiringSoon — the UI resolves this by checking isExpired first.
+    expect(isExpired(daysFromNow(-3), NOW)).toBe(true);
+    // A key 10 days in the future: NOT expired, but expiring soon.
+    expect(isExpired(daysFromNow(10), NOW)).toBe(false);
+    expect(isExpiringSoon(daysFromNow(10), NOW)).toBe(true);
+  });
+});
+
+describe("lastUsedBucket (within-the-last-hour semantics)", () => {
+  const minutesAgo = (n: number) =>
+    new Date(NOW.getTime() - n * 60 * 1000).toISOString();
+
+  it("null last-used is 'never'", () => {
+    expect(lastUsedBucket(null, NOW)).toBe("never");
+  });
+
+  it("under an hour is 'recent' (no sub-hour precision promised)", () => {
+    expect(lastUsedBucket(minutesAgo(5), NOW)).toBe("recent");
+    expect(lastUsedBucket(minutesAgo(59), NOW)).toBe("recent");
+  });
+
+  it("over an hour is 'stale'", () => {
+    expect(lastUsedBucket(minutesAgo(90), NOW)).toBe("stale");
+  });
+});

apps/client/src/features/api-key/utils.ts

@@ -0,0 +1,76 @@
+import { addDays, addYears, differenceInMinutes } from "date-fns";
+
+// The single early-warning window (#501/#506): keys whose expiry is closer than
+// this are visually highlighted in the list. Also used to classify a key as
+// already-expired (a negative "days until" is < 30 too).
+export const EXPIRY_WARNING_DAYS = 30;
+
+// last_used_at is throttled to ~1h server-side, so anything under an hour is
+// shown as the coarse "within the last hour" rather than a false-precise
+// "5 minutes ago".
+export const LAST_USED_THROTTLE_MINUTES = 60;
+
+export type ApiKeyLifetime = "30d" | "90d" | "1y" | "never";
+
+export const DEFAULT_LIFETIME: ApiKeyLifetime = "1y";
+
+// Maps a lifetime choice to the `expiresAt` payload sent to the server: an ISO
+// string for a bounded lifetime, or null for an explicit unlimited key.
+export function lifetimeToExpiresAt(
+  lifetime: ApiKeyLifetime,
+  now: Date = new Date(),
+): string | null {
+  switch (lifetime) {
+    case "30d":
+      return addDays(now, 30).toISOString();
+    case "90d":
+      return addDays(now, 90).toISOString();
+    case "1y":
+      return addYears(now, 1).toISOString();
+    case "never":
+      return null;
+  }
+}
+
+// True when a bounded key's expiry is already in the past (or exactly now). An
+// unlimited key (null) is never expired. This is distinct from "expiring soon":
+// the two states are mutually exclusive at the call site (see api-keys-manager),
+// so an already-expired key is labelled "Expired", not "Expiring soon".
+export function isExpired(
+  expiresAt: string | null,
+  now: Date = new Date(),
+): boolean {
+  if (!expiresAt) return false;
+  const expiry = new Date(expiresAt).getTime();
+  if (Number.isNaN(expiry)) return false;
+  return expiry <= now.getTime();
+}
+
+// True when a bounded key expires within the warning window (or is already
+// expired). An unlimited key (null) is never "expiring soon". Callers that need
+// to distinguish an already-past expiry should check isExpired() first, as this
+// predicate deliberately also covers the already-expired case.
+export function isExpiringSoon(
+  expiresAt: string | null,
+  now: Date = new Date(),
+): boolean {
+  if (!expiresAt) return false;
+  const expiry = new Date(expiresAt).getTime();
+  if (Number.isNaN(expiry)) return false;
+  const msLeft = expiry - now.getTime();
+  return msLeft < EXPIRY_WARNING_DAYS * 24 * 60 * 60 * 1000;
+}
+
+// Classifies last-used recency so the caller can pick the honest label without
+// promising sub-hour precision. Returns "never", "recent" (< 1h) or "stale".
+export function lastUsedBucket(
+  lastUsedAt: string | null,
+  now: Date = new Date(),
+): "never" | "recent" | "stale" {
+  if (!lastUsedAt) return "never";
+  const used = new Date(lastUsedAt);
+  if (Number.isNaN(used.getTime())) return "never";
+  return differenceInMinutes(now, used) < LAST_USED_THROTTLE_MINUTES
+    ? "recent"
+    : "stale";
+}

apps/client/src/features/comment/components/comment-list-with-tabs.test.tsx

@@ -27,11 +27,15 @@ vi.mock("@/features/space/queries/space-query.ts", () => ({
 import {
   buildChildrenByParent,
   CommentEditorWithActions,
+  sortResolvedByResolvedAt,
 } from "./comment-list-with-tabs";
 
 const c = (id: string, parentCommentId: string | null = null): IComment =>
   ({ id, parentCommentId }) as IComment;
 
+const resolvedAtComment = (id: string, resolvedAt: unknown): IComment =>
+  ({ id, resolvedAt }) as unknown as IComment;
+
 describe("buildChildrenByParent (childrenByParent grouping)", () => {
   it("returns an empty map for undefined or empty input", () => {
     expect(buildChildrenByParent(undefined).size).toBe(0);
@@ -71,6 +75,48 @@ describe("buildChildrenByParent (childrenByParent grouping)", () => {
   });
 });
 
+describe("sortResolvedByResolvedAt (Resolved tab order, #542)", () => {
+  it("orders by resolvedAt DESC — newest resolve first — with ISO-STRING values", () => {
+    // At runtime resolvedAt is an ISO string (axios JSON / WS subscription), so
+    // the sort must coerce with new Date(...) before .getTime().
+    const older = resolvedAtComment("older", "2026-07-10T10:00:00.000Z");
+    const newest = resolvedAtComment("newest", "2026-07-12T10:00:00.000Z");
+    const middle = resolvedAtComment("middle", "2026-07-11T10:00:00.000Z");
+
+    const out = sortResolvedByResolvedAt([older, newest, middle]);
+    expect(out.map((x) => x.id)).toEqual(["newest", "middle", "older"]);
+  });
+
+  it("also handles Date instances (optimistic onMutate window)", () => {
+    const older = resolvedAtComment("older", new Date("2026-01-01T00:00:00Z"));
+    const newer = resolvedAtComment("newer", new Date("2026-06-01T00:00:00Z"));
+    expect(sortResolvedByResolvedAt([older, newer]).map((x) => x.id)).toEqual([
+      "newer",
+      "older",
+    ]);
+  });
+
+  it("does not mutate the input array", () => {
+    const a = resolvedAtComment("a", "2026-01-01T00:00:00.000Z");
+    const b = resolvedAtComment("b", "2026-02-01T00:00:00.000Z");
+    const input = [a, b];
+    sortResolvedByResolvedAt(input);
+    expect(input.map((x) => x.id)).toEqual(["a", "b"]);
+  });
+
+  it("keeps stable order for equal resolvedAt timestamps", () => {
+    const ts = "2026-03-03T03:03:03.000Z";
+    const x = resolvedAtComment("x", ts);
+    const y = resolvedAtComment("y", ts);
+    const z = resolvedAtComment("z", ts);
+    expect(sortResolvedByResolvedAt([x, y, z]).map((c) => c.id)).toEqual([
+      "x",
+      "y",
+      "z",
+    ]);
+  });
+});
+
 function renderReplyEditor() {
   return render(
     <MantineProvider>

apps/client/src/features/comment/components/comment-list-with-tabs.tsx

@@ -53,6 +53,22 @@ export function buildChildrenByParent(
   return m;
 }
 
+// Sort the Resolved tab by resolve time, newest first, on a COPY (never mutate
+// the react-query cache array). `resolvedAt` is typed `Date` but at runtime it
+// is an ISO STRING (from the axios-JSON onSuccess and the WS subscription) — a
+// real Date only during the optimistic onMutate window — so it MUST be coerced
+// with `new Date(...)` before `.getTime()`, or a raw `.getTime()` on the string
+// throws / yields NaN. ES2019's stable sort preserves order for equal
+// timestamps. Callers pass a list already filtered to a truthy `resolvedAt`, so
+// the non-null assertion is safe.
+// Exported for unit testing.
+export function sortResolvedByResolvedAt(resolved: IComment[]): IComment[] {
+  return [...resolved].sort(
+    (a, b) =>
+      new Date(b.resolvedAt!).getTime() - new Date(a.resolvedAt!).getTime(),
+  );
+}
+
 function CommentListWithTabs({ onClose }: CommentListWithTabsProps) {
   const { t } = useTranslation();
   const { pageSlug } = useParams();
@@ -91,7 +107,10 @@ function CommentListWithTabs({ onClose }: CommentListWithTabsProps) {
       (comment: IComment) => comment.resolvedAt,
     );
 
-    return { activeComments: active, resolvedComments: resolved };
+    return {
+      activeComments: active,
+      resolvedComments: sortResolvedByResolvedAt(resolved),
+    };
   }, [comments]);
 
   // Index replies by their parent once, instead of an O(n^2) filter per thread.

apps/client/src/features/comment/queries/comment-query.resolve-undo.test.tsx

@@ -0,0 +1,353 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import React from "react";
+import { renderHook, waitFor } from "@testing-library/react";
+import {
+  QueryClient,
+  QueryClientProvider,
+  InfiniteData,
+} from "@tanstack/react-query";
+
+/**
+ * Coverage for the resolve/reopen mutation (#542): the Undo-in-toast reopen and
+ * its double-click guard, the terminal 404 branch (drop from cache + clear the
+ * inline mark, no rollback), and the directional error copy.
+ */
+
+// A fake TipTap editor injected via the mocked pageEditorAtom, so we can assert
+// the mutation clears the inline comment mark (unsetComment / setCommentResolved).
+const editorMock = vi.hoisted(() => ({
+  current: {
+    isDestroyed: false,
+    commands: { unsetComment: vi.fn(), setCommentResolved: vi.fn() },
+  } as {
+    isDestroyed: boolean;
+    commands: {
+      unsetComment: (id: string) => void;
+      setCommentResolved: (id: string, v: boolean) => void;
+    };
+  } | null,
+}));
+
+vi.mock("@mantine/notifications", () => ({
+  notifications: { show: vi.fn(), hide: vi.fn() },
+}));
+
+vi.mock("jotai", () => ({
+  atom: (v: unknown) => v,
+  useAtomValue: () => editorMock.current,
+}));
+
+vi.mock("@/features/comment/services/comment-service", () => ({
+  applySuggestion: vi.fn(),
+  dismissSuggestion: vi.fn(),
+  createComment: vi.fn(),
+  updateComment: vi.fn(),
+  deleteComment: vi.fn(),
+  resolveComment: vi.fn(),
+  getPageComments: vi.fn(),
+}));
+
+import { notifications } from "@mantine/notifications";
+import { resolveComment } from "@/features/comment/services/comment-service";
+import {
+  useResolveCommentMutation,
+  RESOLVE_UNDO_AUTOCLOSE_MS,
+  RQ_KEY,
+} from "@/features/comment/queries/comment-query";
+import { IComment } from "@/features/comment/types/comment.types";
+
+const PAGE_ID = "page-1";
+
+function seededClient(comment: IComment) {
+  const queryClient = new QueryClient({
+    defaultOptions: { mutations: { retry: false } },
+  });
+  const seed: InfiniteData<any> = {
+    pageParams: [undefined],
+    pages: [
+      { items: [comment], meta: { hasNextPage: false, nextCursor: null } },
+    ],
+  };
+  queryClient.setQueryData(RQ_KEY(PAGE_ID), seed);
+  const wrapper = ({ children }: { children: React.ReactNode }) => (
+    <QueryClientProvider client={queryClient}>{children}</QueryClientProvider>
+  );
+  return { queryClient, wrapper };
+}
+
+function items(queryClient: QueryClient): IComment[] {
+  const cache = queryClient.getQueryData(RQ_KEY(PAGE_ID)) as
+    | InfiniteData<any>
+    | undefined;
+  return cache?.pages.flatMap((p) => p.items) ?? [];
+}
+
+const comment = (over?: Partial<IComment>): IComment =>
+  ({
+    id: "c-1",
+    pageId: PAGE_ID,
+    content: "{}",
+    creatorId: "u-1",
+    workspaceId: "ws-1",
+    createdAt: new Date(),
+    resolvedAt: null,
+    ...over,
+  }) as IComment;
+
+// Pull the inline Undo button's onClick out of the success toast's message tree.
+function undoOnClickFromToast(): () => void {
+  const call = vi
+    .mocked(notifications.show)
+    .mock.calls.map((c) => c[0])
+    .find((arg: any) => arg?.autoClose === RESOLVE_UNDO_AUTOCLOSE_MS);
+  expect(call).toBeTruthy();
+  const message: any = (call as any).message;
+  // message = Group( Text, Button ); grab the Button element's onClick.
+  const children = message.props.children as any[];
+  const button = children[1];
+  return button.props.onClick;
+}
+
+describe("useResolveCommentMutation — Undo toast (#542)", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    editorMock.current = {
+      isDestroyed: false,
+      commands: { unsetComment: vi.fn(), setCommentResolved: vi.fn() },
+    };
+  });
+
+  it("resolve shows an Undo toast with autoClose=10000ms; reopen shows NO Undo", async () => {
+    vi.mocked(resolveComment).mockImplementation(async (data) =>
+      comment({
+        resolvedAt: data.resolved ? (new Date() as any) : null,
+      }),
+    );
+    const { wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current.mutateAsync({
+      commentId: "c-1",
+      pageId: PAGE_ID,
+      resolved: true,
+    });
+    await waitFor(() => expect(result.current.isSuccess).toBe(true));
+
+    const resolveToast = vi
+      .mocked(notifications.show)
+      .mock.calls.map((c) => c[0])
+      .find((a: any) => a?.autoClose === RESOLVE_UNDO_AUTOCLOSE_MS);
+    expect(resolveToast).toBeTruthy();
+    expect((resolveToast as any).id).toBe("resolve-undo-c-1");
+    expect((resolveToast as any).autoClose).toBe(10000);
+
+    // Now a reopen → plain toast, no autoClose/Undo, no id.
+    vi.clearAllMocks();
+    await result.current.mutateAsync({
+      commentId: "c-1",
+      pageId: PAGE_ID,
+      resolved: false,
+    });
+    await waitFor(() => expect(result.current.isSuccess).toBe(true));
+    const calls = vi.mocked(notifications.show).mock.calls.map((c) => c[0]);
+    expect(
+      calls.some((a: any) => a?.autoClose === RESOLVE_UNDO_AUTOCLOSE_MS),
+    ).toBe(false);
+    expect(calls).toContainEqual({ message: "Comment re-opened successfully" });
+  });
+
+  it("double/fast Undo click fires reopen EXACTLY once (guard)", async () => {
+    vi.mocked(resolveComment).mockImplementation(async (data) =>
+      comment({ resolvedAt: data.resolved ? (new Date() as any) : null }),
+    );
+    const { wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current.mutateAsync({
+      commentId: "c-1",
+      pageId: PAGE_ID,
+      resolved: true,
+    });
+    await waitFor(() => expect(result.current.isSuccess).toBe(true));
+
+    const onClick = undoOnClickFromToast();
+    // Fire twice synchronously (notifications.hide is not synchronous).
+    onClick();
+    onClick();
+
+    await waitFor(() => {
+      const reopenCalls = vi
+        .mocked(resolveComment)
+        .mock.calls.filter(([d]) => d.resolved === false);
+      expect(reopenCalls).toHaveLength(1);
+    });
+    // The mark was cleared once via setCommentResolved(id, false).
+    expect(editorMock.current!.commands.setCommentResolved).toHaveBeenCalledWith(
+      "c-1",
+      false,
+    );
+    // The toast was hidden.
+    expect(notifications.hide).toHaveBeenCalledWith("resolve-undo-c-1");
+  });
+
+  it("404 → drops the comment from cache, clears the inline mark, no rollback, no Undo", async () => {
+    vi.mocked(resolveComment).mockRejectedValue({ response: { status: 404 } });
+    const { queryClient, wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current
+      .mutateAsync({ commentId: "c-1", pageId: PAGE_ID, resolved: true })
+      .catch(() => undefined);
+    await waitFor(() => expect(result.current.isError).toBe(true));
+
+    // Removed from cache (NOT rolled back to a phantom).
+    expect(items(queryClient)).toHaveLength(0);
+    // Inline mark cleared via unsetComment (mandatory — no panel row left to do it).
+    expect(editorMock.current!.commands.unsetComment).toHaveBeenCalledWith(
+      "c-1",
+    );
+    // Neutral message, red, and crucially NOT the success copy and NO Undo toast.
+    expect(notifications.show).toHaveBeenCalledWith({
+      message: "Comment no longer exists",
+      color: "red",
+    });
+    const calls = vi.mocked(notifications.show).mock.calls.map((c) => c[0]);
+    expect(
+      calls.some((a: any) => a?.autoClose === RESOLVE_UNDO_AUTOCLOSE_MS),
+    ).toBe(false);
+  });
+
+  it("404 does not crash when the editor is gone (read-only / panel closed)", async () => {
+    editorMock.current = null;
+    vi.mocked(resolveComment).mockRejectedValue({ response: { status: 404 } });
+    const { queryClient, wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current
+      .mutateAsync({ commentId: "c-1", pageId: PAGE_ID, resolved: true })
+      .catch(() => undefined);
+    await waitFor(() => expect(result.current.isError).toBe(true));
+    expect(items(queryClient)).toHaveLength(0);
+    expect(notifications.show).toHaveBeenCalledWith({
+      message: "Comment no longer exists",
+      color: "red",
+    });
+  });
+
+  it("non-404 error on REOPEN shows 'Failed to re-open comment' and rolls back", async () => {
+    vi.mocked(resolveComment).mockRejectedValue({ response: { status: 500 } });
+    // Seed a RESOLVED comment (the reopen target).
+    const resolved = comment({ resolvedAt: new Date() as any });
+    const { queryClient, wrapper } = seededClient(resolved);
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current
+      .mutateAsync({ commentId: "c-1", pageId: PAGE_ID, resolved: false })
+      .catch(() => undefined);
+    await waitFor(() => expect(result.current.isError).toBe(true));
+
+    expect(notifications.show).toHaveBeenCalledWith({
+      message: "Failed to re-open comment",
+      color: "red",
+    });
+    expect(notifications.show).not.toHaveBeenCalledWith(
+      expect.objectContaining({ message: "Failed to resolve comment" }),
+    );
+    // Rolled back: the comment is still present and still resolved.
+    expect(items(queryClient)).toHaveLength(1);
+    expect(items(queryClient)[0].resolvedAt).toBeTruthy();
+  });
+
+  it("reopen via Undo FAILS (non-404) → inline mark is NOT left cleared (doc↔panel stay consistent)", async () => {
+    // First resolve succeeds → produces the Undo toast (no mark change on resolve).
+    vi.mocked(resolveComment).mockResolvedValueOnce(
+      comment({ resolvedAt: new Date() as any }),
+    );
+    const { queryClient, wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current.mutateAsync({
+      commentId: "c-1",
+      pageId: PAGE_ID,
+      resolved: true,
+    });
+    await waitFor(() => expect(result.current.isSuccess).toBe(true));
+
+    // Now the reopen fired by Undo fails with a 500.
+    vi.mocked(resolveComment).mockRejectedValue({ response: { status: 500 } });
+    const onClick = undoOnClickFromToast();
+    onClick();
+
+    await waitFor(() => expect(result.current.isError).toBe(true));
+
+    // Core F1 guarantee: the mark-clear now lives in the reopen onSuccess, so a
+    // FAILED reopen must never flip the inline mark to unresolved — otherwise the
+    // doc would show an active highlight the panel still treats as resolved and
+    // the collab mark would diverge with nothing committed on the server.
+    expect(
+      editorMock.current!.commands.setCommentResolved,
+    ).not.toHaveBeenCalledWith("c-1", false);
+    // Cache rolled back: the comment stays resolved and present.
+    expect(items(queryClient)).toHaveLength(1);
+    expect(items(queryClient)[0].resolvedAt).toBeTruthy();
+  });
+
+  it("reopen success with a null editorRef degrades gracefully (no throw, no-op)", async () => {
+    // Read-only view / panel closed: pageEditorAtom is null on the success path.
+    editorMock.current = null;
+    vi.mocked(resolveComment).mockResolvedValue(comment({ resolvedAt: null }));
+    const resolved = comment({ resolvedAt: new Date() as any });
+    const { queryClient, wrapper } = seededClient(resolved);
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current.mutateAsync({
+      commentId: "c-1",
+      pageId: PAGE_ID,
+      resolved: false,
+    });
+    await waitFor(() => expect(result.current.isSuccess).toBe(true));
+
+    // No crash from the reopen mark-clear; the plain reopen toast is still shown.
+    expect(notifications.show).toHaveBeenCalledWith({
+      message: "Comment re-opened successfully",
+    });
+    // Cache updated to reopened (resolvedAt cleared by the server payload).
+    expect(items(queryClient)).toHaveLength(1);
+    expect(items(queryClient)[0].resolvedAt).toBeFalsy();
+  });
+
+  it("non-404 error on RESOLVE shows 'Failed to resolve comment' and rolls back", async () => {
+    vi.mocked(resolveComment).mockRejectedValue({ response: { status: 500 } });
+    const { queryClient, wrapper } = seededClient(comment());
+    const { result } = renderHook(() => useResolveCommentMutation(), {
+      wrapper,
+    });
+
+    await result.current
+      .mutateAsync({ commentId: "c-1", pageId: PAGE_ID, resolved: true })
+      .catch(() => undefined);
+    await waitFor(() => expect(result.current.isError).toBe(true));
+
+    expect(notifications.show).toHaveBeenCalledWith({
+      message: "Failed to resolve comment",
+      color: "red",
+    });
+    // Rolled back to open (previousCache), still present.
+    expect(items(queryClient)).toHaveLength(1);
+    expect(items(queryClient)[0].resolvedAt).toBeFalsy();
+  });
+});

apps/client/src/features/comment/queries/comment-query.ts

@@ -20,12 +20,19 @@ import {
   ISuggestionOutcome,
 } from "@/features/comment/types/comment.types";
 import { notifications } from "@mantine/notifications";
+import { Button, Group, Text } from "@mantine/core";
 import { IPagination } from "@/lib/types.ts";
 import { useTranslation } from "react-i18next";
-import { useEffect, useMemo } from "react";
+import React, { useEffect, useMemo, useRef } from "react";
+import { useAtomValue } from "jotai";
+import { pageEditorAtom } from "@/features/editor/atoms/editor-atoms";
 
 export const RQ_KEY = (pageId: string) => ["comments", pageId];
 
+// How long the resolve success toast (with its inline Undo) stays up before it
+// auto-closes. Policy constant — no env override.
+export const RESOLVE_UNDO_AUTOCLOSE_MS = 10000;
+
 export function useCommentsQuery(params: ICommentParams) {
   const query = useInfiniteQuery({
     queryKey: RQ_KEY(params.pageId),
@@ -376,7 +383,25 @@ export function useResolveCommentMutation() {
   const queryClient = useQueryClient();
   const { t } = useTranslation();
 
-  return useMutation({
+  // Keep the live editor in a ref: the toast's Undo (and the 404 branch) must
+  // clear the inline comment mark AFTER the originating CommentListItem has
+  // unmounted (resolving pulls the comment out of the Open list, so its item is
+  // already gone by the time the 10s toast is clicked). In read-only view
+  // pageEditorAtom is null and the mark converges via the server's
+  // COMMENT_MARK_UPDATE job instead.
+  const editor = useAtomValue(pageEditorAtom);
+  const editorRef = useRef(editor);
+  editorRef.current = editor;
+
+  // Self-reference the mutation so the toast's Undo can re-invoke it (reopen)
+  // long after the triggering component unmounted. Declared BEFORE useMutation
+  // and assigned AFTER; the onClick reads mutationRef.current at CALL time, not
+  // definition time, so there is no initialization cycle.
+  const mutationRef = useRef<{
+    mutate: (vars: IResolveComment) => void;
+  } | null>(null);
+
+  const mutation = useMutation({
     mutationFn: (data: IResolveComment) => resolveComment(data),
     onMutate: async (variables) => {
       await queryClient.cancelQueries({ queryKey: RQ_KEY(variables.pageId) });
@@ -401,7 +426,39 @@ export function useResolveCommentMutation() {
 
       return { previousCache };
     },
-    onError: (_err, variables, context) => {
+    onError: (err: any, variables, context) => {
+      // Terminal 404: the comment was really deleted (missing comment or deleted
+      // page — access denial is 403, resolve is idempotent so no 400). Do NOT
+      // roll back (that would resurrect a phantom row in Resolved); instead drop
+      // it from the cache and clear its now-orphaned inline mark. Mirrors
+      // handleDeleteComment and the dismiss-mutation 404 branch.
+      if (err?.response?.status === 404) {
+        const cache = queryClient.getQueryData(RQ_KEY(variables.pageId)) as
+          | InfiniteData<IPagination<IComment>>
+          | undefined;
+        if (cache) {
+          queryClient.setQueryData(
+            RQ_KEY(variables.pageId),
+            removeCommentFromCache(cache, variables.commentId),
+          );
+        }
+        const ed = editorRef.current;
+        if (ed && !ed.isDestroyed) {
+          try {
+            ed.commands.unsetComment(variables.commentId);
+          } catch {
+            /* editor gone / mark already removed */
+          }
+        }
+        notifications.show({
+          message: t("Comment no longer exists"),
+          color: "red",
+        });
+        return;
+      }
+
+      // Generic failure: roll back the optimistic update and show a DIRECTIONAL
+      // error (resolve vs. reopen), not always "resolve".
       if (context?.previousCache) {
         queryClient.setQueryData(
           RQ_KEY(variables.pageId),
@@ -409,7 +466,9 @@ export function useResolveCommentMutation() {
         );
       }
       notifications.show({
-        message: t("Failed to resolve comment"),
+        message: variables.resolved
+          ? t("Failed to resolve comment")
+          : t("Failed to re-open comment"),
         color: "red",
       });
     },
@@ -430,11 +489,72 @@ export function useResolveCommentMutation() {
         );
       }
 
+      // Reopen keeps the plain toast without an Undo.
+      if (!variables.resolved) {
+        // Clear the inline mark ONLY after the server confirms the reopen, so a
+        // failed reopen never leaves an active highlight the panel still treats
+        // as resolved. Mirrors the 404 branch's editor-liveness guard/try-catch.
+        // The button-triggered reopen already set the mark, so this is an
+        // idempotent no-op there.
+        const ed = editorRef.current;
+        if (ed && !ed.isDestroyed) {
+          try {
+            ed.commands.setCommentResolved(variables.commentId, false);
+          } catch {
+            /* editor gone — server COMMENT_MARK_UPDATE converges it */
+          }
+        }
+        notifications.show({ message: t("Comment re-opened successfully") });
+        return;
+      }
+
+      // Resolve: attach an inline Undo (reopen) to the success toast. Built with
+      // React.createElement because this is a .ts module (no JSX).
+      const { commentId, pageId } = variables;
+      const notificationId = `resolve-undo-${commentId}`;
+      // Double-click guard: notifications.hide is NOT synchronous, so the button
+      // stays clickable for a frame or two — without this a fast double-click
+      // would fire reopen twice.
+      let done = false;
       notifications.show({
-        message: variables.resolved
-          ? t("Comment resolved successfully")
-          : t("Comment re-opened successfully"),
+        id: notificationId,
+        autoClose: RESOLVE_UNDO_AUTOCLOSE_MS,
+        message: React.createElement(
+          Group,
+          { justify: "space-between", wrap: "nowrap", gap: "md" },
+          React.createElement(
+            Text,
+            { size: "sm" },
+            t("Comment resolved successfully"),
+          ),
+          React.createElement(
+            Button,
+            {
+              variant: "subtle",
+              size: "compact-sm",
+              onClick: () => {
+                if (done) return;
+                done = true;
+                // Reopen via the SAME mutation (read at click time — the
+                // originating item is already unmounted).
+                mutationRef.current?.mutate({
+                  commentId,
+                  pageId,
+                  resolved: false,
+                });
+                // The inline mark is cleared in the reopen mutation's onSuccess
+                // (bound to server confirmation), NOT here — clearing it eagerly
+                // would desync the doc from the panel if reopen then fails.
+                notifications.hide(notificationId);
+              },
+            },
+            t("Undo"),
+          ),
+        ),
       });
     },
   });
+
+  mutationRef.current = mutation;
+  return mutation;
 }

apps/client/src/features/page-history/work-time/format-work-time.test.ts

@@ -0,0 +1,45 @@
+import { describe, it, expect } from "vitest";
+import {
+  formatHeadline,
+  formatDayTotal,
+  formatGapMinutes,
+} from "./format-work-time";
+
+const MIN = 60 * 1000;
+// Fake translator: renders the key with {{tokens}} substituted, so the tests
+// assert the rounding + branch selection without depending on the i18n catalogue.
+const t = (key: string, opts?: Record<string, unknown>) =>
+  key.replace(/\{\{(\w+)\}\}/g, (_, k) => String(opts?.[k] ?? ""));
+
+describe("formatHeadline", () => {
+  it("prefixes ≈ and rounds to a 5-minute step", () => {
+    expect(formatHeadline(4 * 60 * MIN + 27 * MIN, t)).toBe("≈ 4h 25m");
+    expect(formatHeadline(90 * MIN, t)).toBe("≈ 1h 30m");
+  });
+
+  it("shows hours only / minutes only cleanly", () => {
+    expect(formatHeadline(120 * MIN, t)).toBe("≈ 2h");
+    expect(formatHeadline(35 * MIN, t)).toBe("≈ 35m");
+  });
+
+  it("floors a tiny non-zero estimate to 5m, never 0", () => {
+    expect(formatHeadline(2 * MIN, t)).toBe("≈ 5m");
+  });
+
+  it("empty string for zero (widget hidden)", () => {
+    expect(formatHeadline(0, t)).toBe("");
+  });
+});
+
+describe("formatDayTotal", () => {
+  it('renders "h m" and shows — for empty days', () => {
+    expect(formatDayTotal(3 * 60 * MIN + 17 * MIN, t)).toBe("3h 17m");
+    expect(formatDayTotal(0, t)).toBe("—");
+  });
+});
+
+describe("formatGapMinutes", () => {
+  it("converts the tGap ms threshold to whole minutes", () => {
+    expect(formatGapMinutes(15 * MIN)).toBe(15);
+  });
+});

apps/client/src/features/page-history/work-time/format-work-time.ts

@@ -0,0 +1,45 @@
+// #395 — display formatting for the work-time estimate. Pure functions that take
+// a translator so ru-RU / en-US wording lives in the i18n catalogue and the
+// rounding logic stays unit-testable.
+
+type Translate = (key: string, opts?: Record<string, unknown>) => string;
+
+const MIN = 60 * 1000;
+
+function hm(totalMinutes: number): { hours: number; minutes: number } {
+  return {
+    hours: Math.floor(totalMinutes / 60),
+    minutes: totalMinutes % 60,
+  };
+}
+
+/**
+ * Headline number (§6.1): an ESTIMATE, so rounded to a coarse 5-minute step and
+ * prefixed with "≈". A non-zero-but-tiny estimate floors to 5m rather than
+ * rounding down to "0" (which would read as "no work"). Zero → empty string
+ * (the caller hides the widget).
+ */
+export function formatHeadline(workMs: number, t: Translate): string {
+  if (workMs <= 0) return "";
+  let minutes = Math.round(workMs / MIN / 5) * 5;
+  if (minutes === 0) minutes = 5;
+  const { hours, minutes: m } = hm(minutes);
+  if (hours > 0 && m > 0) return t("≈ {{hours}}h {{minutes}}m", { hours, minutes: m });
+  if (hours > 0) return t("≈ {{hours}}h", { hours });
+  return t("≈ {{minutes}}m", { minutes: m });
+}
+
+/** Per-day sum (§6.2), rounded to the minute. Zero → "—". */
+export function formatDayTotal(activeMs: number, t: Translate): string {
+  if (activeMs <= 0) return "—";
+  const minutes = Math.max(1, Math.round(activeMs / MIN));
+  const { hours, minutes: m } = hm(minutes);
+  if (hours > 0 && m > 0) return t("{{hours}}h {{minutes}}m", { hours, minutes: m });
+  if (hours > 0) return t("{{hours}}h", { hours });
+  return t("{{minutes}}m", { minutes: m });
+}
+
+/** The inactivity threshold, for the "estimate · gap = N min" caption. */
+export function formatGapMinutes(tGapMs: number): number {
+  return Math.round(tGapMs / MIN);
+}

apps/client/src/features/page-history/work-time/use-page-work-time.ts

@@ -0,0 +1,25 @@
+import { useQuery, UseQueryResult } from "@tanstack/react-query";
+import { IPageWorkTime } from "./work-time.types";
+import { getPageWorkTime, viewerTimezone } from "./work-time-service";
+
+const WORK_TIME_STALE_TIME = 5 * 60 * 1000;
+
+/**
+ * #395 — the "time worked on this article" estimate + per-day punch-card
+ * buckets. The buckets are computed server-side in the viewer's timezone (so a
+ * midnight-crossing session lands on the right calendar day for the reader).
+ * `enabled` is opt-in so the (cheap but non-trivial) projection query only fires
+ * when the number is actually shown.
+ */
+export function usePageWorkTime(
+  pageId: string,
+  enabled = true,
+): UseQueryResult<IPageWorkTime, Error> {
+  const tz = viewerTimezone();
+  return useQuery({
+    queryKey: ["page-work-time", pageId, tz],
+    queryFn: () => getPageWorkTime(pageId, tz),
+    enabled: enabled && !!pageId,
+    staleTime: WORK_TIME_STALE_TIME,
+  });
+}

apps/client/src/features/page-history/work-time/work-time-punch-card.tsx

@@ -0,0 +1,171 @@
+import { Group, Stack, Text } from "@mantine/core";
+import { useTranslation } from "react-i18next";
+import { useMemo } from "react";
+import { IPageWorkTime, IPerDay, IDayWindow } from "./work-time.types";
+import {
+  formatDayTotal,
+  formatGapMinutes,
+  formatHeadline,
+} from "./format-work-time";
+import classes from "./work-time.module.css";
+
+const DAY_MS = 24 * 60 * 60 * 1000;
+// Collapse a run of this many (or more) consecutive edit-free days into a single
+// "× N days" separator (§6.2 long-range) — the row is still always one day.
+const EMPTY_RUN_COLLAPSE = 8;
+
+type Row =
+  | { type: "day"; day: IPerDay }
+  | { type: "gap"; count: number };
+
+function collapseEmptyRuns(perDay: IPerDay[]): Row[] {
+  const rows: Row[] = [];
+  let emptyRun: IPerDay[] = [];
+  const flush = () => {
+    if (emptyRun.length >= EMPTY_RUN_COLLAPSE) {
+      rows.push({ type: "gap", count: emptyRun.length });
+    } else {
+      for (const d of emptyRun) rows.push({ type: "day", day: d });
+    }
+    emptyRun = [];
+  };
+  for (const d of perDay) {
+    if (d.activeMs === 0 && d.agentMs === 0) {
+      emptyRun.push(d);
+    } else {
+      flush();
+      rows.push({ type: "day", day: d });
+    }
+  }
+  flush();
+  return rows;
+}
+
+function dayHeading(day: number): string {
+  return new Date(day).toLocaleDateString(undefined, {
+    weekday: "short",
+    day: "numeric",
+    month: "short",
+  });
+}
+
+function DayTrack({
+  day,
+  pSingle,
+}: {
+  day: IPerDay;
+  pSingle: number;
+}) {
+  const { t } = useTranslation();
+  const ticks = [6, 12, 18];
+  return (
+    <div className={classes.row}>
+      <span className={classes.dayLabel}>{dayHeading(day.day)}</span>
+      <div className={classes.track}>
+        {ticks.map((h) => (
+          <div
+            key={h}
+            className={classes.hourTick}
+            style={{ left: `${(h / 24) * 100}%` }}
+          />
+        ))}
+        {day.windows.map((w: IDayWindow, i) => {
+          const leftPct = ((w.start - day.day) / DAY_MS) * 100;
+          const widthPct = ((w.end - w.start) / DAY_MS) * 100;
+          const isSingle = w.end - w.start <= pSingle;
+          const cls = [
+            classes.window,
+            w.class === "work" ? classes.windowWork : classes.windowAgent,
+            isSingle ? classes.windowSingle : "",
+          ].join(" ");
+          return (
+            <div
+              key={i}
+              className={cls}
+              style={{
+                left: `${Math.max(0, Math.min(100, leftPct))}%`,
+                width: `${Math.max(0, Math.min(100, widthPct))}%`,
+              }}
+            />
+          );
+        })}
+      </div>
+      <span className={classes.daySum}>
+        {formatDayTotal(day.activeMs, t)}
+      </span>
+    </div>
+  );
+}
+
+interface Props {
+  data: IPageWorkTime;
+}
+
+export default function WorkTimePunchCard({ data }: Props) {
+  const { t } = useTranslation();
+  const rows = useMemo(() => collapseEmptyRuns(data.perDay), [data.perDay]);
+  const gapMin = formatGapMinutes(data.config.tGap);
+
+  if (data.workMs <= 0 && data.agentOnlyMs <= 0) {
+    return (
+      <Text size="sm" c="dimmed" py="md">
+        {t("No editing activity recorded yet.")}
+      </Text>
+    );
+  }
+
+  return (
+    <Stack gap="xs">
+      <Group gap="lg">
+        <Text size="sm" fw={500}>
+          {formatHeadline(data.workMs, t)}
+        </Text>
+        {data.agentOnlyMs > 0 && (
+          <Text size="xs" c="dimmed">
+            {t("agent: {{value}}", { value: formatHeadline(data.agentOnlyMs, t) })}
+          </Text>
+        )}
+      </Group>
+
+      <Group gap="md">
+        <Text size="xs" c="dimmed">
+          <span
+            className={`${classes.legendSwatch} ${classes.windowWork}`}
+            style={{ marginRight: 4 }}
+          />
+          {t("Work")}
+        </Text>
+        <Text size="xs" c="dimmed">
+          <span
+            className={`${classes.legendSwatch} ${classes.windowAgent}`}
+            style={{ marginRight: 4 }}
+          />
+          {t("Agent")}
+        </Text>
+      </Group>
+
+      <div>
+        {rows.map((row, i) =>
+          row.type === "day" ? (
+            <DayTrack
+              key={row.day.dayISO}
+              day={row.day}
+              pSingle={data.config.pSingle}
+            />
+          ) : (
+            <div key={`gap-${i}`} className={classes.gapRow}>
+              {t("× {{count}} days without edits", { count: row.count })}
+            </div>
+          ),
+        )}
+      </div>
+
+      <Text size="xs" c="dimmed" mt="xs">
+        {t("Estimate · timezone {{tz}} · inactivity gap {{gap}} min", {
+          tz: data.tz,
+          gap: gapMin,
+        })}
+      </Text>
+    </Stack>
+  );
+}

apps/client/src/features/page-history/work-time/work-time-service.ts

@@ -0,0 +1,23 @@
+import api from "@/lib/api-client";
+import { IPageWorkTime } from "./work-time.types";
+
+/** The viewer's IANA timezone (browser locale) — the punch-card lays days out
+ *  in "my evenings", per §6.3/§10. Falls back to UTC if the runtime hides it. */
+export function viewerTimezone(): string {
+  try {
+    return Intl.DateTimeFormat().resolvedOptions().timeZone || "UTC";
+  } catch {
+    return "UTC";
+  }
+}
+
+export async function getPageWorkTime(
+  pageId: string,
+  tz: string,
+): Promise<IPageWorkTime> {
+  const req = await api.post<IPageWorkTime>("/pages/history/time", {
+    pageId,
+    tz,
+  });
+  return req.data;
+}

apps/client/src/features/page-history/work-time/work-time-stat.tsx

@@ -0,0 +1,69 @@
+import { Modal, Text, Tooltip, UnstyledButton } from "@mantine/core";
+import { useDisclosure } from "@mantine/hooks";
+import { IconClockHour4 } from "@tabler/icons-react";
+import { useTranslation } from "react-i18next";
+import { usePageWorkTime } from "./use-page-work-time";
+import { formatGapMinutes, formatHeadline } from "./format-work-time";
+import WorkTimePunchCard from "./work-time-punch-card";
+
+interface Props {
+  pageId: string;
+}
+
+/**
+ * #395 — the clickable "time worked on this article" headline (§6.1). Renders
+ * the `work` estimate with a "≈" sign and the inactivity threshold in a tooltip
+ * (it is an estimate, not a stopwatch). Clicking opens the daily punch-card
+ * (§6.2). Renders nothing until there is a non-zero human OR agent estimate, so a
+ * brand-new / never-edited page shows no widget. For an agent-only-edited page
+ * (workMs===0, agentOnlyMs>0) the headline shows the agent estimate (labelled
+ * `agent:`, matching the punch-card) so the punch-card stays reachable (#395:
+ * "how much a HUMAN and separately the AGENT").
+ */
+export default function WorkTimeStat({ pageId }: Props) {
+  const { t } = useTranslation();
+  const [opened, { open, close }] = useDisclosure(false);
+  const { data } = usePageWorkTime(pageId);
+
+  if (!data || (data.workMs <= 0 && data.agentOnlyMs <= 0)) return null;
+
+  const agentOnly = data.workMs <= 0;
+  const label = agentOnly
+    ? t("agent: {{value}}", { value: formatHeadline(data.agentOnlyMs, t) })
+    : formatHeadline(data.workMs, t);
+  const gapMin = formatGapMinutes(data.config.tGap);
+
+  return (
+    <>
+      <Tooltip
+        label={t("Estimated time worked (inactivity gap {{gap}} min)", {
+          gap: gapMin,
+        })}
+        position="bottom"
+      >
+        <UnstyledButton
+          onClick={open}
+          aria-label={t("Show time worked on this page")}
+        >
+          <Text
+            size="xs"
+            c="dimmed"
+            style={{ display: "inline-flex", alignItems: "center", gap: 4 }}
+          >
+            <IconClockHour4 size={14} />
+            {label}
+          </Text>
+        </UnstyledButton>
+      </Tooltip>
+
+      <Modal
+        opened={opened}
+        onClose={close}
+        title={t("Time worked on this article")}
+        size="lg"
+      >
+        <WorkTimePunchCard data={data} />
+      </Modal>
+    </>
+  );
+}

apps/client/src/features/page-history/work-time/work-time.module.css

@@ -0,0 +1,82 @@
+/* #395 — 24h × days punch-card. Custom CSS segments on a fixed 24-hour track
+   (position = offset-in-day / 24h, width = duration / 24h), no chart library. */
+
+.row {
+  display: grid;
+  grid-template-columns: 96px 1fr 64px;
+  align-items: center;
+  gap: 12px;
+  padding: 3px 0;
+}
+
+.dayLabel {
+  font-size: var(--mantine-font-size-xs);
+  color: var(--mantine-color-dimmed);
+  white-space: nowrap;
+}
+
+.track {
+  position: relative;
+  height: 16px;
+  border-radius: 4px;
+  background-color: light-dark(
+    var(--mantine-color-gray-1),
+    var(--mantine-color-dark-6)
+  );
+  overflow: hidden;
+}
+
+/* Faint hour grid so the eye can read "morning vs evening". */
+.hourTick {
+  position: absolute;
+  top: 0;
+  bottom: 0;
+  width: 1px;
+  background-color: light-dark(
+    var(--mantine-color-gray-3),
+    var(--mantine-color-dark-4)
+  );
+}
+
+.window {
+  position: absolute;
+  top: 2px;
+  bottom: 2px;
+  border-radius: 3px;
+  min-width: 3px;
+}
+
+.windowWork {
+  background-color: var(--mantine-color-blue-5);
+}
+
+.windowAgent {
+  background-color: var(--mantine-color-grape-5);
+}
+
+/* A lone single-sample (P_single) window: minimal + dimmed, so it neither
+   vanishes nor fakes dense work (§6.2). */
+.windowSingle {
+  opacity: 0.5;
+}
+
+.daySum {
+  font-size: var(--mantine-font-size-xs);
+  text-align: right;
+  white-space: nowrap;
+}
+
+.gapRow {
+  padding: 6px 0 6px 108px;
+  font-size: var(--mantine-font-size-xs);
+  color: var(--mantine-color-dimmed);
+  font-style: italic;
+}
+
+.legendSwatch {
+  display: inline-block;
+  width: 12px;
+  height: 12px;
+  border-radius: 3px;
+  vertical-align: middle;
+}

apps/client/src/features/page-history/work-time/work-time.types.ts

@@ -0,0 +1,37 @@
+// #395 — client-side mirror of the server work-time payload
+// (apps/server/src/core/page/work-time). Shapes returned by POST /pages/history/time.
+
+export type WorkSessionClass = "work" | "agent_only";
+
+export interface IDayWindow {
+  start: number;
+  end: number;
+  class: WorkSessionClass;
+}
+
+export interface IPerDay {
+  day: number;
+  dayISO: string;
+  activeMs: number;
+  agentMs: number;
+  windows: IDayWindow[];
+}
+
+export interface IWorkTimeConfig {
+  tGap: number;
+  agentTGap: number;
+  pIn: number;
+  pOut: number;
+  pSingle: number;
+  excludeGit: boolean;
+  burstCapMs?: number;
+  dedupRoundMs: number;
+}
+
+export interface IPageWorkTime {
+  workMs: number;
+  agentOnlyMs: number;
+  perDay: IPerDay[];
+  config: IWorkTimeConfig;
+  tz: string;
+}

apps/client/src/features/page/components/header/page-header-menu.tsx

@@ -51,6 +51,7 @@ import {
 import { formattedDate } from "@/lib/time.ts";
 import { PageEditModeToggle } from "@/features/user/components/page-state-pref.tsx";
 import MovePageModal from "@/features/page/components/move-page-modal.tsx";
+import WorkTimeStat from "@/features/page-history/work-time/work-time-stat.tsx";
 import { useTimeAgo } from "@/hooks/use-time-ago.tsx";
 import {
   useFavoriteIds,
@@ -265,6 +266,8 @@ function PageActionMenu({ readOnly, onSaveVersion }: PageActionMenuProps) {
 
   return (
     <>
+      {page?.id && <WorkTimeStat pageId={page.id} />}
+
       <Menu
         shadow="xl"
         position="bottom-end"

apps/client/src/features/page/queries/page-query.ts

@@ -665,6 +665,13 @@ export function updateCacheOnMovePage(
   pageData: Partial<IPage>,
 ) {
   invalidatePageTree();
+  // Invalidate the moved page's breadcrumbs (#523). The tree-side child-loss
+  // guard removes the moved node from the local tree when its new parent is an
+  // unloaded branch, so `findBreadcrumbPath` misses it and the breadcrumb bar
+  // falls back to the server `["breadcrumbs", pageId]` query — which this move
+  // must invalidate, otherwise the crumbs keep showing the OLD parent until a
+  // refocus/navigation.
+  queryClient.invalidateQueries({ queryKey: ["breadcrumbs", pageId] });
   // Remove page from old parent's cache
   const oldQueryKey =
     oldParentId === null

apps/client/src/features/page/queries/update-cache-on-move.breadcrumbs.test.ts

@@ -0,0 +1,40 @@
+import { describe, it, expect, beforeEach, vi } from "vitest";
+import type { IPage } from "@/features/page/types/page.types";
+
+// A fresh QueryClient stands in for the app singleton (importing the real
+// @/main.tsx would run ReactDOM.createRoot, which has no DOM root in jsdom).
+vi.mock("@/main.tsx", async () => {
+  const { QueryClient } = await import("@tanstack/react-query");
+  return { queryClient: new QueryClient() };
+});
+
+import { queryClient } from "@/main.tsx";
+import { updateCacheOnMovePage } from "./page-query";
+
+// #523: the tree-side child-loss guard removes the moved node from the local
+// tree when its new parent is an unloaded branch, so `findBreadcrumbPath` misses
+// it and the breadcrumb bar falls back to the server `["breadcrumbs", pageId]`
+// query. That query MUST be invalidated by a move, or the crumbs keep showing
+// the OLD parent until a refocus/navigation.
+describe("updateCacheOnMovePage — breadcrumbs invalidation (#523)", () => {
+  beforeEach(() => {
+    queryClient.clear();
+    vi.restoreAllMocks();
+  });
+
+  it("invalidates the moved page's ['breadcrumbs', pageId] query", () => {
+    const spy = vi.spyOn(queryClient, "invalidateQueries");
+
+    updateCacheOnMovePage("s1", "moved-page", "old-parent", "new-parent", {
+      id: "moved-page",
+    } as Partial<IPage>);
+
+    const invalidatedBreadcrumbs = spy.mock.calls.some(
+      ([arg]) =>
+        Array.isArray((arg as { queryKey?: unknown[] })?.queryKey) &&
+        (arg as { queryKey: unknown[] }).queryKey[0] === "breadcrumbs" &&
+        (arg as { queryKey: unknown[] }).queryKey[1] === "moved-page",
+    );
+    expect(invalidatedBreadcrumbs).toBe(true);
+  });
+});

apps/client/src/features/page/tree/components/doc-tree-row.tsx

@@ -55,7 +55,14 @@ type Props<T extends object> = {
 };
 
 const DRAG_TYPE = 'doc-tree-item';
-const AUTO_EXPAND_MS = 500;
+// Hover-hold before a collapsed row auto-expands during a drag. 2s (not ~0.5s)
+// so merely dragging the cursor THROUGH the tree never expands rows — only a
+// deliberate hold does (#523).
+const AUTO_EXPAND_MS = 2000;
+// How long the "a page just moved in here" cue stays on a collapsed target after
+// a make-child drop. Long enough to notice at a glance during frequent
+// collapsed-drops, short enough not to linger. Code-only UX constant (#523).
+const DROP_LANDED_HIGHLIGHT_MS = 1800;
 
 function DocTreeRowInner<T extends object>(props: Props<T>) {
   const {
@@ -93,7 +100,11 @@ function DocTreeRowInner<T extends object>(props: Props<T>) {
   const rowRef = useRef<HTMLElement>(null);
   const [isDragging, setIsDragging] = useState(false);
   const [instruction, setInstruction] = useState<Instruction | null>(null);
+  // Transient "just received a child" cue: a make-child drop no longer expands
+  // the (collapsed) target, so flash the row instead so the move isn't invisible.
+  const [landedChild, setLandedChild] = useState(false);
   const autoExpandTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  const landedChildTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   const cancelAutoExpand = useCallback(() => {
     if (autoExpandTimerRef.current) {
@@ -249,11 +260,24 @@ function DocTreeRowInner<T extends object>(props: Props<T>) {
               ? getDragLabel(sourceNode)
               : 'item';
             liveRegion.announce(`Moved ${sourceLabel} under ${parentName}.`);
-            // After a make-child drop, expand this row so the user sees the
-            // just-dropped child — especially important when the row had no
-            // children before (chevron just appeared) so the drop would
-            // otherwise be invisible.
-            if (op.kind === 'make-child') onToggle(node.id, true);
+            // Do NOT auto-expand the target on drop: a drop must leave the node
+            // collapsed. Intentional expansion is handled solely by the
+            // hover-hold timer (AUTO_EXPAND_MS). Feedback that the drop landed is
+            // given by the post-move flash + landed-cue highlight + live-region
+            // announce above. When the make-child target is collapsed, flash a
+            // distinct "child moved in here" cue on the row (it stays collapsed).
+            if (op.kind === 'make-child' && !isOpen) {
+              if (landedChildTimerRef.current) {
+                clearTimeout(landedChildTimerRef.current);
+              }
+              setLandedChild(true);
+              landedChildTimerRef.current = setTimeout(() => {
+                setLandedChild(false);
+                landedChildTimerRef.current = null;
+              }, DROP_LANDED_HIGHLIGHT_MS);
+            }
+            // Restore the openness of the MOVED page itself (source) — untouched
+            // by the above; the target is never expanded here.
             if (source.data.isOpenOnDragStart) onToggle(sourceId, true);
           },
         }),
@@ -281,6 +305,17 @@ function DocTreeRowInner<T extends object>(props: Props<T>) {
 
   useEffect(() => () => cancelAutoExpand(), [cancelAutoExpand]);
 
+  // Clear the landed-child cue timer on unmount (mirrors autoExpandTimerRef).
+  useEffect(
+    () => () => {
+      if (landedChildTimerRef.current) {
+        clearTimeout(landedChildTimerRef.current);
+        landedChildTimerRef.current = null;
+      }
+    },
+    [],
+  );
+
   const effectiveInst =
     instruction?.type === 'instruction-blocked'
       ? instruction.desired
@@ -317,6 +352,7 @@ function DocTreeRowInner<T extends object>(props: Props<T>) {
         className={styles.node}
         data-dragging={isDragging || undefined}
         data-selected={isSelected || undefined}
+        data-landed-child={landedChild || undefined}
         data-receiving-drop={
           receivingDrop === 'make-child'
             ? blocked

apps/client/src/features/page/tree/hooks/use-tree-mutation.move-guard.test.tsx

@@ -0,0 +1,149 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+import { renderHook, act } from "@testing-library/react";
+import { Provider, createStore } from "jotai";
+import type { ReactNode } from "react";
+
+import { treeDataAtom } from "@/features/page/tree/atoms/tree-data-atom.ts";
+import { treeModel } from "@/features/page/tree/model/tree-model";
+import type { SpaceTreeNode } from "@/features/page/tree/types";
+
+// --- Boundary mocks: only the network/query + router/i18n surfaces the hook
+// touches. The tree math (treeModel, dropOpToMovePayload) runs for real so the
+// child-loss guard is exercised end-to-end.
+const moveMutate = vi.fn().mockResolvedValue({});
+const updateCacheOnMovePageMock = vi.fn();
+vi.mock("@/features/page/queries/page-query.ts", () => ({
+  useCreatePageMutation: () => ({ mutateAsync: vi.fn() }),
+  useUpdatePageMutation: () => ({ mutateAsync: vi.fn() }),
+  useRemovePageMutation: () => ({ mutateAsync: vi.fn() }),
+  useMovePageMutation: () => ({ mutateAsync: moveMutate }),
+  updateCacheOnMovePage: (...args: unknown[]) =>
+    updateCacheOnMovePageMock(...args),
+}));
+vi.mock("react-router-dom", () => ({
+  useNavigate: () => vi.fn(),
+  useParams: () => ({ spaceSlug: "space", pageSlug: undefined }),
+}));
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (s: string) => s }),
+}));
+vi.mock("@mantine/notifications", () => ({
+  notifications: { show: vi.fn() },
+}));
+
+// Import AFTER mocks so the hook binds to them.
+import { useTreeMutation } from "./use-tree-mutation";
+
+function node(
+  id: string,
+  over: Partial<SpaceTreeNode> = {},
+): SpaceTreeNode {
+  return {
+    id,
+    slugId: `slug-${id}`,
+    name: id.toUpperCase(),
+    position: "a0",
+    spaceId: "space-1",
+    parentPageId: null as unknown as string,
+    hasChildren: false,
+    children: [],
+    ...over,
+  };
+}
+
+function setup(before: SpaceTreeNode[]) {
+  const store = createStore();
+  store.set(treeDataAtom, before);
+  const wrapper = ({ children }: { children: ReactNode }) => (
+    <Provider store={store}>{children}</Provider>
+  );
+  const { result } = renderHook(() => useTreeMutation("space-1"), { wrapper });
+  return { store, result };
+}
+
+describe("useTreeMutation.handleMove — child-loss guard (#523)", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    moveMutate.mockResolvedValue({});
+  });
+
+  it("make-child into an UNLOADED folder does NOT materialize [source] — keeps it lazy-loadable", async () => {
+    // F has children on the server but none are loaded here (canonical unloaded
+    // form: hasChildren + children:[]). X sits at root.
+    const before = [
+      node("F", { position: "a0", hasChildren: true, children: [] }),
+      node("X", { position: "a5" }),
+    ];
+    const { store, result } = setup(before);
+
+    await act(async () => {
+      await result.current.handleMove("X", {
+        kind: "make-child",
+        targetId: "F",
+      });
+    });
+
+    const tree = store.get(treeDataAtom);
+    const f = treeModel.find(tree, "F");
+    // The guard leaves F unloaded (children stay []), so a later expand fetches
+    // the FULL server set (incl. X) instead of showing a misleading partial [X].
+    // MUTATION: dropping the guard (using the `move` result) would put children
+    // === [X] here and redden this.
+    expect(f?.children).toEqual([]);
+    expect(f?.hasChildren).toBe(true);
+    // X is removed from its old (root) slot; it reappears on expand/load of F.
+    expect(treeModel.find(tree, "X")).toBeNull();
+    // The server move is still persisted.
+    expect(moveMutate).toHaveBeenCalledTimes(1);
+  });
+
+  it("make-child into a LOADED folder appends source and KEEPS the existing children", async () => {
+    // F is loaded with one child c1; moving X in must preserve c1 (no loss) and
+    // append X — this path does NOT hit the guard.
+    const before = [
+      node("F", {
+        position: "a0",
+        hasChildren: true,
+        children: [node("c1", { position: "a1", parentPageId: "F" })],
+      }),
+      node("X", { position: "a5" }),
+    ];
+    const { store, result } = setup(before);
+
+    await act(async () => {
+      await result.current.handleMove("X", {
+        kind: "make-child",
+        targetId: "F",
+      });
+    });
+
+    const tree = store.get(treeDataAtom);
+    const f = treeModel.find(tree, "F");
+    expect(f?.children?.map((n) => n.id)).toEqual(["c1", "X"]);
+    expect(f?.hasChildren).toBe(true);
+    // X now lives under F.
+    expect(treeModel.find(tree, "X")?.parentPageId).toBe("F");
+  });
+
+  it("make-child into a genuinely-empty leaf materializes the child (nothing to lose)", async () => {
+    // Leaf L has no server children (hasChildren:false). Dropping X in should
+    // show X immediately — the guard must NOT fire here.
+    const before = [
+      node("L", { position: "a0", hasChildren: false, children: [] }),
+      node("X", { position: "a5" }),
+    ];
+    const { store, result } = setup(before);
+
+    await act(async () => {
+      await result.current.handleMove("X", {
+        kind: "make-child",
+        targetId: "L",
+      });
+    });
+
+    const tree = store.get(treeDataAtom);
+    const l = treeModel.find(tree, "L");
+    expect(l?.children?.map((n) => n.id)).toEqual(["X"]);
+    expect(l?.hasChildren).toBe(true);
+  });
+});

apps/client/src/features/page/tree/hooks/use-tree-mutation.ts

@@ -61,11 +61,48 @@ export function useTreeMutation(spaceId: string): UseTreeMutation {
       if (!source) return;
       const oldParentId = source.parentPageId ?? null;
 
-      // optimistic apply with the new position from the payload
-      let optimistic = treeModel.update(after, sourceId, {
-        position: payload.position,
-        parentPageId: payload.parentPageId,
-      } as Partial<SpaceTreeNode>);
+      // Child-loss guard (#523, twin of #525's realtime `insertByPosition` fix).
+      // We no longer auto-expand the make-child target on drop, so the old
+      // `onToggle(target, true)` — which was ALSO the only trigger of the
+      // corrective lazy-load — is gone. `treeModel.move` materialized
+      // `target.children = [source]` (only the moved node); if the target is an
+      // UNLOADED branch (server has children but none are loaded here), keeping
+      // that partial `[source]` list would defeat the lazy-load gate and hide the
+      // target's OTHER server children (the #159 #1 data-loss class). So for an
+      // unloaded make-child target, build the optimistic tree WITHOUT
+      // materializing source under it: just remove source from its old parent and
+      // flag the target `hasChildren`. The gate stays armed and a later manual
+      // expand fetches the FULL set (incl. the moved page, which the awaited
+      // server move persists). Predicate is the gate's (`isUnloadedBranch`), NOT
+      // `insertByPosition`'s old `=== undefined` (canonical unloaded is `[]`).
+      const target =
+        op.kind === "make-child"
+          ? (treeModel.find(before, op.targetId) as SpaceTreeNode | null)
+          : null;
+      const unloadedMakeChild =
+        op.kind === "make-child" && treeModel.isUnloadedBranch(target);
+
+      let optimistic: SpaceTreeNode[];
+      if (unloadedMakeChild) {
+        // Do NOT materialize [source] into the unloaded target.
+        optimistic = treeModel.remove(before, sourceId);
+        optimistic = treeModel.update(optimistic, op.targetId, {
+          hasChildren: true,
+        } as Partial<SpaceTreeNode>);
+      } else {
+        // optimistic apply with the new position from the payload
+        optimistic = treeModel.update(after, sourceId, {
+          position: payload.position,
+          parentPageId: payload.parentPageId,
+        } as Partial<SpaceTreeNode>);
+        // For make-child onto a previously-childless (loaded) target: flip
+        // hasChildren on so the new parent shows its chevron.
+        if (op.kind === "make-child") {
+          optimistic = treeModel.update(optimistic, op.targetId, {
+            hasChildren: true,
+          } as Partial<SpaceTreeNode>);
+        }
+      }
 
       // If the old parent has no children left, mark hasChildren: false so the
       // chevron disappears. Without this, the empty parent keeps rendering an
@@ -79,14 +116,6 @@ export function useTreeMutation(spaceId: string): UseTreeMutation {
         }
       }
 
-      // For make-child onto a previously-childless target: flip hasChildren on
-      // so the new parent shows its chevron.
-      if (op.kind === "make-child") {
-        optimistic = treeModel.update(optimistic, op.targetId, {
-          hasChildren: true,
-        } as Partial<SpaceTreeNode>);
-      }
-
       setData(optimistic);
 
       try {

apps/client/src/features/page/tree/styles/tree.module.css

@@ -150,6 +150,45 @@
     );
 }
 
+/* "A page just moved in here" cue (#523). A make-child drop no longer expands
+   the collapsed target, so the moved page is momentarily invisible; pulse the
+   target row in a distinct teal (NOT the blue make-child highlight, NOT the
+   neutral post-move flash) so the landing is noticeable while the node stays
+   collapsed. Two short pulses fit inside DROP_LANDED_HIGHLIGHT_MS (1.8s), after
+   which the row clears the attribute and the animation stops. */
+@keyframes landedChildPulse {
+    0% {
+        background-color: light-dark(
+            var(--mantine-color-teal-2),
+            rgba(45, 212, 191, 0.30)
+        );
+        outline-color: light-dark(
+            var(--mantine-color-teal-6),
+            var(--mantine-color-teal-5)
+        );
+    }
+    100% {
+        background-color: transparent;
+        outline-color: transparent;
+    }
+}
+
+.node[data-landed-child="true"] {
+    outline: 2px solid transparent;
+    outline-offset: -1px;
+    animation: landedChildPulse 0.9s ease-out 2;
+}
+
+@media (prefers-reduced-motion: reduce) {
+    .node[data-landed-child="true"] {
+        animation: none;
+        background-color: light-dark(
+            var(--mantine-color-teal-1),
+            rgba(45, 212, 191, 0.18)
+        );
+    }
+}
+
 .dropLine {
     position: absolute;
     left: var(--drop-line-indent, 0);

apps/client/src/features/search/components/search-spotlight.tsx

@@ -1,6 +1,6 @@
 import { Spotlight } from "@mantine/spotlight";
 import { IconSearch } from "@tabler/icons-react";
-import { Group, VisuallyHidden } from "@mantine/core";
+import { Group, Text, VisuallyHidden } from "@mantine/core";
 import { useState, useMemo } from "react";
 import { useDebouncedValue } from "@mantine/hooks";
 import { useTranslation } from "react-i18next";
@@ -84,6 +84,11 @@ export function SearchSpotlight({ spaceId }: SearchSpotlightProps) {
             onFiltersChange={handleFiltersChange}
             spaceId={spaceId}
           />
+          {/* #529: operator hint — matches ANY word by default; "…" for an exact
+              phrase, +term to require, -term to exclude. */}
+          <Text size="xs" c="dimmed" mt={4}>
+            {t('Tip: "exact phrase", +required, -excluded')}
+          </Text>
         </div>
 
         <VisuallyHidden role="status" aria-live="polite">

apps/client/src/features/search/types/search.types.ts

@@ -5,6 +5,9 @@ import { IPage } from "@/features/page/types/page.types.ts";
 
 export interface IPageSearch {
   id: string;
+  // #529 A7 superset: `pageId` aliases `id`; `rank`/`highlight` are null for
+  // substring-only hits (the UI already falls back to the title/snippet).
+  pageId?: string;
   title: string;
   icon: string;
   parentPageId: string;
@@ -12,9 +15,36 @@ export interface IPageSearch {
   creatorId: string;
   createdAt: Date;
   updatedAt: Date;
-  rank: string;
-  highlight: string;
+  rank: string | number | null;
+  highlight: string | null;
   space: Partial<ISpace>;
+  // New #529 fields (present from the native Postgres search driver).
+  snippet?: string;
+  score?: number;
+  path?: string[];
+  matchedFields?: string[];
+  matchedTerms?: string[];
+}
+
+// #529 A5 pagination envelope returned by POST /search (native driver). The web
+// list helpers read `items`; these travel alongside for pagination + diagnostics.
+export interface IPageSearchResponse {
+  items: IPageSearch[];
+  total: number;
+  hasMore: boolean;
+  truncatedAtCap: boolean;
+  offset: number;
+  query?: {
+    raw: string;
+    parsed: {
+      positive: string[];
+      required: string[];
+      excluded: string[];
+      reason?: string;
+    };
+    mode: "or" | "and";
+    match: string;
+  };
 }
 
 export interface SearchSuggestionParams {
@@ -37,6 +67,10 @@ export interface IPageSearchParams {
   query: string;
   spaceId?: string;
   shareId?: string;
+  // #529 A9: match mode (auto default) + pagination.
+  match?: "auto" | "word" | "prefix" | "substring";
+  limit?: number;
+  offset?: number;
 }
 
 export interface IAttachmentSearch {

apps/client/src/lib/telemetry/route-template.ts

@@ -36,6 +36,7 @@ const STATIC_ROUTES = new Set<string>([
   '/setup/register',
   '/settings/account/profile',
   '/settings/account/preferences',
+  '/settings/account/api-keys',
   '/settings/workspace',
   '/settings/ai',
   '/settings/members',

apps/client/src/pages/settings/account/account-api-keys.tsx

@@ -0,0 +1,22 @@
+import SettingsTitle from "@/components/settings/settings-title.tsx";
+import ApiKeysManager from "@/features/api-key/components/api-keys-manager";
+import { getAppName } from "@/lib/config.ts";
+import { Helmet } from "react-helmet-async";
+import { useTranslation } from "react-i18next";
+
+export default function AccountApiKeys() {
+  const { t } = useTranslation();
+
+  return (
+    <>
+      <Helmet>
+        <title>
+          {t("API keys")} - {getAppName()}
+        </title>
+      </Helmet>
+      <SettingsTitle title={t("API keys")} />
+
+      <ApiKeysManager />
+    </>
+  );
+}

apps/client/src/styles/notification-overrides.css

@@ -14,9 +14,11 @@
  */
 
 /*
- * Push the top-anchored toast containers below the top chrome (fixed 45px
- * header + optional 45px format toolbar + ~6px gap) so a toast (z-index 10000)
- * neither covers nor intercepts clicks on the header/toolbar (both z-index 99).
+ * Anchor the top toast containers to the very top edge of the viewport (a small
+ * 8px gap), above the header/search chrome, per product request. The toast
+ * (z-index 10000) therefore renders over the header/toolbar (both z-index 99)
+ * for the few seconds it is visible — intentional, since it is the top-most,
+ * in-the-line-of-sight surface.
  *
  * Scoped to [data-position^='top'] on purpose. Mantine renders ALL SIX position
  * containers simultaneously (`position` only routes toasts into one via the
@@ -34,7 +36,7 @@
  * (the :where() contributes 0), regardless of stylesheet order.
  */
 .mantine-Notifications-root[data-position^='top'] {
-  top: 96px;
+  top: 8px;
 }
 
 [data-mantine-color-scheme='light'] .mantine-Notification-root {

apps/client/vitest.setup.ts

@@ -63,3 +63,15 @@ vi.stubGlobal("matchMedia", (query: string) => ({
   removeEventListener: vi.fn(),
   dispatchEvent: vi.fn(),
 }));
+
+// Mantine's ScrollArea (used by Table.ScrollContainer, ScrollArea, etc.) reads
+// `ResizeObserver` in a layout effect on mount, which jsdom does not implement.
+// A no-op stub lets any test rendering those components mount cleanly.
+vi.stubGlobal(
+  "ResizeObserver",
+  class {
+    observe() {}
+    unobserve() {}
+    disconnect() {}
+  },
+);

apps/server/package.json

@@ -23,7 +23,7 @@
     "migration:reset": "tsx src/database/migrate.ts down-to NO_MIGRATIONS",
     "migration:codegen": "kysely-codegen --dialect=postgres --camel-case --env-file=../../.env --out-file=./src/database/types/db.d.ts",
     "lint": "eslint \"{src,apps,libs,test}/**/*.ts\" --fix",
-    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build && pnpm --filter @docmost/token-estimate build",
+    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build && pnpm --filter @docmost/token-estimate build && pnpm --filter @docmost/mcp build",
     "test": "jest",
     "test:int": "jest --config test/jest-integration.json",
     "test:watch": "jest --watch",

apps/server/src/collaboration/collaboration.util.ts

@@ -141,7 +141,57 @@ export function htmlToJson(html: string) {
   }
 }
 
-export function jsonToText(tiptapJson: JSONContent) {
+/**
+ * Deterministic text-serializer overrides for the `format:"text"` page read
+ * (#502). Non-text nodes render to a STABLE placeholder instead of their
+ * (structure-dependent) inner text, so a machine diff of two text reads is
+ * driven only by the page's actual prose — output stability across package
+ * versions IS the contract (pinned by a snapshot test). Returning a string from
+ * a `textSerializer` also stops `generateText` descending into the node, so a
+ * table renders as ONE token rather than its flattened cell text.
+ *
+ * Only nodes with no meaningful flat-text form are overridden; every other node
+ * (paragraph/heading/list/code/blockquote/callout/…) keeps its natural text so
+ * a config written as markdown reads back byte-identical.
+ */
+const TEXT_READ_SERIALIZERS: Record<string, (props: { node: any }) => string> =
+  {
+    // Image atom: no inner text -> a fixed placeholder.
+    image: () => '[image]',
+    // Table: `[table RxC]` where R = row count, C = the first row's cell count
+    // (a table's columns are uniform per the schema). Computed from the PM node,
+    // so it is independent of cell contents.
+    table: ({ node }) => {
+      const rows = node?.childCount ?? 0;
+      const cols = rows > 0 ? (node.child(0)?.childCount ?? 0) : 0;
+      return `[table ${rows}x${cols}]`;
+    },
+  };
+
+/**
+ * Serialize a ProseMirror/TipTap document to plain text.
+ *
+ * Default (no options): the long-standing search-index behavior — bare
+ * concatenated node text with `generateText`'s default `\n\n` block separator.
+ * This feeds the page `textContent` tsvector and MUST NOT change.
+ *
+ * `deterministic:true` (#502 `format:"text"` page read): a flat, machine-diffable
+ * rendering — one line per block (`\n` block separator; `hardBreak` already
+ * serializes to `\n`), inline marks/anchors/autoformat dropped, and non-text
+ * nodes replaced by the stable placeholders above (`[image]`, `[table RxC]`).
+ */
+export function jsonToText(
+  // `any` (like jsonToHtml/jsonToMarkdown) so a loosely-typed DB `page.content`
+  // (JsonValue) can be passed straight through, as the controller does.
+  tiptapJson: any,
+  options?: { deterministic?: boolean },
+) {
+  if (options?.deterministic) {
+    return generateText(tiptapJson, tiptapExtensions, {
+      blockSeparator: '\n',
+      textSerializers: TEXT_READ_SERIALIZERS,
+    });
+  }
   return generateText(tiptapJson, tiptapExtensions);
 }
 

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

@@ -736,6 +736,53 @@ describe('PersistenceExtension.onStoreDocument — Approach-A boundary snapshot'
       expect(historyQueue.remove).toHaveBeenCalledWith(PAGE_ID);
       expect(historyQueue.remove).not.toHaveBeenCalledWith(SLUG);
     });
+
+    // #370 F2 — an effectively-empty page is a REACHABLE no-op (agent calls
+    // save_page_version on a blank page): the version tx short-circuits with
+    // nothing to pin. The handler MUST still broadcast a TERMINAL reply
+    // (version.skipped, reason:'empty') so the client resolves at once instead of
+    // waiting out its 20s ack timeout and misreporting a healthy server as
+    // unreachable. MUTATION: drop the `else if (skipped)` broadcast → no terminal
+    // reply is sent → this reddens.
+    it('empty page → no version written, broadcasts a terminal version.skipped(empty)', async () => {
+      const emptyDoc = { type: 'doc', content: [{ type: 'paragraph' }] };
+      const document = ydocFor(emptyDoc);
+      pageRepo.findById.mockResolvedValue({
+        ...persistedHumanPage('IGNORED'),
+        content: emptyDoc,
+      });
+
+      await emitSave(document, 'agent');
+
+      // Nothing pinned.
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      expect(pageHistoryRepo.updateHistoryKind).not.toHaveBeenCalled();
+      // But a terminal reply WAS sent so the client never times out. The flush
+      // (onStoreDocument) emits its own `page.updated`; the version.skipped is the
+      // LAST broadcast (dropping the skip branch leaves page.updated last → reds).
+      const calls = (document as any).broadcastStateless.mock.calls;
+      const msg = JSON.parse(calls[calls.length - 1][0]);
+      expect(msg).toEqual({ type: 'version.skipped', reason: 'empty' });
+    });
+
+    // #370 F2 — the page row is gone (deleted / never persisted). Same rule: a
+    // terminal reply MUST be sent (version.skipped, reason:'page-not-found') so the
+    // client surfaces a truthful "not found" immediately rather than a health
+    // timeout. onStoreDocument's own `!page` guard returns early without throwing,
+    // so the handler reaches the version tx and its `!page` skip branch.
+    it('page not found → broadcasts a terminal version.skipped(page-not-found)', async () => {
+      const document = ydocFor(doc('GONE'));
+      pageRepo.findById.mockResolvedValue(null);
+
+      await emitSave(document, 'agent');
+
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      expect((document as any).broadcastStateless).toHaveBeenCalledTimes(1);
+      const msg = JSON.parse(
+        (document as any).broadcastStateless.mock.calls[0][0],
+      );
+      expect(msg).toEqual({ type: 'version.skipped', reason: 'page-not-found' });
+    });
   });
 
   // #370 — the in-memory idle-burst marker must be dropped on doc unload (like

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

@@ -68,6 +68,19 @@ export const INTENTIONAL_CLEAR_MESSAGE_TYPE = 'intentional-clear';
  */
 export const SAVE_VERSION_MESSAGE_TYPE = 'save-version';
 
+/**
+ * #370 F2 — wire format of the server→client REPLY to a save-version signal, sent
+ * over the same stateless channel. `version.saved` means a version was created or
+ * promoted; `version.skipped` is a TERMINAL "nothing was pinned" reply for the two
+ * reachable no-op cases (an effectively-empty page, or the page row is gone) so the
+ * client resolves at once instead of waiting out its ack timeout and misreporting a
+ * healthy server as unreachable. EXACTLY ONE of these is broadcast per handled
+ * save. The MCP client duplicates these literals (it cannot import server code) —
+ * keep the two in sync (see packages/mcp/src/lib/collaboration.ts).
+ */
+export const VERSION_SAVED_MESSAGE_TYPE = 'version.saved';
+export const VERSION_SKIPPED_MESSAGE_TYPE = 'version.skipped';
+
 /**
  * #251 — how long an intentional-clear signal stays "pending" before it is
  * ignored. The signal is set on the clearing keystroke but consumed by the
@@ -643,6 +656,10 @@ export class PersistenceExtension implements Extension {
     let result:
       | { historyId: string; kind: PageHistoryKind; alreadySaved: boolean }
       | undefined;
+    // #370 F2 — set when there is nothing to version (empty page / page gone), so
+    // the tail broadcasts a terminal `version.skipped` instead of staying silent
+    // and forcing the client to time out. Mutually exclusive with `result`.
+    let skipped: 'empty' | 'page-not-found' | undefined;
 
     // #370 F8-twin — the contributor set popped from Redis (destructive SPOP)
     // must be restored if the version row does not durably land. The inner
@@ -668,11 +685,20 @@ export class PersistenceExtension implements Extension {
           includeContent: true,
           trx,
         });
-        if (!page) return;
+        if (!page) {
+          // The page row is gone (deleted/never persisted). Nothing to version —
+          // record it so the tail sends a terminal skip reply (#370 F2).
+          skipped = 'page-not-found';
+          return;
+        }
         versionedPageId = page.id;
         // Never version an effectively-empty page (mirrors the processor's
-        // first-history guard); there is nothing intentional to pin.
-        if (isEmptyParagraphDoc(page.content as any)) return;
+        // first-history guard); there is nothing intentional to pin. Record the
+        // skip so the client gets a terminal reply rather than a timeout (#370 F2).
+        if (isEmptyParagraphDoc(page.content as any)) {
+          skipped = 'empty';
+          return;
+        }
 
         const lastHistory = await this.pageHistoryRepo.findPageLastHistory(
           page.id,
@@ -747,15 +773,26 @@ export class PersistenceExtension implements Extension {
     }
     this.idleBurstStart.delete(documentName);
 
+    // EXACTLY ONE terminal reply per handled save (#370 F2): a real save/promote
+    // broadcasts `version.saved`; the two no-op early returns (empty page / page
+    // gone) broadcast `version.skipped` so the client never waits out its ack
+    // timeout. A genuine failure threw above and rejected before reaching here.
     if (result) {
       document.broadcastStateless(
         JSON.stringify({
-          type: 'version.saved',
+          type: VERSION_SAVED_MESSAGE_TYPE,
           historyId: result.historyId,
           kind: result.kind,
           alreadySaved: result.alreadySaved,
         }),
       );
+    } else if (skipped) {
+      document.broadcastStateless(
+        JSON.stringify({
+          type: VERSION_SKIPPED_MESSAGE_TYPE,
+          reason: skipped,
+        }),
+      );
     }
   }
 

apps/server/src/collaboration/json-to-text-deterministic.spec.ts

@@ -0,0 +1,116 @@
+import { jsonToText } from './collaboration.util';
+
+// #502 READ contract: `jsonToText(json, { deterministic: true })` — the flat,
+// machine-diffable text rendering behind getPage `format:"text"`. Block-per-line
+// (`\n` separator), inline marks/anchors dropped, hardBreak -> `\n`, and non-text
+// nodes replaced by STABLE placeholders (`[image]`, `[table RxC]`). Output
+// stability across package versions IS the contract, so it is pinned by a
+// snapshot below. The DEFAULT (no options) path is the search-index serializer
+// and MUST be unchanged — asserted separately.
+
+const doc = (...content: any[]) => ({ type: 'doc', content });
+const para = (...content: any[]) => ({ type: 'paragraph', content });
+const text = (t: string, marks?: any[]) =>
+  marks ? { type: 'text', text: t, marks } : { type: 'text', text: t };
+
+describe('jsonToText — default (search index) behavior is unchanged', () => {
+  it('uses the `\\n\\n` block separator and drops non-text nodes to empty', () => {
+    const d = doc(para(text('alpha')), para(text('beta')));
+    expect(jsonToText(d)).toBe('alpha\n\nbeta');
+  });
+
+  it('an image contributes no text in the default (tsvector) mode', () => {
+    const d = doc(para(text('cap')), { type: 'image', attrs: { src: 's' } });
+    // No `[image]` placeholder leaks into the search index.
+    expect(jsonToText(d)).not.toContain('[image]');
+  });
+});
+
+describe('jsonToText — deterministic:true (getPage format:"text")', () => {
+  it('renders one line per block with `\\n` separators, marks dropped', () => {
+    const d = doc(
+      { type: 'heading', attrs: { level: 2 }, content: [text('Title')] },
+      para(
+        text('hello ', [{ type: 'bold' }]),
+        text('world', [{ type: 'italic' }]),
+      ),
+    );
+    expect(jsonToText(d, { deterministic: true })).toBe('Title\nhello world');
+  });
+
+  it('a hardBreak renders as a newline', () => {
+    const d = doc(para(text('a'), { type: 'hardBreak' }, text('b')));
+    expect(jsonToText(d, { deterministic: true })).toBe('a\nb');
+  });
+
+  it('an image node -> the stable `[image]` placeholder', () => {
+    const d = doc(para(text('before')), { type: 'image', attrs: { src: 's' } });
+    expect(jsonToText(d, { deterministic: true })).toBe('before\n[image]');
+  });
+
+  it('a table -> `[table RxC]` (rows x columns) and its cell text is NOT flattened in', () => {
+    const cell = (t: string) => ({
+      type: 'tableCell',
+      content: [para(text(t))],
+    });
+    const row = (...cells: any[]) => ({ type: 'tableRow', content: cells });
+    const table = {
+      type: 'table',
+      content: [
+        row(cell('CELLONE'), cell('CELLTWO'), cell('CELLTHREE')),
+        row(cell('CELLFOUR'), cell('CELLFIVE'), cell('CELLSIX')),
+      ],
+    };
+    const d = doc(para(text('grid:')), table);
+    const out = jsonToText(d, { deterministic: true });
+    expect(out).toBe('grid:\n[table 2x3]');
+    expect(out).not.toContain('CELL'); // cell text is not flattened into the read
+  });
+
+  it('SNAPSHOT: a mixed document renders to a stable, deterministic string', () => {
+    const cell = (t: string) => ({
+      type: 'tableCell',
+      content: [para(text(t))],
+    });
+    const d = doc(
+      { type: 'heading', attrs: { level: 1 }, content: [text('Config')] },
+      para(text('key = ', [{ type: 'bold' }]), text('value')),
+      {
+        type: 'bulletList',
+        content: [
+          { type: 'listItem', content: [para(text('one'))] },
+          { type: 'listItem', content: [para(text('two'))] },
+        ],
+      },
+      { type: 'image', attrs: { src: 's' } },
+      {
+        type: 'table',
+        content: [{ type: 'tableRow', content: [cell('x'), cell('y')] }],
+      },
+    );
+    expect(jsonToText(d, { deterministic: true })).toMatchInlineSnapshot(`
+     "Config
+     key = value
+
+
+     one
+
+     two
+     [image]
+     [table 1x2]"
+    `);
+  });
+
+  it('SCENARIO: a config written as a code block reads back byte-identical (diff empty)', () => {
+    const config =
+      'export TICKET_LIFETIME=$2592000\nservers:\n  - www.internal.host';
+    const d = doc({
+      type: 'codeBlock',
+      attrs: { language: 'yaml' },
+      content: [text(config)],
+    });
+    // The code block is one block; its text (dollars, bare domain, newlines) is
+    // preserved verbatim, so a read-as-text of a stored config diffs empty.
+    expect(jsonToText(d, { deterministic: true })).toBe(config);
+  });
+});

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

@@ -35,6 +35,7 @@ import {
 import { PaginationOptions } from '@docmost/db/pagination/pagination-options';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
 import { UserThrottlerGuard } from '../../integrations/throttle/user-throttler.guard';
 import { AI_CHAT_THROTTLER } from '../../integrations/throttle/throttler-names';
@@ -43,6 +44,8 @@ import {
   AiChatRunHooks,
   AiChatService,
   AiChatStreamBody,
+  rowHasInlineParts,
+  hydrateAssistantParts,
 } from './ai-chat.service';
 import { AiChatRunService } from './ai-chat-run.service';
 import { AiTranscriptionService } from './ai-transcription.service';
@@ -129,8 +132,39 @@ export class AiChatController {
     // production. Only touched on the resumable-stream (flag-on) path.
     private readonly streamRegistry?: AiChatStreamRegistryService,
     private readonly environment?: EnvironmentService,
+    // #492: reconstruct a #492 mid-run record's parts from the steps table before
+    // returning rows to the client / export. OPTIONAL so positional controller
+    // specs compile unchanged; when absent, hydration is skipped (old-era rows
+    // already carry inline parts, so nothing to reconstruct).
+    private readonly aiChatRunStepRepo?: AiChatRunStepRepo,
   ) {}
 
+  /**
+   * Reconstruct parts for any assistant rows that don't carry them INLINE — a
+   * #492 mid-run record whose per-step parts live in `ai_chat_run_steps` (the
+   * append-persist backend). Every FINISHED row (old-era + #492) and every old-era
+   * streaming snapshot already has inline `metadata.parts`, so the common path
+   * fetches NOTHING and returns the rows untouched; only an actively-streaming
+   * new-style row triggers the batch step fetch. Consumers (seed/poll/export) read
+   * `metadata.parts` off the returned rows exactly as before — the era switch is
+   * invisible to them (reconstructRunParts contract).
+   */
+  private async withReconstructedParts(
+    rows: AiChatMessage[],
+    workspaceId: string,
+  ): Promise<AiChatMessage[]> {
+    if (!this.aiChatRunStepRepo) return rows;
+    const needy = rows.filter(
+      (r) => r.role === 'assistant' && !rowHasInlineParts(r),
+    );
+    if (needy.length === 0) return rows;
+    const stepsByMessage = await this.aiChatRunStepRepo.findByMessageIds(
+      needy.map((r) => r.id),
+      workspaceId,
+    );
+    return hydrateAssistantParts(rows, stepsByMessage);
+  }
+
   /** List the requesting user's chats in this workspace (paginated). */
   @HttpCode(HttpStatus.OK)
   @Post('chats')
@@ -184,11 +218,17 @@ export class AiChatController {
     @AuthWorkspace() workspace: Workspace,
   ) {
     await this.assertOwnedChat(dto.chatId, user, workspace);
-    return this.aiChatMessageRepo.findByChat(
+    const page = await this.aiChatMessageRepo.findByChat(
       dto.chatId,
       workspace.id,
       pagination,
     );
+    // #492: reconstruct parts for any active new-style row so the client seed sees
+    // `metadata.parts` unchanged (a no-op for the finished rows that fill a page).
+    return {
+      ...page,
+      items: await this.withReconstructedParts(page.items, workspace.id),
+    };
   }
 
   /**
@@ -225,7 +265,10 @@ export class AiChatController {
       workspace.id,
     );
     return {
-      rows,
+      // #492: the delta of an actively-streaming new-style row carries its parts
+      // reconstructed from the steps table, so the degraded poll shows persisted
+      // progress exactly as the pre-#492 full-row snapshot did.
+      rows: await this.withReconstructedParts(rows, workspace.id),
       cursor,
       run: run ? { id: run.id, status: run.status } : null,
     };
@@ -247,8 +290,10 @@ export class AiChatController {
     @AuthWorkspace() workspace: Workspace,
   ): Promise<{ markdown: string }> {
     const chat = await this.assertOwnedChat(dto.chatId, user, workspace);
-    const rows = await this.aiChatMessageRepo.findAllByChat(
-      dto.chatId,
+    const rows = await this.withReconstructedParts(
+      await this.aiChatMessageRepo.findAllByChat(dto.chatId, workspace.id),
+      // #492: an interrupted-but-still-active turn exports its persisted steps
+      // (reconstructed from the steps table) just like the pre-#492 full row did.
       workspace.id,
     );
     const markdown = buildChatMarkdown({
@@ -288,7 +333,13 @@ export class AiChatController {
           workspace.id,
         )
       : undefined;
-    return { run, message: message ?? null };
+    // #492: reconnect to an IN-FLIGHT run reconstructs the projection row's parts
+    // from the steps table (the row itself carries only the step marker mid-run);
+    // a finished run's row already has inline parts, so this is a no-op.
+    const [hydrated] = message
+      ? await this.withReconstructedParts([message], workspace.id)
+      : [undefined];
+    return { run, message: hydrated ?? null };
   }
 
   /**

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

@@ -370,10 +370,12 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
     );
   });
 
-  // #490 reactive branch: a provider CONTEXT-OVERFLOW 400 in onError is classified,
-  // records a distinguishable cause, and stamps metadata.replayOverflow so the NEXT
-  // turn's budgeter trims aggressively (the recovery that un-bricks the chat).
-  it('#490: a context-overflow 400 stamps replayOverflow on the finalized row', async () => {
+  // #490/#520 reactive branch: a provider CONTEXT-OVERFLOW 400 in onError is
+  // classified, records a distinguishable cause, and stamps the consecutive-overflow
+  // COUNTER (metadata.replayOverflowCount) so the NEXT turn's budgeter trims with
+  // escalating aggression (the recovery that un-bricks the chat). This is a fresh
+  // chat (empty history -> prior streak 0), so the first overflow stamps count 1.
+  it('#490/#520: a context-overflow 400 stamps replayOverflowCount=1 on the finalized row', async () => {
     jest
       .spyOn(Logger.prototype, 'error')
       .mockImplementation(() => undefined as never);
@@ -397,11 +399,14 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       metadata: Record<string, unknown>;
     };
     expect(patch.status).toBe('error');
-    expect(patch.metadata.replayOverflow).toBe(true);
+    // First overflow on a fresh chat -> k = prior(0) + 1 = 1.
+    expect(patch.metadata.replayOverflowCount).toBe(1);
+    // The legacy boolean is no longer written (the counter supersedes it).
+    expect('replayOverflow' in patch.metadata).toBe(false);
     expect(patch.metadata.error).toContain('контекстное окно');
   });
 
-  it('#490: a non-overflow error does NOT stamp replayOverflow', async () => {
+  it('#490/#520: a non-overflow error does NOT stamp the overflow counter', async () => {
     jest
       .spyOn(Logger.prototype, 'error')
       .mockImplementation(() => undefined as never);
@@ -412,6 +417,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       status: string;
       metadata: Record<string, unknown>;
     };
+    expect('replayOverflowCount' in patch.metadata).toBe(false);
     expect('replayOverflow' in patch.metadata).toBe(false);
   });
 });

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

@@ -30,13 +30,16 @@ import {
   STEP_LIMIT_NO_ANSWER_MARKER,
   OUTPUT_DEGENERATION_ERROR,
   lastAssistantContextTokens,
-  lastAssistantReplayOverflow,
+  lastAssistantReplayOverflowCount,
   seedActivatedTools,
 } from './ai-chat.service';
 import type { AiChatMessage, Workspace } from '@docmost/db/types/entity.types';
 import { buildSystemPrompt } from './ai-chat.prompt';
 import type { McpClientsService } from './external-mcp/mcp-clients.service';
-import { resolveEffectiveReplayThreshold } from './history-budget';
+import {
+  resolveEffectiveReplayThreshold,
+  REPLAY_MIN_FLOOR_TOKENS,
+} from './history-budget';
 
 /**
  * Unit tests for compactToolOutput: the pure helper that shrinks tool outputs
@@ -554,49 +557,129 @@ describe('seedActivatedTools', () => {
   });
 });
 
-describe('lastAssistantReplayOverflow', () => {
+describe('lastAssistantReplayOverflowCount', () => {
   const row = (
     role: string,
     metadata: Record<string, unknown> | null,
   ): AiChatMessage => ({ role, metadata }) as unknown as AiChatMessage;
 
-  it('is true only when the LAST assistant turn overflowed', () => {
+  it('reads the consecutive-overflow count from the LAST assistant turn', () => {
     expect(
-      lastAssistantReplayOverflow([
-        row('assistant', { replayOverflow: true }),
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: 3 }),
         row('user', null),
       ]),
-    ).toBe(true);
-    // A recovered (later, non-overflow) assistant turn clears it.
+    ).toBe(3);
+    // A recovered (later, non-overflow) assistant turn resets it to 0 — the read
+    // stops at the most recent assistant row, which carries no count.
     expect(
-      lastAssistantReplayOverflow([
-        row('assistant', { replayOverflow: true }),
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: 3 }),
         row('user', null),
         row('assistant', { contextTokens: 5 }),
       ]),
-    ).toBe(false);
-    expect(lastAssistantReplayOverflow([])).toBe(false);
+    ).toBe(0);
+    expect(lastAssistantReplayOverflowCount([])).toBe(0);
   });
 
-  // #490 reactive recovery: a prior turn stamped `replayOverflow` must make the
-  // NEXT turn's effective budget the AGGRESSIVE 0.5x cut — that harder trim is
-  // what un-bricks a chat that just 400'd on the context window. This exercises
-  // the exact wiring the service uses: read the stamp, then scale the threshold.
-  it('#490: a prior replayOverflow drives the next turn to the 0.5x aggressive budget', () => {
+  // BACK-COMPAT (#520): an in-flight row written by the pre-#520 boolean stamp
+  // (`replayOverflow: true`, no count) reads as k=1 — the old single 0.5× behavior —
+  // so a chat mid-recovery across the deploy does not regress.
+  it('#520 back-compat: a legacy boolean replayOverflow reads as k=1', () => {
+    expect(
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+      ]),
+    ).toBe(1);
+    // A legacy row with the flag absent/false is k=0.
+    expect(
+      lastAssistantReplayOverflowCount([row('assistant', { contextTokens: 5 })]),
+    ).toBe(0);
+  });
+
+  // A corrupt/negative persisted count never yields a negative k.
+  it('clamps a corrupt negative count to 0', () => {
+    expect(
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: -4 }),
+      ]),
+    ).toBe(0);
+  });
+
+  // #490/#520 reactive recovery: the prior consecutive-overflow count `k` drives
+  // the next turn's effective budget to an ESCALATING cut (0.5**k) — each further
+  // consecutive 400 tightens it, which is what un-bricks a chat that keeps
+  // overflowing. This exercises the exact wiring the service uses: read the count,
+  // then scale the threshold.
+  it('#490/#520: the prior count drives the next turn to the escalating aggressive budget', () => {
     const history = [
-      row('assistant', { replayOverflow: true }),
+      row('assistant', { replayOverflowCount: 1 }),
       row('user', null),
     ];
-    const priorOverflowed = lastAssistantReplayOverflow(history);
-    expect(priorOverflowed).toBe(true);
-    // Base budget 100k -> aggressive recovery halves it to 50k this turn.
-    expect(resolveEffectiveReplayThreshold(100_000, priorOverflowed)).toBe(50_000);
+    const k = lastAssistantReplayOverflowCount(history);
+    expect(k).toBe(1);
+    // Base budget 100k -> first-overflow recovery halves it to 50k this turn.
+    expect(resolveEffectiveReplayThreshold(100_000, k)).toBe(50_000);
+    // A second consecutive overflow (k=2) quarters it.
+    expect(resolveEffectiveReplayThreshold(100_000, 2)).toBe(25_000);
     // Odd base floors, not rounds.
-    expect(resolveEffectiveReplayThreshold(99_999, true)).toBe(49_999);
-    // No prior overflow -> the base budget is used verbatim (no aggressive cut).
-    expect(resolveEffectiveReplayThreshold(100_000, false)).toBe(100_000);
+    expect(resolveEffectiveReplayThreshold(99_999, 1)).toBe(49_999);
+    // No prior overflow (k=0) -> the base budget is used verbatim (no cut).
+    expect(resolveEffectiveReplayThreshold(100_000, 0)).toBe(100_000);
     // An explicit off-switch (null) is never overridden, even on recovery.
-    expect(resolveEffectiveReplayThreshold(null, true)).toBeNull();
+    expect(resolveEffectiveReplayThreshold(null, 3)).toBeNull();
+  });
+
+  // #520 escalation table + convergence: the cut deepens each consecutive overflow
+  // and is CLAMPED at the floor so it converges (un-bricks even against a small
+  // real window), instead of the old fixed single 0.5× that stuck at 50k forever.
+  it('#520: escalates and converges to the floor, un-bricking a small real window', () => {
+    const base = 100_000;
+    expect(resolveEffectiveReplayThreshold(base, 0)).toBe(base);
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBe(50_000);
+    expect(resolveEffectiveReplayThreshold(base, 2)).toBe(25_000);
+    expect(resolveEffectiveReplayThreshold(base, 3)).toBe(12_500);
+
+    // Residual-brick regression (#520): with the flat-default base (100k) and a real
+    // model window of ~40k, the OLD fixed 0.5× stuck at 50k forever (> 40k -> 400s
+    // again, never recovers). The escalating cut drops BELOW 40k after enough
+    // consecutive overflows -> the history finally fits -> the chat un-bricks.
+    const realWindow = 40_000;
+    // k=1 (50k) still exceeds the window — the old behavior's terminal state.
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBeGreaterThan(realWindow);
+    // But escalation converges under the window within a couple more turns.
+    const converged = [2, 3, 4, 5].some(
+      (k) => (resolveEffectiveReplayThreshold(base, k) as number) < realWindow,
+    );
+    expect(converged).toBe(true);
+
+    // Convergence is bounded BELOW by the floor: a large k never trims below it.
+    for (const k of [4, 8, 20, 100]) {
+      expect(resolveEffectiveReplayThreshold(base, k)).toBe(REPLAY_MIN_FLOOR_TOKENS);
+      expect(
+        resolveEffectiveReplayThreshold(base, k) as number,
+      ).toBeGreaterThanOrEqual(REPLAY_MIN_FLOOR_TOKENS);
+    }
+  });
+
+  // The floor never RAISES a legitimately small configured budget above itself —
+  // that would re-overflow the very window it was configured for. Under #520 Option B
+  // recovery MAY cut it BELOW itself (down to floor(0.5×base) = the old 0.5× cut).
+  it('#520: never inflates a small configured budget above itself (may cut below)', () => {
+    const small = 5_000; // below the floor
+    const floor = Math.min(REPLAY_MIN_FLOOR_TOKENS, Math.floor(small * 0.5)); // 2500
+    expect(resolveEffectiveReplayThreshold(small, 0)).toBe(small);
+    // Even under escalation the effective threshold never exceeds the base, and never
+    // drops below the absolute floor.
+    for (const k of [1, 2, 3, 10]) {
+      const t = resolveEffectiveReplayThreshold(small, k) as number;
+      expect(t).toBeLessThanOrEqual(small);
+      expect(t).toBeGreaterThanOrEqual(floor);
+    }
+    // Option B: on overflow it DOES cut below the configured budget (old floor==budget
+    // behavior would keep this at `small`).
+    expect(resolveEffectiveReplayThreshold(small, 1)).toBe(floor);
   });
 });
 
@@ -930,21 +1013,50 @@ describe('flushAssistant', () => {
     expect(flushed.metadata.error).toBe('boom');
   });
 
-  // #490 observability: the replay budgeter's decision is stamped on the turn.
-  it('records replayTrimmedToTokens + replayOverflow when provided', () => {
+  // #490/#520 observability: the replay budgeter's decision is stamped on the turn,
+  // now including the consecutive-overflow COUNTER (#520) the next turn escalates on.
+  it('records replayTrimmedToTokens + replayOverflowCount when provided', () => {
     const f = flushAssistant([], '', 'error', {
       error: 'ctx',
       replayTrimmedToTokens: 42_000,
-      replayOverflow: true,
+      replayOverflowCount: 2,
     });
     expect(f.metadata.replayTrimmedToTokens).toBe(42_000);
-    expect(f.metadata.replayOverflow).toBe(true);
+    expect(f.metadata.replayOverflowCount).toBe(2);
   });
 
   it('omits the replay metadata when not provided', () => {
     const f = flushAssistant([], '', 'completed', { finishReason: 'stop' });
     expect('replayTrimmedToTokens' in f.metadata).toBe(false);
-    expect('replayOverflow' in f.metadata).toBe(false);
+    expect('replayOverflowCount' in f.metadata).toBe(false);
+    expect('replayBelowConfiguredBudget' in f.metadata).toBe(false);
+  });
+
+  // #520 Option B observability: the turn is marked when recovery replayed BELOW the
+  // admin-configured budget, so the UI/telemetry can surface "config too large".
+  it('stamps replayBelowConfiguredBudget when recovery cut below the configured budget', () => {
+    const f = flushAssistant([], '', 'error', {
+      error: 'ctx',
+      replayOverflowCount: 2,
+      replayBelowConfiguredBudget: true,
+    });
+    // MUTATION SENTINEL: dropping the metadata stamp in flushAssistant reddens this.
+    expect(f.metadata.replayBelowConfiguredBudget).toBe(true);
+  });
+
+  it('omits replayBelowConfiguredBudget on a normal in-budget replay', () => {
+    // false/omitted must NOT leave the flag on the turn (only the below-budget case).
+    const off = flushAssistant([], '', 'completed', {
+      replayBelowConfiguredBudget: false,
+    });
+    expect('replayBelowConfiguredBudget' in off.metadata).toBe(false);
+  });
+
+  // A clean finalize (no overflow -> count 0/omitted) leaves NO counter, which the
+  // next turn reads as k=0 — the reset that ends a recovery streak.
+  it('omits replayOverflowCount for a zero/absent count (reset semantics)', () => {
+    const zero = flushAssistant([], '', 'completed', { replayOverflowCount: 0 });
+    expect('replayOverflowCount' in zero.metadata).toBe(false);
   });
 
   // #274 observability: the page-change diff the agent saw this turn is persisted

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

@@ -22,6 +22,7 @@ import { AiSettingsService } from '../../integrations/ai/ai-settings.service';
 import { describeProviderError } from '../../integrations/ai/ai-error.util';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { AiChatPageSnapshotRepo } from '@docmost/db/repos/ai-chat/ai-chat-page-snapshot.repo';
 import { AiAgentRoleRepo } from '@docmost/db/repos/ai-agent-roles/ai-agent-roles.repo';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
@@ -140,9 +141,10 @@ const OUTPUT_DEGENERATION_ERROR =
 
 // Prefix recorded on the assistant row when the provider rejected the turn for
 // CONTEXT OVERFLOW (#490): the replayed history exceeded the model's window. The
-// row is ALSO stamped `metadata.replayOverflow` so the NEXT turn's budgeter trims
-// aggressively (the reactive recovery — the overflowing turn had no usage signal
-// to trigger preventive trimming, so the classified 400 is what un-bricks it).
+// row is ALSO stamped `metadata.replayOverflowCount` (the consecutive-overflow
+// counter, #520) so the NEXT turn's budgeter trims with escalating aggression (the
+// reactive recovery — the overflowing turn had no usage signal to trigger
+// preventive trimming, so the classified 400 is what un-bricks it).
 export const CONTEXT_OVERFLOW_ERROR_PREFIX =
   'Диалог превысил контекстное окно модели; история будет агрессивно ' +
   'сокращена на следующем ходу.';
@@ -518,6 +520,12 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
     // constructions compile unchanged; Nest always injects the real singleton, so
     // reconcile sees the SAME in-memory active/zombie maps the runner mutates.
     private readonly aiChatRunService?: AiChatRunService,
+    // #492 append-persist: per-step INSERT into the lightweight steps table (the
+    // O(Σ steps) replacement for the O(n²) full-row `metadata.parts` rewrite).
+    // OPTIONAL so existing positional constructions (int-specs) compile unchanged;
+    // Nest injects the real singleton. When ABSENT the per-step path falls back to
+    // the pre-#492 full-row flush (no regression, only no WAL win).
+    private readonly aiChatRunStepRepo?: AiChatRunStepRepo,
   ) {}
 
   // #487: periodic reconcile timer (single-process phase 1). Started in
@@ -1114,8 +1122,34 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
         chatId,
         workspace.id,
       );
+      // #492: HYDRATE needy assistant rows from the steps table BEFORE the replay
+      // map. A #492 mid-run assistant row carries only a step marker
+      // (metadata.parts:[]); its real per-step parts live in `ai_chat_run_steps`.
+      // The graceful terminal callbacks (onFinish/onError/onAbort -> flushAssistant)
+      // assemble the full inline parts, so a normally-ended turn already has them.
+      // But a HARD crash mid-run (SIGKILL/OOM) fires NO terminal callback, so the
+      // row stays parts:[]; without this, rowToUiMessage falls back to an empty
+      // text part and the partial tool-calls/results/text — durable in the steps
+      // table — would DROP OUT of the model's replay context (regressing #183
+      // step-granular durability for the model consumer). Mirrors the controller's
+      // withReconstructedParts EXACTLY (same needy predicate + hydration helper).
+      // Guarded on the optional repo: absent (positional test builds) degrades to
+      // the current behavior rather than crashing.
+      let replayHistory = oldHistory;
+      if (this.aiChatRunStepRepo) {
+        const needy = oldHistory.filter(
+          (r) => r.role === 'assistant' && !rowHasInlineParts(r),
+        );
+        if (needy.length > 0) {
+          const stepsByMessage = await this.aiChatRunStepRepo.findByMessageIds(
+            needy.map((r) => r.id),
+            workspace.id,
+          );
+          replayHistory = hydrateAssistantParts(oldHistory, stepsByMessage);
+        }
+      }
       const uiMessages: Array<Omit<UIMessage, 'id'> & { id: string }> = [
-        ...oldHistory.map(rowToUiMessage),
+        ...replayHistory.map(rowToUiMessage),
         {
           id: 'pending-user',
           role: 'user',
@@ -1154,7 +1188,9 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       // hint — confirm it against the persisted history (the preceding assistant
       // turn must really be aborted/streaming) so a spoofed flag cannot inject the
       // interrupt note onto an ordinary turn. The partial output the model needs is
-      // already in `messages` (the aborted assistant row replays via findRecent).
+      // already in `messages`: a #492 mid-run row's per-step parts live only in the
+      // `ai_chat_run_steps` table and were hydrated into the replay history above,
+      // so the aborted assistant turn replays WITH its partial parts intact.
       // Append the new user turn (shape-only) so index -2 is the prior assistant.
       const interrupted = isInterruptResume(
         [...oldHistory, { role: 'user', status: null, metadata: null }],
@@ -1193,27 +1229,52 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       }
       // Last turn's provider-reported context size (authoritative when present).
       const priorContextTokens = lastAssistantContextTokens(oldHistory);
-      // Reactive recovery (#490): if the LAST turn was rejected for context
-      // overflow (stamped by onError), trim AGGRESSIVELY this turn — the
-      // overflowing turn produced no usage signal, so a normal-threshold trim may
-      // not shrink enough to fit. This is what un-bricks a chat that just 400'd.
-      const priorOverflowed = lastAssistantReplayOverflow(oldHistory);
+      // Reactive recovery (#490/#520): `k` = how many CONSECUTIVE preceding turns
+      // were rejected for context overflow (stamped by onError). Each consecutive
+      // overflow trims MORE aggressively (resolveEffectiveReplayThreshold scales the
+      // budget by 0.5**k, clamped at the floor) so recovery ESCALATES until the
+      // history fits — the overflowing turn produced no usage signal, so a single
+      // fixed cut may not shrink enough when the real model window is small. This is
+      // what un-bricks a chat that keeps 400'ing on the context window.
+      const priorOverflowCount = lastAssistantReplayOverflowCount(oldHistory);
       const effectiveThreshold = resolveEffectiveReplayThreshold(
         replayBudget.thresholdTokens,
-        priorOverflowed,
+        priorOverflowCount,
       );
-      if (priorOverflowed) {
-        this.logger.warn(
-          `AI chat (chat ${chatId}): previous turn hit context overflow; ` +
-            `applying aggressive replay budget (${effectiveThreshold} tokens).`,
-        );
+      // #520 Option B: recovery MAY cut the replay budget BELOW the admin-configured
+      // window when the provider keeps proving non-fit (the anti-brick invariant of
+      // the reactive branch beats a config that a 400 has disproved). Surface it so
+      // the admin sees their window is factually too large, and mark the turn so the
+      // UI/telemetry can show it. Only true when actually below (not on in-budget
+      // escalation edge cases); k>0 implies below for any budget >= 2 tokens.
+      const replayBelowConfiguredBudget =
+        replayBudget.thresholdTokens != null &&
+        effectiveThreshold != null &&
+        effectiveThreshold < replayBudget.thresholdTokens;
+      if (priorOverflowCount > 0) {
+        if (replayBelowConfiguredBudget) {
+          this.logger.warn(
+            `AI chat (chat ${chatId}): configured replay budget ` +
+              `(${replayBudget.thresholdTokens} tokens) does not fit the model — ` +
+              `replaying BELOW it at ${effectiveThreshold} tokens after ` +
+              `${priorOverflowCount} consecutive context overflow(s) ` +
+              `(escalation level ${priorOverflowCount}). Lower chatContextWindow ` +
+              `to match the model's real context window.`,
+          );
+        } else {
+          this.logger.warn(
+            `AI chat (chat ${chatId}): ${priorOverflowCount} consecutive context ` +
+              `overflow(s); applying escalated aggressive replay budget ` +
+              `(${effectiveThreshold} tokens).`,
+          );
+        }
       }
       const preTrim = trimHistoryForReplay(
         messages,
         effectiveThreshold,
         // A prior OVERFLOW means the provider count is stale/absent — force the
         // char-estimate path by ignoring priorContextTokens on recovery.
-        priorOverflowed ? undefined : priorContextTokens,
+        priorOverflowCount > 0 ? undefined : priorContextTokens,
       );
       messages = preTrim.messages;
       // Observability (#490): record the budgeter's decision on the turn so the UI
@@ -1559,17 +1620,57 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
         // connection when finalize runs, so the SQL `WHERE status='streaming'`
         // (not this flag) is what prevents it clobbering the terminal row.
         if (finalized) return null;
-        // Build the flush ONCE so the returned count is EXACTLY the persisted
-        // `stepsPersisted` (both derive from capturedSteps.length at this instant).
-        const flushed = flushAssistant(capturedSteps, '', 'streaming', {
-          pageChanged,
-          partsCache,
-        });
-        const stepsPersisted = flushed.metadata.stepsPersisted as number;
+        // The count derives from capturedSteps.length at THIS instant, so the
+        // returned value is EXACTLY the persisted `stepsPersisted` the ring rotates
+        // on (whether we take the append-persist path or the legacy fallback).
+        const stepsPersisted = capturedSteps.length;
         try {
-          await this.aiChatMessageRepo.update(assistantId, workspace.id, flushed, {
-            onlyIfStreaming: true,
-          });
+          if (this.aiChatRunStepRepo) {
+            // #492 APPEND-PERSIST: write only THIS finished step's parts to the
+            // steps table (O(step) WAL), then bump the row's CHEAP step marker —
+            // NO growing `metadata.parts` blob (that O(n²) full-row rewrite is
+            // exactly what this removes). The full `metadata.parts` is assembled
+            // once at finalize; a mid-run resume seed is reconstructed from the
+            // step rows (reconstructRunParts). The INSERT is idempotent
+            // (ON CONFLICT DO NOTHING), so a re-fired step never doubles the parts.
+            const index = stepsPersisted - 1;
+            if (index >= 0) {
+              const stepParts = assistantParts(
+                [capturedSteps[index]],
+                '',
+                partsCache,
+              );
+              await this.aiChatRunStepRepo.insertStep(
+                assistantId,
+                workspace.id,
+                index,
+                stepParts,
+              );
+            }
+            // Marker UPDATE: advance stepsPersisted + keep the toolTrace era marker
+            // (bumps updatedAt so the delta poll observes the step, and carries the
+            // frontier a resuming client attaches from). Scoped onlyIfStreaming so a
+            // late marker never clobbers the terminal finalize.
+            await this.aiChatMessageRepo.update(
+              assistantId,
+              workspace.id,
+              { metadata: stepMarkerMetadata(stepsPersisted) },
+              { onlyIfStreaming: true },
+            );
+          } else {
+            // Legacy fallback (no steps table wired — positional test builds): the
+            // pre-#492 full-row flush, so parts still land inline on the row.
+            const flushed = flushAssistant(capturedSteps, '', 'streaming', {
+              pageChanged,
+              partsCache,
+            });
+            await this.aiChatMessageRepo.update(
+              assistantId,
+              workspace.id,
+              flushed,
+              { onlyIfStreaming: true },
+            );
+          }
           return stepsPersisted;
         } catch (err) {
           this.logger.warn(
@@ -1828,6 +1929,10 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
                 pageChanged,
                 partsCache,
                 replayTrimmedToTokens,
+                // #520 Option B: mark the turn when recovery replayed BELOW the
+                // configured budget (only when it actually did).
+                replayBelowConfiguredBudget:
+                  replayBelowConfiguredBudget || undefined,
               }),
             );
             // #184/#487: the RUN is finalized ALWAYS (never gated on the message).
@@ -1906,7 +2011,16 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
                 pageChanged,
                 partsCache,
                 replayTrimmedToTokens,
-                replayOverflow: overflow || undefined,
+                // #520: escalate the consecutive-overflow counter so the NEXT turn
+                // trims MORE aggressively (0.5**k). k grows by 1 each consecutive
+                // overflow; a clean finalize omits the field, resetting it to 0.
+                replayOverflowCount: overflow
+                  ? priorOverflowCount + 1
+                  : undefined,
+                // #520 Option B: mark the turn when THIS replay was already below the
+                // configured budget (only when it actually was).
+                replayBelowConfiguredBudget:
+                  replayBelowConfiguredBudget || undefined,
               }),
             );
             // #184: settle the RUN as failed, carrying the provider/transport cause.
@@ -2338,22 +2452,39 @@ export function seedActivatedTools(
 }
 
 /**
- * Whether the most recent assistant turn was rejected for CONTEXT OVERFLOW
- * (#490): its row carries `metadata.replayOverflow` (stamped by the stream's
- * onError). The next turn's budgeter reads this to trim aggressively — the
- * reactive recovery. Only the LAST assistant turn matters (an older overflow was
- * already recovered), so we stop at the first assistant row scanning backwards.
+ * How many CONSECUTIVE recent turns were rejected for CONTEXT OVERFLOW (#490/#520):
+ * `k`, read from the most recent assistant row's `metadata.replayOverflowCount`
+ * (stamped by the stream's onError, incremented each consecutive overflow and reset
+ * to 0 on any clean finalize). The next turn's budgeter feeds this to
+ * {@link resolveEffectiveReplayThreshold} to trim with ESCALATING aggression — the
+ * reactive recovery. Only the LAST assistant turn matters (its count already carries
+ * the consecutive streak; an older overflow followed by a clean turn was recovered),
+ * so we stop at the first assistant row scanning backwards.
+ *
+ * BACK-COMPAT: a row written by the pre-#520 boolean stamp (`replayOverflow: true`,
+ * no count) is read as k=1 — the old single 0.5× behavior — so in-flight chats do
+ * not regress across the deploy.
  */
-export function lastAssistantReplayOverflow(
+export function lastAssistantReplayOverflowCount(
   history: ReadonlyArray<AiChatMessage>,
-): boolean {
+): number {
   for (let i = history.length - 1; i >= 0; i--) {
     const row = history[i];
     if (row.role !== 'assistant') continue;
-    const meta = (row.metadata ?? {}) as { replayOverflow?: unknown };
-    return meta.replayOverflow === true;
+    const meta = (row.metadata ?? {}) as {
+      replayOverflowCount?: unknown;
+      replayOverflow?: unknown;
+    };
+    if (typeof meta.replayOverflowCount === 'number') {
+      // Guard against a corrupt/negative persisted value.
+      return meta.replayOverflowCount > 0
+        ? Math.floor(meta.replayOverflowCount)
+        : 0;
+    }
+    // Back-compat: legacy boolean stamp -> one overflow (0.5× cut).
+    return meta.replayOverflow === true ? 1 : 0;
   }
-  return false;
+  return 0;
 }
 
 /** The last message with role 'user' from a useChat payload, if any. */
@@ -2749,6 +2880,122 @@ export function rowToUiMessage(row: AiChatMessage): Omit<UIMessage, 'id'> & {
   return { id: row.id, role, parts: parts as UIMessage['parts'] };
 }
 
+/**
+ * Cheap step-marker metadata for the #492 per-step UPDATE. Advances
+ * `stepsPersisted` (the resume attach frontier) and keeps the `toolTraceVersion`
+ * era marker, WITHOUT the growing `parts` blob (those live in the steps table
+ * now; the full `metadata.parts` is assembled once at finalize by flushAssistant).
+ * `parts: []` is kept for shape stability — it reads as an empty inline-parts row,
+ * which is exactly the discriminator that routes reconstruction to the steps table.
+ */
+export function stepMarkerMetadata(
+  stepsPersisted: number,
+): Record<string, unknown> {
+  return { parts: [], toolTraceVersion: 2, stepsPersisted };
+}
+
+/**
+ * Whether an assistant row already carries its full UI parts INLINE on the row
+ * (`metadata.parts`). TRUE for every FINISHED row — old-era rows AND #492 rows,
+ * whose full parts are assembled once at finalize — and for old-era streaming
+ * snapshots (the pre-#492 per-step full-row flush). FALSE for a #492 MID-RUN
+ * record, whose per-step parts live in the `ai_chat_run_steps` table. This is the
+ * era discriminator the reconstruct seam branches on — no schema flag needed.
+ */
+export function rowHasInlineParts(row: { metadata?: unknown }): boolean {
+  const meta = (row.metadata ?? {}) as { parts?: unknown };
+  return Array.isArray(meta.parts) && meta.parts.length > 0;
+}
+
+/**
+ * Concatenate persisted per-step parts (in `stepIndex` order) into the turn's UI
+ * parts (#492). Reproduces EXACTLY what flushAssistant → assistantParts would have
+ * written to `metadata.parts` for those finished steps, since each step row stored
+ * `assistantParts([step])` at persist time.
+ */
+export function assembleStepParts(
+  stepRows: ReadonlyArray<{ stepIndex: number; parts: unknown }>,
+): UIMessage['parts'] {
+  const parts: Array<Record<string, unknown>> = [];
+  for (const step of [...stepRows].sort((a, b) => a.stepIndex - b.stepIndex)) {
+    if (Array.isArray(step.parts)) {
+      parts.push(...(step.parts as Array<Record<string, unknown>>));
+    }
+  }
+  return parts as UIMessage['parts'];
+}
+
+/**
+ * reconstructRunParts (#492) — the single backend-switch seam. Given an assistant
+ * ROW and its persisted step rows, return the turn's UI `parts` + the persisted
+ * step count, reading from the ROW when it already carries inline parts (old-era
+ * records AND every finished record) and from the STEPS TABLE otherwise (a #492
+ * mid-run record). The higher-level consumers (attach seed, delta poll, export)
+ * route their row→parts through this / {@link hydrateAssistantParts}, so old and
+ * new records reconstruct identically WITHOUT the consumers branching on the era.
+ */
+export function reconstructRunParts(
+  row: { metadata?: unknown; content?: string | null },
+  stepRows: ReadonlyArray<{ stepIndex: number; parts: unknown }>,
+): { parts: UIMessage['parts']; stepsPersisted: number } {
+  if (rowHasInlineParts(row)) {
+    const meta = row.metadata as {
+      parts: UIMessage['parts'];
+      stepsPersisted?: number;
+    };
+    return {
+      parts: meta.parts,
+      stepsPersisted:
+        typeof meta.stepsPersisted === 'number'
+          ? meta.stepsPersisted
+          : stepRows.length,
+    };
+  }
+  if (stepRows.length > 0) {
+    return {
+      parts: assembleStepParts(stepRows),
+      stepsPersisted: stepRows.length,
+    };
+  }
+  // No inline parts and no step rows: an old-era seed / empty streaming row. Fall
+  // back to a single text part from `content` (mirrors rowToUiMessage).
+  return {
+    parts: textPart(row.content ?? '') as UIMessage['parts'],
+    stepsPersisted: 0,
+  };
+}
+
+/**
+ * Fill each assistant row's `metadata.parts` from its step rows when the row does
+ * not already carry them inline (a #492 mid-run record), so a consumer that reads
+ * `metadata.parts` off the RAW row (the client seed/poll, the Markdown export)
+ * sees the reconstructed parts with NO change to itself. Rows that already have
+ * inline parts (old-era + finished) and non-assistant rows pass through untouched.
+ * Pure: returns new row objects, never mutates the inputs.
+ */
+export function hydrateAssistantParts<
+  T extends { id: string; role?: string; metadata?: unknown },
+>(
+  rows: ReadonlyArray<T>,
+  stepsByMessage: Map<
+    string,
+    ReadonlyArray<{ stepIndex: number; parts: unknown }>
+  >,
+): T[] {
+  return rows.map((row) => {
+    if (row.role !== 'assistant' || rowHasInlineParts(row)) return row;
+    const steps = stepsByMessage.get(row.id);
+    if (!steps || steps.length === 0) return row;
+    return {
+      ...row,
+      metadata: {
+        ...((row.metadata ?? {}) as Record<string, unknown>),
+        parts: assembleStepParts(steps),
+      },
+    };
+  });
+}
+
 /**
  * The persisted-row patch shape produced by {@link flushAssistant}. It is the
  * SAME shape the assistant repo insert/update consume (content + toolCalls +
@@ -2893,9 +3140,16 @@ export function flushAssistant(
     // the (estimated) token size it trimmed to — the UI can show "replay truncated
     // at N tokens". Omitted when nothing was trimmed.
     replayTrimmedToTokens?: number;
-    // #490 reactive branch: set when the provider rejected this turn for context
-    // overflow. Stamped into metadata so the NEXT turn's budgeter trims aggressively.
-    replayOverflow?: boolean;
+    // #490/#520 reactive branch: the consecutive context-overflow count for THIS
+    // turn (prior streak + 1) when the provider rejected it for context overflow.
+    // Stamped into metadata so the NEXT turn's budgeter trims with escalating
+    // aggression (0.5**k). Omitted (undefined) on a clean turn, which resets k to 0.
+    replayOverflowCount?: number;
+    // #520 Option B observability: true when recovery replayed this turn's history
+    // BELOW the admin-configured budget (the configured window did not fit and the
+    // anti-brick invariant cut past it). Omitted on a normal in-budget replay so the
+    // flag marks ONLY the "config is factually too large" case.
+    replayBelowConfiguredBudget?: boolean;
   },
 ): AssistantFlush {
   const finished = capturedSteps ?? [];
@@ -2948,7 +3202,10 @@ export function flushAssistant(
     metadata.maxContextTokens = extra.maxContextTokens;
   if (extra?.replayTrimmedToTokens)
     metadata.replayTrimmedToTokens = extra.replayTrimmedToTokens;
-  if (extra?.replayOverflow) metadata.replayOverflow = true;
+  if (extra?.replayOverflowCount && extra.replayOverflowCount > 0)
+    metadata.replayOverflowCount = extra.replayOverflowCount;
+  if (extra?.replayBelowConfiguredBudget)
+    metadata.replayBelowConfiguredBudget = true;
   if (extra?.error) metadata.error = extra.error;
   // Persist the page-change diff the agent saw this turn (#274 observability),
   // so history / the Markdown export can show what the user changed. Only when

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

@@ -1,6 +1,11 @@
 import { randomBytes } from 'crypto';
 import { Client } from 'pg';
-import { flushAssistant, serializeSteps } from './ai-chat.service';
+import {
+  flushAssistant,
+  serializeSteps,
+  lastAssistantReplayOverflowCount,
+} from './ai-chat.service';
+import type { AiChatMessage } from '@docmost/db/types/entity.types';
 
 /**
  * #490 write-volume regression — an OBSERVABLE-PROPERTY test on a LIVE Postgres,
@@ -207,3 +212,111 @@ describe('#490 write-volume on a live Postgres (pg_current_wal_lsn delta)', () =
     expect(v2).toBeLessThan(v1 * 0.75);
   }, 120_000);
 });
+
+/**
+ * #520 reactive-recovery COUNTER lifecycle on a LIVE Postgres — proves the
+ * consecutive-overflow count survives a real jsonb metadata round-trip (the persist
+ * path), not just an in-memory object. flushAssistant BUILDS the row metadata, we
+ * WRITE it to a jsonb column, READ it back, then reconstruct the assistant row and
+ * run lastAssistantReplayOverflowCount over it — exactly the read the next turn does.
+ *
+ * The lifecycle proven end-to-end through pg:
+ *   - consecutive overflows INCREMENT k (1 -> 2 -> 3);
+ *   - a CLEAN finalize omits the field, which the reader treats as a RESET to 0;
+ *   - a legacy boolean row (`replayOverflow: true`) reads back as k=1 (back-compat).
+ */
+describe('#520 overflow-counter lifecycle on a live Postgres (jsonb round-trip)', () => {
+  let client: Client | undefined;
+  let available = false;
+
+  beforeAll(async () => {
+    try {
+      client = new Client(CONN);
+      await client.connect();
+      await client.query('SELECT 1');
+      available = true;
+    } catch {
+      available = false;
+      client = undefined;
+    }
+  });
+
+  afterAll(async () => {
+    await client?.end().catch(() => undefined);
+  });
+
+  // Round-trip an arbitrary metadata object through a real jsonb column and read it
+  // back as the reconstructed assistant row the next turn would load.
+  async function roundTrip(
+    c: Client,
+    metadata: unknown,
+  ): Promise<AiChatMessage> {
+    await c.query('UPDATE _wal_counter SET metadata=$1 WHERE id=1', [
+      JSON.stringify(metadata),
+    ]);
+    const back = (await c.query('SELECT metadata FROM _wal_counter WHERE id=1'))
+      .rows[0].metadata as Record<string, unknown>;
+    return { role: 'assistant', metadata: back } as unknown as AiChatMessage;
+  }
+
+  it('increments across consecutive overflows, resets on a clean turn, and honors the legacy boolean', async () => {
+    if (!available || !client) {
+      console.warn('SKIP: gitmost-test-pg not reachable; skipping counter test.');
+      return;
+    }
+    const c = client;
+    await c.query('DROP TABLE IF EXISTS _wal_counter');
+    await c.query('CREATE TABLE _wal_counter(id int primary key, metadata jsonb)');
+    await c.query("INSERT INTO _wal_counter VALUES (1, '{}'::jsonb)");
+
+    // Turn 1 overflow: prior streak 0 -> stamp k=1 (as the service does: prior+1).
+    let prior = lastAssistantReplayOverflowCount([]); // fresh chat
+    expect(prior).toBe(0);
+    let row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(1);
+
+    // Turn 2 overflow: prior 1 -> stamp k=2.
+    row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(2);
+
+    // Turn 3 overflow: prior 2 -> stamp k=3.
+    row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(3);
+
+    // Turn 4 CLEAN finalize: no overflow -> the field is omitted -> reset to 0.
+    row = await roundTrip(
+      c,
+      flushAssistant([], 'all good', 'completed', { finishReason: 'stop' })
+        .metadata,
+    );
+    expect('replayOverflowCount' in (row.metadata as object)).toBe(false);
+    expect(lastAssistantReplayOverflowCount([row])).toBe(0);
+
+    // Back-compat: a row persisted by the pre-#520 boolean stamp reads back as k=1.
+    row = await roundTrip(c, { replayOverflow: true });
+    expect(lastAssistantReplayOverflowCount([row])).toBe(1);
+
+    await c.query('DROP TABLE IF EXISTS _wal_counter');
+  }, 60_000);
+});

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

@@ -32,6 +32,14 @@ describe('EmbeddingIndexerService.reindexWorkspace fail-fast', () => {
     const pageEmbeddingRepo = {};
     const aiService = {
       getEmbeddingModel: jest.fn().mockResolvedValue('some-model'),
+      // #530: reindexWorkspace's pre-check now resolves the provider (workspace
+      // or global). Resolve it so the batch control flow under test proceeds.
+      resolveEmbeddingProvider: jest.fn().mockResolvedValue({
+        model: 'some-model',
+        queryPrefix: '',
+        docPrefix: '',
+        fingerprint: 'fp-test',
+      }),
     };
     // Progress is a best-effort cosmetic store; mock its async methods so the
     // batch control flow can be tested without Redis.
@@ -108,6 +116,14 @@ describe('EmbeddingIndexerService.reindexWorkspace progress', () => {
     const pageEmbeddingRepo = {};
     const aiService = {
       getEmbeddingModel: jest.fn().mockResolvedValue('some-model'),
+      // #530: reindexWorkspace's pre-check now resolves the provider (workspace
+      // or global). Resolve it so the batch control flow under test proceeds.
+      resolveEmbeddingProvider: jest.fn().mockResolvedValue({
+        model: 'some-model',
+        queryPrefix: '',
+        docPrefix: '',
+        fingerprint: 'fp-test',
+      }),
     };
     const reindexProgress = {
       start: jest.fn().mockResolvedValue(undefined),
@@ -174,7 +190,7 @@ describe('EmbeddingIndexerService.reindexWorkspace progress', () => {
     const { service, aiService, reindexProgress } = makeService();
     // Embeddings not configured: reindexWorkspace returns early WITHOUT starting
     // a fresh record, but the finally must still clear the enqueue-time seed.
-    aiService.getEmbeddingModel = jest
+    aiService.resolveEmbeddingProvider = jest
       .fn()
       .mockRejectedValue(new AiEmbeddingNotConfiguredException());
 
@@ -187,3 +203,87 @@ describe('EmbeddingIndexerService.reindexWorkspace progress', () => {
     expect(reindexProgress.clear).toHaveBeenCalledWith(WORKSPACE_ID);
   });
 });
+
+/**
+ * #530 PR-1: reindexPage must (a) prepend the provider's DOC prefix to each chunk
+ * BEFORE embedding (so stored vectors live in the same prefixed space as a
+ * prefixed query), and (b) stamp the ACTIVE fingerprint on every inserted row (so
+ * search only fuses same-generation vectors). Uses lightweight mocks; the tx is
+ * stubbed to run its callback inline.
+ */
+describe('EmbeddingIndexerService.reindexPage doc-prefix + fingerprint (#530)', () => {
+  const WORKSPACE_ID = 'ws-1';
+  const SPACE_ID = 'space-1';
+  const PAGE_ID = 'page-1';
+
+  function makeService(docPrefix: string) {
+    const pageRepo = {
+      findById: jest.fn().mockResolvedValue({
+        id: PAGE_ID,
+        workspaceId: WORKSPACE_ID,
+        spaceId: SPACE_ID,
+        title: 'Заголовок',
+        // No ProseMirror content -> the plain-text fallback path (single chunk).
+        content: null,
+        textContent: 'простой текст страницы',
+        deletedAt: null,
+      }),
+    };
+    const insertChunks = jest.fn().mockResolvedValue(undefined);
+    const pageEmbeddingRepo = {
+      deleteByPage: jest.fn().mockResolvedValue(undefined),
+      insertChunks,
+    };
+    const embedWithModel = jest.fn().mockResolvedValue([[0.1, 0.2, 0.3]]);
+    const aiService = {
+      resolveEmbeddingProvider: jest.fn().mockResolvedValue({
+        model: { modelId: 'e5-small' },
+        queryPrefix: 'query: ',
+        docPrefix,
+        fingerprint: 'fp-gen-1',
+      }),
+      embedWithModel,
+    };
+    const reindexProgress = {};
+    // Stub the tx so executeTx runs its callback inline against a fake trx.
+    const db = {
+      transaction: () => ({ execute: (cb: any) => cb({}) }),
+    };
+    const service = new EmbeddingIndexerService(
+      pageRepo as unknown as PageRepo,
+      pageEmbeddingRepo as unknown as PageEmbeddingRepo,
+      aiService as unknown as AiService,
+      reindexProgress as unknown as EmbeddingReindexProgressService,
+      db as unknown as KyselyDB,
+    );
+    return { service, embedWithModel, insertChunks };
+  }
+
+  it('prepends the doc prefix to each chunk and stamps the fingerprint on rows', async () => {
+    const { service, embedWithModel, insertChunks } = makeService('passage: ');
+    await service.reindexPage(PAGE_ID);
+
+    // Embedded values are DOC-prefixed; the model is the resolved provider model.
+    expect(embedWithModel).toHaveBeenCalledTimes(1);
+    const [modelArg, wsArg, valuesArg] = embedWithModel.mock.calls[0];
+    expect(modelArg).toEqual({ modelId: 'e5-small' });
+    expect(wsArg).toBe(WORKSPACE_ID);
+    expect(valuesArg).toEqual(['passage: простой текст страницы']);
+
+    // Inserted rows carry the active fingerprint and the ORIGINAL (un-prefixed)
+    // content (the prefix is an embedding-space artifact, not stored text).
+    expect(insertChunks).toHaveBeenCalledTimes(1);
+    const rows = insertChunks.mock.calls[0][0];
+    expect(rows).toHaveLength(1);
+    expect(rows[0].fingerprint).toBe('fp-gen-1');
+    expect(rows[0].content).toBe('простой текст страницы');
+    expect(rows[0].modelName).toBe('e5-small');
+  });
+
+  it('does not prefix when the provider has an empty doc prefix', async () => {
+    const { service, embedWithModel } = makeService('');
+    await service.reindexPage(PAGE_ID);
+    const [, , valuesArg] = embedWithModel.mock.calls[0];
+    expect(valuesArg).toEqual(['простой текст страницы']);
+  });
+});

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

@@ -108,15 +108,24 @@ export class EmbeddingIndexerService {
       return;
     }
 
-    // Resolve embeddings config WITHOUT crashing the queue when unconfigured.
+    // Resolve the embeddings provider WITHOUT crashing the queue when
+    // unconfigured. #530: resolveEmbeddingProvider prefers the workspace provider
+    // and falls back to the GLOBAL env provider (TEI sidecar), and yields the
+    // doc-prefix + the active fingerprint stored per row so search filters by the
+    // active generation.
     let modelName = 'unknown';
+    let provider: Awaited<
+      ReturnType<AiService['resolveEmbeddingProvider']>
+    >;
     try {
-      const model = await this.aiService.getEmbeddingModel(workspaceId);
+      provider = await this.aiService.resolveEmbeddingProvider(workspaceId);
       // Record the model id per row so a future migration can detect + re-index
       // rows produced by a different model (see the migration header). The SDK
       // type is `string | EmbeddingModel{V2,V3}`; model objects carry `modelId`.
       modelName =
-        typeof model === 'string' ? model : (model.modelId ?? 'unknown');
+        typeof provider.model === 'string'
+          ? provider.model
+          : (provider.model.modelId ?? 'unknown');
     } catch (err) {
       if (err instanceof AiEmbeddingNotConfiguredException) {
         // No embeddings provider for this workspace: NO-OP (§6.7). The page can
@@ -145,8 +154,18 @@ export class EmbeddingIndexerService {
       return;
     }
 
-    // Embed all chunks in one batch.
-    const vectors = await this.aiService.embedTexts(workspaceId, chunks);
+    // #530: prepend the provider's DOC prefix to each chunk (e5-style
+    // "passage: "; empty for a non-e5 provider) so stored vectors live in the
+    // same prefixed space as a prefixed query, then embed with the RESOLVED
+    // provider model (which may be the global TEI sidecar, not a workspace one).
+    const prefixedChunks = provider.docPrefix
+      ? chunks.map((c) => provider.docPrefix + c)
+      : chunks;
+    const vectors = await this.aiService.embedWithModel(
+      provider.model,
+      workspaceId,
+      prefixedChunks,
+    );
 
     // The column is dimension-agnostic, so ANY model dimension is stored as-is.
     // Defensive sanity check only: all chunks of ONE page come from the SAME
@@ -170,6 +189,7 @@ export class EmbeddingIndexerService {
       vectors,
       { pageId, workspaceId, spaceId },
       modelName,
+      provider.fingerprint,
     );
 
     // HARD replace in one transaction: delete then insert so search never
@@ -216,7 +236,9 @@ export class EmbeddingIndexerService {
     // (seeded at enqueue time); the finally cleans that too.
     try {
       try {
-        await this.aiService.getEmbeddingModel(workspaceId);
+        // #530: resolve via the same path reindexPage uses (workspace provider,
+        // else the global TEI sidecar) so a global-only deployment is NOT skipped.
+        await this.aiService.resolveEmbeddingProvider(workspaceId);
       } catch (err) {
         if (err instanceof AiEmbeddingNotConfiguredException) {
           this.logger.log(
@@ -348,6 +370,7 @@ export class EmbeddingIndexerService {
     vectors: number[][],
     ids: { pageId: string; workspaceId: string; spaceId: string },
     modelName: string,
+    fingerprint: string | null,
   ): PageEmbeddingChunkRow[] {
     const rows: PageEmbeddingChunkRow[] = [];
     let cursor = 0;
@@ -370,6 +393,8 @@ export class EmbeddingIndexerService {
         // Provenance for a future re-index sweep on model change.
         modelName,
         modelDimensions: embedding.length,
+        // #530: the active generation fingerprint this row belongs to.
+        fingerprint,
         embedding,
       });
     }

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

@@ -1,14 +1,146 @@
 import type { ModelMessage } from 'ai';
 import {
   resolveReplayBudget,
+  resolveEffectiveReplayThreshold,
   isContextOverflowError,
   estimateMessagesTokens,
   trimHistoryForReplay,
   REPLAY_BUDGET_DEFAULT_TOKENS,
+  REPLAY_BUDGET_WINDOW_FRACTION,
+  REPLAY_MIN_FLOOR_TOKENS,
   REPLAY_TRUNCATION_MARKER,
   REPLAY_TURN_COLLAPSED_MARKER,
 } from './history-budget';
 
+describe('resolveEffectiveReplayThreshold (#520 iterative escalation)', () => {
+  // The escalation table: each consecutive overflow (k) deepens the cut by 0.5×.
+  it('scales the base by 0.5**k, flooring (not rounding) fractional tokens', () => {
+    const base = 100_000;
+    expect(resolveEffectiveReplayThreshold(base, 0)).toBe(base); // k=0: unchanged
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBe(50_000); // 0.5×
+    expect(resolveEffectiveReplayThreshold(base, 2)).toBe(25_000); // 0.25×
+    expect(resolveEffectiveReplayThreshold(base, 3)).toBe(12_500); // 0.125×
+    // Floors, not rounds.
+    expect(resolveEffectiveReplayThreshold(99_999, 1)).toBe(49_999);
+  });
+
+  it('passes a null base (trimming OFF) through unchanged for any k', () => {
+    for (const k of [0, 1, 2, 5, 100]) {
+      expect(resolveEffectiveReplayThreshold(null, k)).toBeNull();
+    }
+  });
+
+  // The crux of #520: convergence. A large k is clamped at REPLAY_MIN_FLOOR_TOKENS,
+  // so the escalation CONVERGES to a small-but-usable budget instead of trimming to
+  // zero — and, unlike the old fixed 0.5× that stuck at 50k, it drops far enough to
+  // fit a small real model window.
+  it('clamps a large k at the floor (converges, never below)', () => {
+    const base = 100_000;
+    for (const k of [4, 6, 10, 50, 200]) {
+      const t = resolveEffectiveReplayThreshold(base, k) as number;
+      expect(t).toBe(REPLAY_MIN_FLOOR_TOKENS);
+      expect(t).toBeGreaterThanOrEqual(REPLAY_MIN_FLOOR_TOKENS);
+    }
+  });
+
+  // Residual-brick regression (#520): flat-default base 100k, real window ~40k. The
+  // OLD fixed single 0.5× stuck at 50k > 40k forever (re-overflows every turn — the
+  // brick). The iterative cut drops BELOW 40k after a couple more consecutive
+  // overflows, so the history finally fits and the chat un-bricks.
+  it('un-bricks: escalation drops below a small real window the fixed 0.5× never could', () => {
+    const base = 100_000;
+    const realWindow = 40_000;
+    // The old terminal state: 0.5× = 50k, still above the window.
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBeGreaterThan(realWindow);
+    // Escalation converges under the window.
+    const converged = [2, 3, 4, 5].some(
+      (k) => (resolveEffectiveReplayThreshold(base, k) as number) < realWindow,
+    );
+    expect(converged).toBe(true);
+    // MUTATION SENTINEL: reverting `** k` to `** 1` (fixed 0.5×) makes every k yield
+    // 50k, so `converged` above would be FALSE and this test reddens. Removing the
+    // floor reddens the clamp test instead.
+  });
+
+  // "Don't INFLATE above itself" is still an invariant: the floor never RAISES a
+  // legitimately small configured budget above itself (that would re-overflow the
+  // very window it was set for). Under #520 Option B the floor is min(FLOOR,
+  // floor(0.5×base)), so for a base BELOW the floor recovery MAY cut it further —
+  // down to floor(0.5×base), i.e. AT LEAST the old single 0.5× cut — but never above.
+  it('never inflates a small configured budget above itself (may cut below it, #520 Option B)', () => {
+    const small = 5_000; // below REPLAY_MIN_FLOOR_TOKENS
+    const floor = Math.min(REPLAY_MIN_FLOOR_TOKENS, Math.floor(small * 0.5)); // 2500
+    expect(resolveEffectiveReplayThreshold(small, 0)).toBe(small); // k=0 unchanged
+    for (const k of [1, 2, 3, 10]) {
+      const t = resolveEffectiveReplayThreshold(small, k) as number;
+      // Invariant: never RAISED above the configured budget.
+      expect(t).toBeLessThanOrEqual(small);
+      // Option B: never trimmed below the absolute floor (converges).
+      expect(t).toBeGreaterThanOrEqual(floor);
+    }
+    // The old single 0.5× cut is reached (and pinned) — recovery is never WORSE than
+    // before, and DOES cut below the configured budget on overflow (Option B). Under
+    // the old floor==budget behavior this would be `small` (5000), not `floor`.
+    expect(resolveEffectiveReplayThreshold(small, 1)).toBe(floor);
+  });
+
+  // #520 Option B: with a configured window (NOT the flat default), repeated overflow
+  // escalates the replay budget BELOW the configured budget, down to the absolute
+  // floor — the chat must not brick even when the configured window is too large for
+  // the model. A CLEAN turn (k back to 0) restores the full configured budget.
+  it('escalates below the configured budget down to the floor, and resets on a clean turn', () => {
+    // Configured context window 8000 -> budget = floor(0.7 × 8000) = 5600 (the code's
+    // window->budget ratio; compute it, do not hardcode).
+    const window = 8_000;
+    const budget = resolveReplayBudget(window).thresholdTokens as number;
+    expect(budget).toBe(Math.floor(REPLAY_BUDGET_WINDOW_FRACTION * window)); // 5600
+
+    // The absolute floor for this budget = min(FLOOR, floor(0.5 × budget)) = 2800
+    // (a "small window": 0.5×budget < REPLAY_MIN_FLOOR_TOKENS), which is BELOW the
+    // 5600 configured budget.
+    const floor = Math.min(
+      REPLAY_MIN_FLOOR_TOKENS,
+      Math.floor(budget * 0.5),
+    ); // 2800
+    expect(floor).toBe(2_800);
+    expect(floor).toBeLessThan(budget); // below the configured budget
+
+    // k=0 -> full configured budget (no overflow yet).
+    expect(resolveEffectiveReplayThreshold(budget, 0)).toBe(budget);
+    // k=1 -> cut BELOW the configured budget to the 0.5× floor (2800). This is the
+    // MUTATION SENTINEL: with the OLD floor==configured-budget behavior this would be
+    // max(2800, 5600) = 5600 (never below budget), so the assertion reddens.
+    expect(resolveEffectiveReplayThreshold(budget, 1)).toBe(floor);
+    expect(resolveEffectiveReplayThreshold(budget, 1)).toBeLessThan(budget);
+    // k>=2 stays at the floor (converges, never below it).
+    for (const k of [2, 3, 5, 20]) {
+      expect(resolveEffectiveReplayThreshold(budget, k)).toBe(floor);
+    }
+
+    // A CLEAN turn resets k to 0 -> the full configured budget is restored (recovery
+    // is not sticky; it only bites while the provider keeps proving non-fit).
+    expect(resolveEffectiveReplayThreshold(budget, 0)).toBe(budget);
+  });
+
+  // #520 Option B, LARGE window: a big configured budget escalates in DISTINCT steps
+  // below itself, converging to the fixed absolute floor REPLAY_MIN_FLOOR_TOKENS.
+  it('a large configured budget escalates below itself down to REPLAY_MIN_FLOOR_TOKENS', () => {
+    const budget = resolveReplayBudget(200_000).thresholdTokens as number; // 140000
+    // The floor for a large window is the fixed REPLAY_MIN_FLOOR_TOKENS (8k), well
+    // BELOW the configured budget.
+    expect(
+      Math.min(REPLAY_MIN_FLOOR_TOKENS, Math.floor(budget * 0.5)),
+    ).toBe(REPLAY_MIN_FLOOR_TOKENS);
+    const seq = [0, 1, 2, 3, 4, 5, 6].map(
+      (k) => resolveEffectiveReplayThreshold(budget, k) as number,
+    );
+    expect(seq).toEqual([140_000, 70_000, 35_000, 17_500, 8_750, 8_000, 8_000]);
+    // Every escalated step (k>=1) is below the configured budget; convergence floor.
+    for (const t of seq.slice(1)) expect(t).toBeLessThan(budget);
+    expect(seq[seq.length - 1]).toBe(REPLAY_MIN_FLOOR_TOKENS);
+  });
+});
+
 describe('resolveReplayBudget', () => {
   it('uses floor(0.7 x window) for a configured window (no cap)', () => {
     // 0.7 x 60k = 42k

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

@@ -22,10 +22,35 @@ export const REPLAY_BUDGET_DEFAULT_TOKENS = 100_000;
 /** Fraction of a configured context window used as the budget. */
 export const REPLAY_BUDGET_WINDOW_FRACTION = 0.7;
 /**
- * Fraction of the normal budget used for the REACTIVE re-trim after a provider
- * context-overflow 400 — the preventive estimate under-counted, so cut harder.
+ * Per-step fraction of the normal budget applied on the REACTIVE re-trim after a
+ * provider context-overflow 400 — the preventive estimate under-counted, so cut
+ * harder. This is now applied ITERATIVELY: with `k` consecutive overflow turns the
+ * budget is scaled by `fraction ** k` (k=1 -> 0.5×, k=2 -> 0.25×, …), so recovery
+ * ESCALATES turn over turn until the replayed history finally fits, instead of the
+ * old single fixed 0.5× cut that could never un-brick a chat whose real model
+ * window is smaller than 0.5 × the (unconfigured, flat-default) base budget (#520).
  */
 export const REPLAY_AGGRESSIVE_FRACTION = 0.5;
+/**
+ * Absolute lower bound (tokens) on the escalating reactive budget: the iterative
+ * cut is clamped here so it CONVERGES (a fixed floor, not an ever-shrinking value
+ * that would eventually trim everything). Rationale: below ~8k tokens a chat cannot
+ * carry meaningful recent context, and even a small real model window comfortably
+ * fits this much — keep-recent-turns still applies on top, so a handful of recent
+ * turns survive.
+ *
+ * #520 Option B: for a LARGE configured window this floor sits BELOW the configured
+ * replay budget, and recovery is allowed to escalate PAST the configured budget down
+ * to it — "the chat must not brick on overflow" is an INVARIANT of the reactive
+ * branch, and a configured window is a claim about model capacity, not a promise
+ * about replay size; once the provider proves non-fit with a 400, reality beats the
+ * config. This does NOT change the NORMAL path's upper bound (#510 Option A still
+ * respects the configured budget while it fits) — it is survival once non-fit is
+ * proven. For a SMALL configured window the effective floor is instead the old
+ * single 0.5× cut (never worse than before, and never RAISED above the budget); see
+ * {@link resolveEffectiveReplayThreshold}.
+ */
+export const REPLAY_MIN_FLOOR_TOKENS = 8_000;
 /**
  * Turns (a user message + its assistant/tool replies) kept FULL at the tail,
  * including the current one — never trimmed. Older turns are compacted first.
@@ -85,22 +110,48 @@ export function resolveReplayBudget(rawContextWindow: unknown): ReplayBudget {
 }
 
 /**
- * The effective replay threshold for THIS turn, given the base budget and whether
- * the PREVIOUS turn hit a context-overflow 400 (the reactive-recovery signal,
- * `metadata.replayOverflow`). On recovery the base budget is scaled down by
- * {@link REPLAY_AGGRESSIVE_FRACTION}: the overflowing turn produced no usage
- * signal, so the preventive estimate under-counted and a normal-threshold trim may
- * not shrink enough to fit — this harder cut is what un-bricks the chat.
+ * The effective replay threshold for THIS turn, given the base budget and `k` — the
+ * number of CONSECUTIVE preceding turns that hit a context-overflow 400 (the
+ * reactive-recovery signal `metadata.replayOverflowCount`, read from the last
+ * assistant row). On recovery the base budget is scaled down ITERATIVELY by
+ * {@link REPLAY_AGGRESSIVE_FRACTION} ** k and clamped at {@link REPLAY_MIN_FLOOR_TOKENS}:
+ *   - k=0 -> base unchanged (no overflow: nothing to recover from).
+ *   - k=1 -> floor(0.5 × base); k=2 -> floor(0.25 × base); … each further consecutive
+ *     overflow tightens the cut, so recovery ESCALATES until the history fits.
+ *   - the escalation is clamped at the floor so it CONVERGES — this is what un-bricks
+ *     a chat whose real model window is smaller than a single 0.5× cut of the base
+ *     (e.g. an unconfigured window: flat-default base 100k, real window <50k) (#520).
+ *
+ * The overflowing turn produced no usage signal, so the preventive estimate
+ * under-counted and a normal-threshold (or single fixed 0.5×) trim may not shrink
+ * enough to fit; the escalating cut is what recovers such a chat.
  *
  * A `null` base budget (trimming OFF) is passed through unchanged: an explicit
  * off-switch is never overridden by the recovery path.
+ *
+ * The absolute floor is `min(REPLAY_MIN_FLOOR_TOKENS, floor(0.5 × base))` (#520
+ * Option B):
+ *   - LARGE window (floor(0.5 × base) >= REPLAY_MIN_FLOOR_TOKENS) -> the fixed
+ *     REPLAY_MIN_FLOOR_TOKENS, which is BELOW the configured budget: escalation is
+ *     ALLOWED to cut past the configured budget down to this absolute minimum, so a
+ *     chat whose real model window is far smaller than the configured window still
+ *     un-bricks (the invariant beats a config reality disproved with a 400).
+ *   - SMALL window (floor(0.5 × base) < REPLAY_MIN_FLOOR_TOKENS) -> floor(0.5 ×
+ *     base), i.e. AT LEAST the old single 0.5× cut: recovery is never WORSE than the
+ *     pre-#520 behavior, and the floor is still never RAISED above the configured
+ *     budget itself (that would re-overflow the very window it was set for).
  */
 export function resolveEffectiveReplayThreshold(
   thresholdTokens: number | null,
-  priorOverflowed: boolean,
+  k: number,
 ): number | null {
-  if (!priorOverflowed || thresholdTokens == null) return thresholdTokens;
-  return Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION);
+  if (thresholdTokens == null || k <= 0) return thresholdTokens;
+  const scaled = Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION ** k);
+  const floor = Math.min(
+    REPLAY_MIN_FLOOR_TOKENS,
+    Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION),
+  );
+  return Math.max(scaled, floor);
 }
 
 /**

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

@@ -106,6 +106,7 @@ function __assertClientCallContract(client: DocmostClientLike): void {
   void client.sharePage(s, true);
   void client.unsharePage(s);
   void client.restorePageVersion(s);
+  void client.savePageVersion(s);
   void client.transformPage(s, s, { dryRun: true });
   void client.stashPage(s);
   // --- write (image / footnote), in-app since #410 ---

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

@@ -58,6 +58,7 @@ type DocmostClientMethod =
   | 'sharePage'
   | 'unsharePage'
   | 'restorePageVersion'
+  | 'savePageVersion'
   | 'transformPage'
   | 'stashPage'
   // --- write (image / footnote), in-app since #410 ---

apps/server/src/core/page/dto/create-page.dto.ts

@@ -10,6 +10,12 @@ import { Transform } from 'class-transformer';
 
 export type ContentFormat = 'json' | 'markdown' | 'html';
 
+// READ-only rendering formats for `getPage` (#502). A superset of the writable
+// `ContentFormat` with `text` — a flat, deterministic, machine-diffable text
+// rendering — added. Kept SEPARATE from `ContentFormat` so the write path
+// (createPage/updatePage `parseProsemirrorContent`) can never be handed `text`.
+export type PageReadFormat = ContentFormat | 'text';
+
 export class CreatePageDto {
   @IsOptional()
   @IsString()

apps/server/src/core/page/dto/page.dto.ts

@@ -5,10 +5,11 @@ import {
   IsOptional,
   IsString,
   IsUUID,
+  MaxLength,
 } from 'class-validator';
 import { Transform } from 'class-transformer';
 
-import { ContentFormat } from './create-page.dto';
+import { PageReadFormat } from './create-page.dto';
 import { IsPageIdOrSlugId } from './page-identity.validator';
 
 export class PageIdDto {
@@ -43,8 +44,19 @@ export class PageInfoDto extends PageIdDto {
 
   @IsOptional()
   @Transform(({ value }) => value?.toLowerCase())
-  @IsIn(['json', 'markdown', 'html'])
-  format?: ContentFormat;
+  @IsIn(['json', 'markdown', 'html', 'text'])
+  format?: PageReadFormat;
+}
+
+export class PageWorkTimeDto extends PageIdDto {
+  // Viewer IANA timezone for the per-day punch-card buckets (§6.3). Optional —
+  // falls back to UTC server-side. Length-capped so a bogus value cannot bloat
+  // the request; the value is only ever handed to Intl.DateTimeFormat, which
+  // throws on an unknown zone (caught by the controller → 400).
+  @IsOptional()
+  @IsString()
+  @MaxLength(64)
+  tz?: string;
 }
 
 export class DeletePageDto extends PageIdDto {

apps/server/src/core/page/page.controller.getpage-format.spec.ts

@@ -0,0 +1,76 @@
+import { NotFoundException } from '@nestjs/common';
+import { PageController } from './page.controller';
+import { jsonToText, jsonToMarkdown } from '../../collaboration/collaboration.util';
+
+// #502 READ: getPage `format:"text"` routes through the deterministic jsonToText
+// path (placeholders for non-text nodes, block-per-line), while `format:"json"`
+// (or none) returns the raw content and `format:"markdown"` still converts. This
+// pins the CONTROLLER wiring with lightweight mocks (no DB needed — the jsonToText
+// output contract itself is pinned in json-to-text-deterministic.spec.ts).
+
+const CONTENT = {
+  type: 'doc',
+  content: [
+    { type: 'heading', attrs: { level: 1 }, content: [{ type: 'text', text: 'Config' }] },
+    {
+      type: 'paragraph',
+      content: [
+        { type: 'text', text: 'key=', marks: [{ type: 'bold' }] },
+        { type: 'text', text: 'value' },
+      ],
+    },
+    { type: 'image', attrs: { src: 's' } },
+  ],
+};
+
+function makeController(page: any): PageController {
+  const pageRepo = { findById: jest.fn().mockResolvedValue(page) } as any;
+  const pageAccessService = {
+    validateCanViewWithPermissions: jest
+      .fn()
+      .mockResolvedValue({ canEdit: true, hasRestriction: false }),
+  } as any;
+  // Only pageRepo + pageAccessService are exercised by getPage; the rest are
+  // never touched on this path, so undefined placeholders are fine.
+  return new PageController(
+    undefined as any, // pageService
+    pageRepo,
+    undefined as any, // pageHistoryService
+    undefined as any, // spaceAbility
+    pageAccessService,
+    undefined as any, // backlinkService
+    undefined as any, // labelService
+    undefined as any, // auditService
+  );
+}
+
+const user = { id: 'u1' } as any;
+
+describe('PageController.getPage — format:"text" (#502)', () => {
+  it('returns deterministic flat text with placeholders for non-text nodes', async () => {
+    const controller = makeController({ id: 'p1', content: CONTENT });
+    const res: any = await controller.getPage({ pageId: 'p1', format: 'text' } as any, user);
+    expect(res.content).toBe(jsonToText(CONTENT, { deterministic: true }));
+    expect(res.content).toBe('Config\nkey=value\n[image]');
+    expect(res.permissions).toEqual({ canEdit: true, hasRestriction: false });
+  });
+
+  it('markdown format still converts to markdown (unchanged)', async () => {
+    const controller = makeController({ id: 'p1', content: CONTENT });
+    const res: any = await controller.getPage({ pageId: 'p1', format: 'markdown' } as any, user);
+    expect(res.content).toBe(jsonToMarkdown(CONTENT));
+  });
+
+  it('json / no format returns the raw ProseMirror content object', async () => {
+    const controller = makeController({ id: 'p1', content: CONTENT });
+    const res: any = await controller.getPage({ pageId: 'p1' } as any, user);
+    expect(res.content).toEqual(CONTENT); // untouched object, not a string
+  });
+
+  it('missing page -> NotFoundException', async () => {
+    const controller = makeController(null);
+    await expect(
+      controller.getPage({ pageId: 'nope', format: 'text' } as any, user),
+    ).rejects.toBeInstanceOf(NotFoundException);
+  });
+});

apps/server/src/core/page/page.controller.spec.ts

@@ -1,3 +1,8 @@
+import {
+  BadRequestException,
+  ForbiddenException,
+  NotFoundException,
+} from '@nestjs/common';
 import { PageController } from './page.controller';
 
 // Direct instantiation with stub deps. The Test.createTestingModule form failed
@@ -22,4 +27,88 @@ describe('PageController', () => {
   it('should be defined', () => {
     expect(controller).toBeDefined();
   });
+
+  // #395 — the work-time endpoint must be gated exactly like /history.
+  describe('getPageWorkTime', () => {
+    const user = { id: 'u1' } as any;
+
+    function build(overrides: {
+      page?: any;
+      validate?: jest.Mock;
+      compute?: jest.Mock;
+    }) {
+      const pageRepo = { findById: jest.fn().mockResolvedValue(overrides.page) };
+      const pageAccessService = {
+        validateCanView: overrides.validate ?? jest.fn().mockResolvedValue(undefined),
+      };
+      const pageHistoryService = {
+        computeWorkTime:
+          overrides.compute ?? jest.fn().mockResolvedValue({ workMs: 0 }),
+      };
+      const c = new PageController(
+        {} as any,
+        pageRepo as any,
+        pageHistoryService as any,
+        {} as any,
+        pageAccessService as any,
+        {} as any,
+        {} as any,
+        {} as any,
+      );
+      return { c, pageRepo, pageAccessService, pageHistoryService };
+    }
+
+    it('404s when the page does not exist', async () => {
+      const { c } = build({ page: null });
+      await expect(
+        c.getPageWorkTime({ pageId: 'p1' } as any, user),
+      ).rejects.toBeInstanceOf(NotFoundException);
+    });
+
+    it('enforces validateCanView before computing, then delegates with tz', async () => {
+      const validate = jest.fn().mockResolvedValue(undefined);
+      const compute = jest.fn().mockResolvedValue({ workMs: 42 });
+      const { c } = build({ page: { id: 'pg' }, validate, compute });
+      const out = await c.getPageWorkTime(
+        { pageId: 'pg', tz: 'Europe/Moscow' } as any,
+        user,
+      );
+      expect(validate).toHaveBeenCalledWith({ id: 'pg' }, user);
+      expect(compute).toHaveBeenCalledWith('pg', 'Europe/Moscow');
+      expect(out).toEqual({ workMs: 42 });
+    });
+
+    it('propagates a denied view gate and does NOT reach compute (security)', async () => {
+      // If validateCanView is moved AFTER computeWorkTime, the timeline of a page
+      // the caller may not see would be read/estimated before the gate — this
+      // locks the order: a rejecting gate must short-circuit before any compute.
+      const validate = jest.fn().mockRejectedValue(new ForbiddenException());
+      const compute = jest.fn().mockResolvedValue({ workMs: 1 });
+      const { c, pageHistoryService } = build({
+        page: { id: 'pg' },
+        validate,
+        compute,
+      });
+      await expect(
+        c.getPageWorkTime({ pageId: 'pg' } as any, user),
+      ).rejects.toBeInstanceOf(ForbiddenException);
+      expect(pageHistoryService.computeWorkTime).not.toHaveBeenCalled();
+    });
+
+    it('maps an unknown-timezone RangeError to a 400', async () => {
+      const compute = jest.fn().mockRejectedValue(new RangeError('bad tz'));
+      const { c } = build({ page: { id: 'pg' }, compute });
+      await expect(
+        c.getPageWorkTime({ pageId: 'pg', tz: 'X/Y' } as any, user),
+      ).rejects.toBeInstanceOf(BadRequestException);
+    });
+
+    it('does not swallow a non-RangeError from the service', async () => {
+      const compute = jest.fn().mockRejectedValue(new Error('db down'));
+      const { c } = build({ page: { id: 'pg' }, compute });
+      await expect(
+        c.getPageWorkTime({ pageId: 'pg' } as any, user),
+      ).rejects.toThrow('db down');
+    });
+  });
 });

apps/server/src/core/page/page.controller.ts

@@ -21,6 +21,7 @@ import {
   PageHistoryIdDto,
   PageIdDto,
   PageInfoDto,
+  PageWorkTimeDto,
 } from './dto/page.dto';
 import { PageHistoryService } from './services/page-history.service';
 import { AuthUser } from '../../common/decorators/auth-user.decorator';
@@ -49,6 +50,7 @@ import { AddLabelsDto, RemoveLabelDto } from '../label/dto/label.dto';
 import {
   jsonToHtml,
   jsonToMarkdown,
+  jsonToText,
 } from '../../collaboration/collaboration.util';
 import { AuditEvent, AuditResource } from '../../common/events/audit-events';
 import {
@@ -93,10 +95,16 @@ export class PageController {
     const permissions = { canEdit, hasRestriction };
 
     if (dto.format && dto.format !== 'json' && page.content) {
-      const contentOutput =
-        dto.format === 'markdown'
-          ? jsonToMarkdown(page.content)
-          : jsonToHtml(page.content);
+      let contentOutput: string;
+      if (dto.format === 'markdown') {
+        contentOutput = jsonToMarkdown(page.content);
+      } else if (dto.format === 'text') {
+        // #502: flat, deterministic, machine-diffable text (block-per-line,
+        // inline marks/anchors dropped, non-text nodes -> stable placeholders).
+        contentOutput = jsonToText(page.content, { deterministic: true });
+      } else {
+        contentOutput = jsonToHtml(page.content);
+      }
       return {
         ...page,
         content: contentOutput,
@@ -524,6 +532,32 @@ export class PageController {
     return this.pageHistoryService.findHistoryByPageId(page.id, pagination);
   }
 
+  @HttpCode(HttpStatus.OK)
+  @Post('/history/time')
+  async getPageWorkTime(
+    @Body() dto: PageWorkTimeDto,
+    @AuthUser() user: User,
+  ) {
+    const page = await this.pageRepo.findById(dto.pageId);
+    if (!page) {
+      throw new NotFoundException('Page not found');
+    }
+
+    // Same view gate as /history and /history/info.
+    await this.pageAccessService.validateCanView(page, user);
+
+    try {
+      return await this.pageHistoryService.computeWorkTime(page.id, dto.tz);
+    } catch (e) {
+      // Intl.DateTimeFormat throws RangeError on an unknown IANA zone; surface
+      // it as a 400 rather than a 500.
+      if (e instanceof RangeError) {
+        throw new BadRequestException('Invalid timezone');
+      }
+      throw e;
+    }
+  }
+
   @HttpCode(HttpStatus.OK)
   @Post('/history/info')
   async getPageHistoryInfo(
@@ -818,6 +852,11 @@ export class PageController {
       throw new NotFoundException('Page not found');
     }
 
+    // Target-only validateCanView is intentional: getPageBreadCrumbs returns
+    // the full ancestor chain WITHOUT per-ancestor permission filtering. Safe
+    // because page restrictions inherit down the tree, so any ancestor the
+    // caller could not view would already hide the target here — see the
+    // getPageBreadCrumbs docstring / #471.
     await this.pageAccessService.validateCanView(page, user);
 
     return this.pageService.getPageBreadCrumbs(page.id);

apps/server/src/core/page/services/page-history.service.ts

@@ -3,6 +3,23 @@ import { PageHistoryRepo } from '@docmost/db/repos/page/page-history.repo';
 import { PageHistory } from '@docmost/db/types/entity.types';
 import { PaginationOptions } from '@docmost/db/pagination/pagination-options';
 import { CursorPaginationResult } from '@docmost/db/pagination/cursor-pagination';
+import {
+  computeWorkTime,
+  bucketByDay,
+  DEFAULT_WORK_TIME_CONFIG,
+  WorkTimeConfig,
+  PerDay,
+} from '../work-time';
+
+export interface PageWorkTime {
+  workMs: number;
+  agentOnlyMs: number;
+  perDay: PerDay[];
+  /** the config actually used, so the UI can show "≈" + the T_gap threshold. */
+  config: WorkTimeConfig;
+  /** the tz the per-day buckets were computed in (echoed back for the label). */
+  tz: string;
+}
 
 @Injectable()
 export class PageHistoryService {
@@ -23,4 +40,33 @@ export class PageHistoryService {
       paginationOptions,
     );
   }
+
+  /**
+   * #395 — estimate time worked on a page (§5) and bucket it into the viewer's
+   * calendar days for the punch-card (§6.3). Reads only the cheap history
+   * projection (no `content`); the estimate itself is a pure, deterministic
+   * function so it is unit-tested exhaustively without a DB.
+   *
+   * `tz` is the viewer's IANA zone (browser locale) — it moves which day a
+   * session lands in and where its windows sit, but never the total (§10).
+   */
+  async computeWorkTime(
+    pageId: string,
+    tz = 'UTC',
+    config?: Partial<WorkTimeConfig>,
+  ): Promise<PageWorkTime> {
+    const rows = await this.pageHistoryRepo.findTimelineByPageId(pageId);
+    const result = computeWorkTime(rows, config);
+    const usedConfig: WorkTimeConfig = { ...DEFAULT_WORK_TIME_CONFIG, ...config };
+    // `bucketByDay` consumes the pure core's un-bucketed sessions here; the
+    // full session list is NOT shipped on the response (no client reads it).
+    const perDay = bucketByDay(result.sessions, tz);
+    return {
+      workMs: result.workMs,
+      agentOnlyMs: result.agentOnlyMs,
+      perDay,
+      config: usedConfig,
+      tz,
+    };
+  }
 }

apps/server/src/core/page/services/page.service.ts

@@ -1071,6 +1071,29 @@ export class PageService {
     });
   }
 
+  /**
+   * Walk the ancestor chain of `childPageId` up to the space root, filtered
+   * ONLY by `deletedAt` (+ MAX_PAGE_TREE_DEPTH) — WITHOUT per-ancestor
+   * permission filtering. Callers that expose this to a user (the
+   * `/breadcrumbs` endpoint) validate `validateCanView` on the TARGET page
+   * only, then return the whole chain of ancestor titles (#471).
+   *
+   * This is safe — NOT a title leak — because page restrictions inherit DOWN
+   * the tree: to view a page the caller must hold permission on EVERY
+   * restricted ancestor (`validateCanView` -> `canUserAccessPage` checks the
+   * full ancestor chain — see page-access.service.ts / page-permission.repo.ts
+   * `canUserEditPage`). A restricted ancestor the caller may not see would
+   * therefore already hide the TARGET page itself, so every ancestor reachable
+   * here is one the caller is already entitled to view (content stays gated
+   * regardless — getPage/getNode re-check permissions).
+   *
+   * Note the guarantee is the narrow "may view a descendant => may view its
+   * ancestors", NOT "space membership sees every page" — restricted subtrees do
+   * hide pages from members. Per-ancestor permission filtering here was
+   * considered and declined as redundant given the inheritance invariant above
+   * (#471). The same chain feeds the web-UI breadcrumb bar under identical CASL
+   * scope.
+   */
   async getPageBreadCrumbs(childPageId: string, trx?: KyselyTransaction) {
     const ancestors = await dbOrTx(this.db, trx)
       .withRecursive('page_ancestors', (db) =>

apps/server/src/core/page/work-time/bucket-by-day.spec.ts

@@ -0,0 +1,129 @@
+import { bucketByDay, zonedDayStart } from './bucket-by-day';
+import { computeWorkTime } from './compute-work-time';
+import { WorkSession, TimelineSample } from './work-time.types';
+
+const MIN = 60 * 1000;
+const HOUR = 60 * MIN;
+
+function work(start: number, end: number): WorkSession {
+  return { start, end, class: 'work' };
+}
+function agent(start: number, end: number): WorkSession {
+  return { start, end, class: 'agent_only' };
+}
+
+function sumActive(perDay: ReturnType<typeof bucketByDay>): number {
+  return perDay.reduce((a, d) => a + d.activeMs, 0);
+}
+
+describe('bucketByDay', () => {
+  it('Σ activeMs == workMs — the §6.3 consistency invariant', () => {
+    const rows: TimelineSample[] = [
+      { createdAt: '2026-07-04T03:40:00Z', lastUpdatedById: 'h', lastUpdatedSource: 'user', lastUpdatedAiChatId: null, kind: null },
+      { createdAt: '2026-07-04T03:49:00Z', lastUpdatedById: 'h', lastUpdatedSource: 'user', lastUpdatedAiChatId: null, kind: null },
+      { createdAt: '2026-07-04T18:11:00Z', lastUpdatedById: 'h', lastUpdatedSource: 'user', lastUpdatedAiChatId: null, kind: null },
+      { createdAt: '2026-07-06T15:34:00Z', lastUpdatedById: 'h', lastUpdatedSource: 'user', lastUpdatedAiChatId: null, kind: null },
+    ];
+    const r = computeWorkTime(rows);
+    const perDay = bucketByDay(r.sessions, 'UTC');
+    expect(sumActive(perDay)).toBe(r.workMs);
+  });
+
+  it('empty input → no days', () => {
+    expect(bucketByDay([], 'UTC')).toEqual([]);
+  });
+
+  it('midnight-crossing session splits across two days, sum preserved (§9#9)', () => {
+    const start = Date.UTC(2026, 0, 10, 23, 14);
+    const end = Date.UTC(2026, 0, 11, 0, 40);
+    const perDay = bucketByDay([work(start, end)], 'UTC');
+    expect(perDay).toHaveLength(2);
+    expect(perDay[0].dayISO).toBe('2026-01-10');
+    expect(perDay[1].dayISO).toBe('2026-01-11');
+    expect(perDay[0].activeMs).toBe(46 * MIN); // 23:14 → 24:00
+    expect(perDay[1].activeMs).toBe(40 * MIN); // 00:00 → 00:40
+    expect(sumActive(perDay)).toBe(end - start);
+  });
+
+  it('empty days between active days are emitted, not skipped (§9#12)', () => {
+    const d1 = work(Date.UTC(2026, 0, 10, 10, 0), Date.UTC(2026, 0, 10, 11, 0));
+    const d3 = work(Date.UTC(2026, 0, 12, 10, 0), Date.UTC(2026, 0, 12, 11, 0));
+    const perDay = bucketByDay([d1, d3], 'UTC');
+    expect(perDay.map((d) => d.dayISO)).toEqual([
+      '2026-01-10',
+      '2026-01-11',
+      '2026-01-12',
+    ]);
+    expect(perDay[1].activeMs).toBe(0);
+    expect(perDay[1].windows).toEqual([]);
+  });
+
+  it('agent_only windows are drawn but excluded from activeMs', () => {
+    const w = work(Date.UTC(2026, 0, 10, 9, 0), Date.UTC(2026, 0, 10, 10, 0));
+    const a = agent(Date.UTC(2026, 0, 10, 14, 0), Date.UTC(2026, 0, 10, 14, 30));
+    const perDay = bucketByDay([w, a], 'UTC');
+    expect(perDay).toHaveLength(1);
+    expect(perDay[0].activeMs).toBe(1 * HOUR);
+    expect(perDay[0].agentMs).toBe(30 * MIN);
+    expect(perDay[0].windows.map((x) => x.class)).toEqual(['work', 'agent_only']);
+  });
+
+  it('work and agent_only are unioned SEPARATELY (agent does not swallow work)', () => {
+    // Overlapping work + agent windows on the same day.
+    const w = work(Date.UTC(2026, 0, 10, 9, 0), Date.UTC(2026, 0, 10, 11, 0));
+    const a = agent(Date.UTC(2026, 0, 10, 10, 0), Date.UTC(2026, 0, 10, 12, 0));
+    const perDay = bucketByDay([w, a], 'UTC');
+    expect(perDay[0].activeMs).toBe(2 * HOUR);
+    expect(perDay[0].agentMs).toBe(2 * HOUR);
+  });
+
+  it('overlapping same-class sessions are UNIONed, not summed (no double-count)', () => {
+    // Two work sessions that overlap 10:00–10:30 on one day.
+    const a = work(Date.UTC(2026, 0, 10, 9, 0), Date.UTC(2026, 0, 10, 10, 30));
+    const b = work(Date.UTC(2026, 0, 10, 10, 0), Date.UTC(2026, 0, 10, 11, 0));
+    const perDay = bucketByDay([a, b], 'UTC');
+    expect(perDay).toHaveLength(1);
+    // Union 09:00–11:00 = 2h, NOT 90m + 60m = 150m.
+    expect(perDay[0].activeMs).toBe(2 * HOUR);
+    // The drawn windows are also merged to one, so the punch-card cannot render
+    // an overlapping double bar.
+    expect(perDay[0].windows).toHaveLength(1);
+    expect(perDay[0].windows[0].start).toBe(a.start);
+    expect(perDay[0].windows[0].end).toBe(b.end);
+  });
+
+  it('DST fall-back: a full 25-hour day still balances (§9#14)', () => {
+    // America/New_York ends DST 2026-11-01 (25h day).
+    const tz = 'America/New_York';
+    const dayStart = zonedDayStart(Date.UTC(2026, 10, 1, 12, 0), tz);
+    const nextStart = zonedDayStart(dayStart + 26 * HOUR, tz);
+    expect(nextStart - dayStart).toBe(25 * HOUR);
+    const perDay = bucketByDay([work(dayStart, nextStart)], tz);
+    expect(perDay).toHaveLength(1);
+    expect(perDay[0].dayISO).toBe('2026-11-01');
+    expect(perDay[0].activeMs).toBe(25 * HOUR);
+    expect(sumActive(perDay)).toBe(nextStart - dayStart);
+  });
+
+  it('DST spring-forward: a full 23-hour day still balances (§9#14)', () => {
+    // America/New_York starts DST 2026-03-08 (23h day).
+    const tz = 'America/New_York';
+    const dayStart = zonedDayStart(Date.UTC(2026, 2, 8, 12, 0), tz);
+    const nextStart = zonedDayStart(dayStart + 26 * HOUR, tz);
+    expect(nextStart - dayStart).toBe(23 * HOUR);
+    const perDay = bucketByDay([work(dayStart, nextStart)], tz);
+    expect(perDay).toHaveLength(1);
+    expect(perDay[0].activeMs).toBe(23 * HOUR);
+    expect(sumActive(perDay)).toBe(nextStart - dayStart);
+  });
+
+  it('tz changes the day a session lands in but not the total', () => {
+    const start = Date.UTC(2026, 0, 10, 2, 0); // 02:00 UTC
+    const end = Date.UTC(2026, 0, 10, 3, 0);
+    const utc = bucketByDay([work(start, end)], 'UTC');
+    const ny = bucketByDay([work(start, end)], 'America/New_York'); // 21:00 prev day
+    expect(utc[0].dayISO).toBe('2026-01-10');
+    expect(ny[0].dayISO).toBe('2026-01-09');
+    expect(sumActive(utc)).toBe(sumActive(ny));
+  });
+});

apps/server/src/core/page/work-time/bucket-by-day.ts

@@ -0,0 +1,180 @@
+import { WorkSession, PerDay, DayWindow } from './work-time.types';
+
+/**
+ * Merge intervals into a disjoint, sorted union. Overlapping OR touching
+ * intervals are joined. Empty input → [].
+ */
+function union(intervals: Array<[number, number]>): Array<[number, number]> {
+  if (intervals.length === 0) return [];
+  const sorted = [...intervals].sort((a, b) => a[0] - b[0]);
+  const out: Array<[number, number]> = [];
+  let [curStart, curEnd] = sorted[0];
+  for (let i = 1; i < sorted.length; i++) {
+    const [s, e] = sorted[i];
+    if (s <= curEnd) {
+      if (e > curEnd) curEnd = e;
+    } else {
+      out.push([curStart, curEnd]);
+      curStart = s;
+      curEnd = e;
+    }
+  }
+  out.push([curStart, curEnd]);
+  return out;
+}
+
+// Cache one Intl formatter per tz — constructing them is comparatively costly.
+const fmtCache = new Map<string, Intl.DateTimeFormat>();
+
+function partsFmt(tz: string): Intl.DateTimeFormat {
+  let fmt = fmtCache.get(tz);
+  if (!fmt) {
+    fmt = new Intl.DateTimeFormat('en-US', {
+      timeZone: tz,
+      year: 'numeric',
+      month: '2-digit',
+      day: '2-digit',
+      hour: '2-digit',
+      minute: '2-digit',
+      second: '2-digit',
+      hour12: false,
+    });
+    fmtCache.set(tz, fmt);
+  }
+  return fmt;
+}
+
+interface WallParts {
+  year: number;
+  month: number;
+  day: number;
+  hour: number;
+  minute: number;
+  second: number;
+}
+
+/** Wall-clock parts of an instant in `tz` (DST-correct, via Intl). */
+function wallParts(ms: number, tz: string): WallParts {
+  const parts = partsFmt(tz).formatToParts(new Date(ms));
+  const get = (type: string) =>
+    Number(parts.find((p) => p.type === type)?.value ?? '0');
+  let hour = get('hour');
+  // Intl emits "24" for midnight under some engines/locales; normalize to 0.
+  if (hour === 24) hour = 0;
+  return {
+    year: get('year'),
+    month: get('month'),
+    day: get('day'),
+    hour,
+    minute: get('minute'),
+    second: get('second'),
+  };
+}
+
+/** tz offset (wall − real) at an instant, in ms. */
+function offset(ms: number, tz: string): number {
+  const p = wallParts(ms, tz);
+  const asUTC = Date.UTC(p.year, p.month - 1, p.day, p.hour, p.minute, p.second);
+  return asUTC - ms;
+}
+
+/**
+ * Epoch-ms of the local-midnight day start of `ms` in `tz`. DST-correct: takes
+ * the calendar day of the instant, its wall-midnight, then converts back with
+ * the offset that actually applies AT that midnight (refined once). The rare
+ * tz-with-a-DST-transition-exactly-at-midnight case is a documented edge (§9#14).
+ */
+export function zonedDayStart(ms: number, tz: string): number {
+  const p = wallParts(ms, tz);
+  const wallMidnightAsUTC = Date.UTC(p.year, p.month - 1, p.day, 0, 0, 0);
+  let start = wallMidnightAsUTC - offset(ms, tz);
+  // Refine with the offset at the computed midnight (DST may differ from `ms`).
+  start = wallMidnightAsUTC - offset(start, tz);
+  return start;
+}
+
+/** The next local midnight after `dayStart` (handles 23/25h DST days). */
+function nextDayStart(dayStart: number, tz: string): number {
+  // +26h always lands inside the NEXT calendar day (day length ∈ [23h,25h]),
+  // never two days ahead; startOf('day') of it is the next midnight.
+  return zonedDayStart(dayStart + 26 * 60 * 60 * 1000, tz);
+}
+
+function isoDay(dayStart: number, tz: string): string {
+  const p = wallParts(dayStart, tz);
+  const pad = (n: number) => String(n).padStart(2, '0');
+  return `${p.year}-${pad(p.month)}-${pad(p.day)}`;
+}
+
+/** Clip a union to [lo, hi) and emit windows of `class`. */
+function clip(
+  merged: Array<[number, number]>,
+  lo: number,
+  hi: number,
+  cls: DayWindow['class'],
+): DayWindow[] {
+  const out: DayWindow[] = [];
+  for (const [s, e] of merged) {
+    const start = Math.max(s, lo);
+    const end = Math.min(e, hi);
+    if (end > start) out.push({ start, end, class: cls });
+  }
+  return out;
+}
+
+/**
+ * #395 §6.3 — bucket sessions into calendar days of `tz` for the punch-card.
+ * Pure and deterministic. `work` and `agent_only` are unioned SEPARATELY (else
+ * agent windows would swallow work windows on overlap), then each union is split
+ * at tz midnight boundaries (`startOf('day')` in tz, NOT "+24h" — DST-safe §9#14)
+ * and clipped to each day.
+ *
+ * By construction Σ perDay.activeMs == workMs: the days are a partition of the
+ * `work` union — no loss, no dup, even on 23/25h DST days. `agent_only` windows
+ * are drawn but NOT in activeMs. Empty days between the first and last active day
+ * are emitted (empty track + "—") so the rhythm/pauses stay visible.
+ */
+export function bucketByDay(sessions: WorkSession[], tz: string): PerDay[] {
+  const uWork = union(
+    sessions.filter((s) => s.class === 'work').map((s) => [s.start, s.end]),
+  );
+  const uAgent = union(
+    sessions
+      .filter((s) => s.class === 'agent_only')
+      .map((s) => [s.start, s.end]),
+  );
+
+  if (uWork.length === 0 && uAgent.length === 0) return [];
+
+  const minStart = Math.min(
+    uWork.length ? uWork[0][0] : Infinity,
+    uAgent.length ? uAgent[0][0] : Infinity,
+  );
+  const maxEnd = Math.max(
+    uWork.length ? uWork[uWork.length - 1][1] : -Infinity,
+    uAgent.length ? uAgent[uAgent.length - 1][1] : -Infinity,
+  );
+
+  const perDay: PerDay[] = [];
+  let dayStart = zonedDayStart(minStart, tz);
+  // Guard against a pathological non-advancing boundary.
+  let guard = 0;
+  while (dayStart < maxEnd && guard < 100000) {
+    guard++;
+    const dayEnd = nextDayStart(dayStart, tz);
+    const workWin = clip(uWork, dayStart, dayEnd, 'work');
+    const agentWin = clip(uAgent, dayStart, dayEnd, 'agent_only');
+    const activeMs = workWin.reduce((a, w) => a + (w.end - w.start), 0);
+    const agentMs = agentWin.reduce((a, w) => a + (w.end - w.start), 0);
+    const windows = [...workWin, ...agentWin].sort((a, b) => a.start - b.start);
+    perDay.push({
+      day: dayStart,
+      dayISO: isoDay(dayStart, tz),
+      activeMs,
+      agentMs,
+      windows,
+    });
+    dayStart = dayEnd;
+  }
+  return perDay;
+}

apps/server/src/core/page/work-time/compute-work-time.spec.ts

@@ -0,0 +1,358 @@
+import { computeWorkTime } from './compute-work-time';
+import { bucketByDay } from './bucket-by-day';
+import { TimelineSample, WorkSession } from './work-time.types';
+
+const MIN = 60 * 1000;
+
+/** Union wall-clock of a set of intervals (touching intervals merge). */
+function unionMs(intervals: Array<[number, number]>): number {
+  if (intervals.length === 0) return 0;
+  const sorted = [...intervals].sort((a, b) => a[0] - b[0]);
+  let total = 0;
+  let [cs, ce] = sorted[0];
+  for (let i = 1; i < sorted.length; i++) {
+    const [s, e] = sorted[i];
+    if (s <= ce) {
+      if (e > ce) ce = e;
+    } else {
+      total += ce - cs;
+      cs = s;
+      ce = e;
+    }
+  }
+  return total + (ce - cs);
+}
+
+const ivsOf = (sessions: WorkSession[], cls?: string): Array<[number, number]> =>
+  sessions
+    .filter((x) => cls == null || x.class === cls)
+    .map((x) => [x.start, x.end] as [number, number]);
+
+function s(
+  iso: string,
+  opts: {
+    source?: string | null;
+    chat?: string | null;
+    kind?: string | null;
+    by?: string | null;
+  } = {},
+): TimelineSample {
+  return {
+    createdAt: `${iso}Z`,
+    lastUpdatedById: opts.by ?? 'human-1',
+    lastUpdatedSource: opts.source === undefined ? 'user' : opts.source,
+    lastUpdatedAiChatId: opts.chat ?? null,
+    kind: opts.kind ?? null,
+  };
+}
+
+// §7 config: T_gap=30m, P_in+P_out=10m, P_single=2m.
+const S7 = { tGap: 30 * MIN, agentTGap: 30 * MIN, pIn: 5 * MIN, pOut: 5 * MIN, pSingle: 2 * MIN };
+
+describe('computeWorkTime', () => {
+  it('§7 fixture — sessionizes 20-ish samples to ≈1h32m, not the ≈60h naive span', () => {
+    const rows: TimelineSample[] = [
+      // S1: multi-sample morning session
+      s('2026-07-04T03:40:00'),
+      s('2026-07-04T03:45:00'),
+      s('2026-07-04T03:49:00'),
+      // S2: agent burst (one run) then human supervising → class work
+      s('2026-07-04T15:43:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T15:47:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T15:50:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T16:13:00'),
+      // S3: single
+      s('2026-07-04T18:11:00'),
+      // S4: multi-sample evening session
+      s('2026-07-04T19:38:00'),
+      s('2026-07-04T19:44:00'),
+      s('2026-07-04T19:54:00'),
+      // S5 / S6: two singles two days later, 44m apart → two sessions at T_gap=30
+      s('2026-07-06T15:34:00'),
+      s('2026-07-06T16:18:00'),
+    ];
+
+    const r = computeWorkTime(rows, S7);
+
+    // 19 + 40 + 2 + 26 + 2 + 2 = 91 minutes.
+    expect(r.workMs).toBe(91 * MIN);
+    expect(r.agentOnlyMs).toBe(0);
+    expect(r.sessions).toHaveLength(6);
+    expect(r.sessions.every((x) => x.class === 'work')).toBe(true);
+
+    const naiveSpan =
+      new Date('2026-07-06T16:18:00Z').getTime() -
+      new Date('2026-07-04T03:40:00Z').getTime();
+    expect(naiveSpan).toBeGreaterThan(60 * 60 * MIN); // ≈60h
+    expect(r.workMs).toBeLessThan(naiveSpan / 30); // dramatically smaller
+  });
+
+  it('n=0 → zero, no sessions', () => {
+    const r = computeWorkTime([]);
+    expect(r).toEqual({ workMs: 0, agentOnlyMs: 0, sessions: [] });
+  });
+
+  it('n=1 human → one P_single work session', () => {
+    const r = computeWorkTime([s('2026-07-04T10:00:00')], S7);
+    expect(r.sessions).toHaveLength(1);
+    expect(r.sessions[0].class).toBe('work');
+    expect(r.workMs).toBe(2 * MIN);
+    expect(r.agentOnlyMs).toBe(0);
+    // pre-roll only: [t − P_single, t]
+    expect(r.sessions[0].end).toBe(new Date('2026-07-04T10:00:00Z').getTime());
+  });
+
+  it('n=1 agent → one P_single agent_only session, work=0 (§9#2)', () => {
+    const r = computeWorkTime(
+      [s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' })],
+      S7,
+    );
+    expect(r.sessions).toHaveLength(1);
+    expect(r.sessions[0].class).toBe('agent_only');
+    expect(r.workMs).toBe(0);
+    expect(r.agentOnlyMs).toBe(2 * MIN);
+  });
+
+  it('MUST close the last session — the newest session is not lost (§9#1)', () => {
+    // Two singles a day apart: without the post-loop close, the 2nd is dropped.
+    const rows = [s('2026-07-04T10:00:00'), s('2026-07-05T10:00:00')];
+    const r = computeWorkTime(rows, S7);
+    expect(r.sessions).toHaveLength(2);
+    const lastStart = Math.max(...r.sessions.map((x) => x.start));
+    expect(lastStart).toBe(
+      new Date('2026-07-05T10:00:00Z').getTime() - 2 * MIN,
+    );
+    expect(r.workMs).toBe(4 * MIN);
+  });
+
+  it('agent-burst collapse: density does not inflate — length = wall-clock', () => {
+    const span = ['00', '01', '02', '03', '04', '05', '06'];
+    const dense: TimelineSample[] = span.map((sec) =>
+      s(`2026-07-04T10:00:${sec}`, { source: 'agent', chat: 'c1', kind: 'agent' }),
+    );
+    const sparse: TimelineSample[] = [
+      s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:00:06', { source: 'agent', chat: 'c1', kind: 'agent' }),
+    ];
+    const rDense = computeWorkTime(dense, S7);
+    const rSparse = computeWorkTime(sparse, S7);
+    // Same 6-second wall-clock span → same estimate regardless of snapshot count.
+    expect(rDense.agentOnlyMs).toBe(rSparse.agentOnlyMs);
+    expect(rDense.sessions).toHaveLength(1);
+    expect(rDense.sessions[0].class).toBe('agent_only');
+  });
+
+  it('supervisory agent time inside a human session counts as work, not agent', () => {
+    const rows = [
+      s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:05:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:12:00'), // human within T_gap
+    ];
+    const r = computeWorkTime(rows, S7);
+    expect(r.sessions).toHaveLength(1);
+    expect(r.sessions[0].class).toBe('work');
+    expect(r.agentOnlyMs).toBe(0);
+    expect(r.workMs).toBeGreaterThan(0);
+  });
+
+  it('a DIFFERENT aiChatId breaks the burst — two agent runs, idle gap excluded', () => {
+    // Run c1 ends 10:05, run c2 starts 10:20 (15m > agentTGap 7m) → two sessions.
+    const rows = [
+      s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:05:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:20:00', { source: 'agent', chat: 'c2', kind: 'agent' }),
+      s('2026-07-04T10:25:00', { source: 'agent', chat: 'c2', kind: 'agent' }),
+    ];
+    const r = computeWorkTime(rows); // default agentTGap = 7m
+    expect(r.sessions).toHaveLength(2);
+    expect(r.sessions.every((x) => x.class === 'agent_only')).toBe(true);
+    // The 15m idle gap between the two runs is NOT counted.
+    const run1 = 5 * MIN + 5 * MIN + 5 * MIN; // pIn + span + pOut
+    expect(r.agentOnlyMs).toBe(2 * run1);
+  });
+
+  it('idle pulse (same/null run) is a full activity sample that continues a burst', () => {
+    const rows = [
+      s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      // idle flush 4m later, null run id → continues the burst, not a new one
+      s('2026-07-04T10:04:00', { source: 'agent', chat: null, kind: 'idle' }),
+      s('2026-07-04T10:08:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+    ];
+    const r = computeWorkTime(rows);
+    expect(r.sessions).toHaveLength(1);
+    // burst span 10:00→10:08 (+pIn/pOut) = 8 + 10 = 18m
+    expect(r.agentOnlyMs).toBe(18 * MIN);
+  });
+
+  it('a USER-sourced idle breaks an agent burst → session is work, not agent_only', () => {
+    // A human supervision idle inherits source=user (aiChatId:null) and must NOT
+    // be swallowed into the agent burst. Δ=3m is within the default agentTGap so
+    // the two samples stay one session — but its class flips to `work`.
+    const rows = [
+      s('2026-07-04T10:00:00', { source: 'agent', chat: 'c1', kind: 'agent' }),
+      s('2026-07-04T10:03:00', { source: 'user', chat: null, kind: 'idle' }),
+    ];
+    const r = computeWorkTime(rows);
+    expect(r.sessions).toHaveLength(1);
+    expect(r.sessions[0].class).toBe('work');
+    expect(r.workMs).toBeGreaterThan(0);
+    // The human idle is NOT captured as agent_only time.
+    expect(r.agentOnlyMs).toBe(0);
+    // Σ over `work` sessions == workMs and Σ over `agent_only` == agentOnlyMs.
+    const sum = (cls: string) =>
+      r.sessions
+        .filter((x) => x.class === cls)
+        .reduce((acc, x) => acc + (x.end - x.start), 0);
+    expect(sum('work')).toBe(r.workMs);
+    expect(sum('agent_only')).toBe(r.agentOnlyMs);
+  });
+
+  it('idle pulse keeps a human writing session visible (not excluded)', () => {
+    const rows = [
+      s('2026-07-04T10:00:00'),
+      s('2026-07-04T10:08:00', { kind: 'idle' }), // pulse within T_gap
+      s('2026-07-04T10:15:00'),
+    ];
+    const r = computeWorkTime(rows);
+    expect(r.sessions).toHaveLength(1);
+    expect(r.sessions[0].class).toBe('work');
+    // span 10:00→10:15 + pIn/pOut = 15 + 10 = 25m
+    expect(r.workMs).toBe(25 * MIN);
+  });
+
+  it('git-source samples are excluded (§10 excludeGit)', () => {
+    const rows = [
+      s('2026-07-04T10:00:00', { source: 'git', kind: 'boundary' }),
+      s('2026-07-04T10:01:00', { source: 'git', kind: 'boundary' }),
+    ];
+    expect(computeWorkTime(rows).workMs).toBe(0);
+    // ...but honoured off:
+    expect(
+      computeWorkTime(rows, { excludeGit: false }).workMs,
+    ).toBeGreaterThan(0);
+  });
+
+  it('rejects an invalid config (tGap < pIn + pOut)', () => {
+    expect(() =>
+      computeWorkTime([s('2026-07-04T10:00:00')], {
+        tGap: 5 * MIN,
+        pIn: 5 * MIN,
+        pOut: 5 * MIN,
+      }),
+    ).toThrow(/tGap/);
+  });
+
+  it('rejects an invalid config (2·agentTGap < pIn + pOut)', () => {
+    // tGap (default 15m) still ≥ pIn+pOut, so only the 2·agentTGap guard trips.
+    // Without it a short session of one class between two of the other could
+    // produce a NON-adjacent cross-class overlap the adjacent-only clip misses.
+    expect(() =>
+      computeWorkTime([s('2026-07-04T10:00:00')], {
+        agentTGap: 2 * MIN,
+        pIn: 5 * MIN,
+        pOut: 5 * MIN,
+      }),
+    ).toThrow(/agentTGap/);
+  });
+
+  // F1 — cross-class double-count. On the DEFAULT config agentTGap (7m) < pIn+pOut
+  // (10m), so a `work` session ending in an agent segment and a nearby separate
+  // `agent_only` run (gap in (7m,10m]) used to produce OVERLAPPING padded
+  // intervals — the same wall-clock counted into BOTH workMs and agentOnlyMs. The
+  // cross-class padding clip must make the two per-class unions disjoint.
+  it('does NOT double-count wall-clock across work/agent_only (§F1)', () => {
+    // user@0s ; agent(chatX)@60s (breaks into a work session with the human) ;
+    // agent(chatY)@560s,590s (a separate agent_only run). Raw gap between the work
+    // session (ends 60s) and the agent run (starts 560s) is 500s ∈ (agentTGap,
+    // pIn+pOut] once padded — the classic overlap window.
+    const rows: TimelineSample[] = [
+      s('2026-07-04T00:00:00'), // user @ 0s
+      s('2026-07-04T00:01:00', { source: 'agent', chat: 'cX', kind: 'agent' }), // @ 60s
+      s('2026-07-04T00:09:20', { source: 'agent', chat: 'cY', kind: 'agent' }), // @ 560s
+      s('2026-07-04T00:09:50', { source: 'agent', chat: 'cY', kind: 'agent' }), // @ 590s
+    ];
+    const r = computeWorkTime(rows); // DEFAULT config
+
+    // Both classes present.
+    expect(r.workMs).toBeGreaterThan(0);
+    expect(r.agentOnlyMs).toBeGreaterThan(0);
+
+    // Per-class metrics are exactly their own union (union, not Σ).
+    expect(r.workMs).toBe(unionMs(ivsOf(r.sessions, 'work')));
+    expect(r.agentOnlyMs).toBe(unionMs(ivsOf(r.sessions, 'agent_only')));
+
+    // The F1 invariant: work-union and agent-union are cross-class-disjoint, so
+    // the union of ALL padded intervals equals workMs + agentOnlyMs (no overlap).
+    // With the clip disabled this fails (union < sum by the 100s overlap).
+    expect(unionMs(ivsOf(r.sessions))).toBe(r.workMs + r.agentOnlyMs);
+  });
+
+  // F1 property/fuzz — random timelines across several timezones must uphold the
+  // work-time invariants. Backs the (corrected) PR claim of a real fuzz test.
+  it('property: random timelines uphold union & cross-class-disjoint invariants', () => {
+    // Deterministic LCG (numerical-recipes constants) so a failure is reproducible.
+    let seed = 0x9e3779b9 >>> 0;
+    const rand = () => {
+      seed = (Math.imul(seed, 1664525) + 1013904223) >>> 0;
+      return seed / 0x100000000;
+    };
+    const pick = <T>(arr: T[]): T => arr[Math.floor(rand() * arr.length)];
+
+    const tzs = [
+      'UTC',
+      'America/New_York',
+      'Europe/Moscow',
+      'Australia/Lord_Howe', // 30-min DST offset — a nasty bucket stress
+    ];
+    const base = Date.UTC(2026, 5, 1, 0, 0, 0); // 2026-06-01Z
+    const chats = ['c1', 'c2', 'c3'];
+
+    for (let iter = 0; iter < 250; iter++) {
+      const tz = pick(tzs);
+      const n = 2 + Math.floor(rand() * 18); // 2..19 rows
+      const rows: TimelineSample[] = [];
+      // Walk time forward by a random inter-sample gap. The gap distribution is
+      // centred on the DANGEROUS band — a bit under to a bit over pIn+pOut (10m)
+      // AND straddling agentTGap (7m) — so adjacent samples routinely split into
+      // separate sessions whose ±P padding would overlap if a class boundary sits
+      // there. Mixing user/agent classes at these gaps reliably manufactures the
+      // work-ending-in-agent → agent_only cross-class boundary F1 is about, plus
+      // dense within-class runs (occasional 0–2m gaps) that exercise the union.
+      let t = base + Math.floor(rand() * 60 * MIN);
+      for (let i = 0; i < n; i++) {
+        const roll = rand();
+        const gap =
+          roll < 0.25
+            ? Math.floor(rand() * 2 * MIN) // dense burst (same-class union)
+            : roll < 0.85
+              ? 5 * MIN + Math.floor(rand() * 8 * MIN) // 5–13m: the split band
+              : 20 * MIN + Math.floor(rand() * 40 * MIN); // long idle → new day-ish
+        t += gap;
+        const iso = new Date(t).toISOString().slice(0, 19); // 'YYYY-MM-DDTHH:MM:SS'
+        const isAgent = rand() < 0.5;
+        rows.push(
+          isAgent
+            ? s(iso, { source: 'agent', chat: pick(chats), kind: 'agent' })
+            : s(iso, { source: 'user', kind: rand() < 0.3 ? 'idle' : 'manual' }),
+        );
+      }
+
+      const r = computeWorkTime(rows); // DEFAULT config
+
+      const workIvs = ivsOf(r.sessions, 'work');
+      const agentIvs = ivsOf(r.sessions, 'agent_only');
+
+      // (1) each metric is exactly its per-class union (catches a union→Σ regress).
+      expect(r.workMs).toBe(unionMs(workIvs));
+      expect(r.agentOnlyMs).toBe(unionMs(agentIvs));
+
+      // (2) NO cross-class overlap: union(all) == workMs + agentOnlyMs (F1).
+      expect(unionMs(ivsOf(r.sessions))).toBe(r.workMs + r.agentOnlyMs);
+
+      // (3) bucket invariant: Σ per-day activeMs == workMs (§6.3).
+      const perDay = bucketByDay(r.sessions, tz);
+      const sumActive = perDay.reduce((a, d) => a + d.activeMs, 0);
+      expect(sumActive).toBe(r.workMs);
+    }
+  });
+});

apps/server/src/core/page/work-time/compute-work-time.ts

@@ -0,0 +1,274 @@
+import {
+  TimelineSample,
+  WorkSession,
+  WorkTimeResult,
+} from './work-time.types';
+import { WorkTimeConfig, resolveWorkTimeConfig } from './work-time.config';
+
+/** A normalized activity sample (one history row), createdAt as epoch-ms. */
+interface NormSample {
+  t: number;
+  isAgent: boolean;
+  aiChatId: string | null;
+  kind: string | null;
+}
+
+/**
+ * A collapsed segment: either a scalar sample (t_start == t_end) or an
+ * agent-burst spanning several agent samples of one run (§5.1). It participates
+ * in sessionization as a single "sample".
+ */
+interface Segment {
+  tStart: number;
+  tEnd: number;
+  isAgent: boolean;
+}
+
+function toMs(v: Date | string | number): number {
+  if (v instanceof Date) return v.getTime();
+  if (typeof v === 'number') return v;
+  return new Date(v).getTime();
+}
+
+/**
+ * Normalize raw rows → sorted, deduped activity samples. `git` is dropped when
+ * configured; every other kind (incl. `idle` — the continuous-work pulse §3) is
+ * a real activity sample. Sort is by createdAt ASC; samples whose timestamps
+ * fall in the same `dedupRoundMs` bucket collapse to one (§9#7: a synchronous
+ * boundary row + the immediate agent snapshot can share a createdAt). A merged
+ * sample is human unless EVERY member is an agent, so supervision never gets
+ * mis-attributed to the agent.
+ */
+function normalize(
+  rows: TimelineSample[],
+  config: WorkTimeConfig,
+): NormSample[] {
+  const samples: NormSample[] = [];
+  for (const row of rows) {
+    const source = row.lastUpdatedSource;
+    if (config.excludeGit && source === 'git') continue;
+    samples.push({
+      t: toMs(row.createdAt),
+      isAgent: source === 'agent',
+      aiChatId: row.lastUpdatedAiChatId ?? null,
+      kind: row.kind ?? null,
+    });
+  }
+  samples.sort((a, b) => a.t - b.t);
+
+  if (config.dedupRoundMs <= 0 || samples.length < 2) return samples;
+
+  const deduped: NormSample[] = [];
+  for (const s of samples) {
+    const prev = deduped[deduped.length - 1];
+    if (prev && s.t - prev.t < config.dedupRoundMs) {
+      // Merge into the previous sample. Human wins the class; keep the earliest
+      // t; keep a non-null aiChatId if either has one (so a bare boundary row
+      // does not erase the run id).
+      prev.isAgent = prev.isAgent && s.isAgent;
+      prev.aiChatId = prev.aiChatId ?? s.aiChatId;
+      // Prefer the more specific kind (a real kind over a null/boundary) only
+      // matters for burst continuation; keep prev.kind (earliest) as-is.
+      continue;
+    }
+    deduped.push({ ...s });
+  }
+  return deduped;
+}
+
+/**
+ * Collapse consecutive same-run agent samples into one burst segment (§5.1) so a
+ * dense burst (8 snapshots in 7 minutes) contributes its wall-clock, not a count
+ * × block. A burst is broken by any sample NOT continuing the same aiChatId
+ * agent run: a non-agent sample, a `boundary` (actor transition), or a DIFFERENT
+ * aiChatId. Only an AGENT-sourced `idle` pulse with the SAME or a null aiChatId
+ * continues the burst (its label lags the real edit ≤ maxWait, well within
+ * rounding); a user-sourced `idle` (a human supervision pulse) breaks it.
+ */
+function collapse(samples: NormSample[], config: WorkTimeConfig): Segment[] {
+  const segments: Segment[] = [];
+  let burst: { chatId: string | null; tStart: number; tEnd: number } | null =
+    null;
+
+  const flush = () => {
+    if (!burst) return;
+    let tEnd = burst.tEnd;
+    if (config.burstCapMs != null && tEnd - burst.tStart > config.burstCapMs) {
+      tEnd = burst.tStart + config.burstCapMs;
+    }
+    segments.push({ tStart: burst.tStart, tEnd, isAgent: true });
+    burst = null;
+  };
+
+  for (const s of samples) {
+    // An agent-sourced idle pulse continues the current agent burst (same or
+    // null run id). A user-sourced idle (human supervision) must NOT be swallowed
+    // here — it falls through to the human branch so the session flips to `work`.
+    if (
+      burst &&
+      s.kind === 'idle' &&
+      s.isAgent &&
+      (s.aiChatId === burst.chatId || s.aiChatId == null)
+    ) {
+      burst.tEnd = s.t;
+      continue;
+    }
+    if (s.isAgent && s.kind !== 'boundary') {
+      if (burst && burst.chatId === s.aiChatId) {
+        burst.tEnd = s.t;
+      } else {
+        flush();
+        burst = { chatId: s.aiChatId, tStart: s.t, tEnd: s.t };
+      }
+      continue;
+    }
+    // A human sample, a boundary, or an agent-boundary: breaks the burst and is
+    // itself a zero-width segment (its class follows its own source).
+    flush();
+    segments.push({ tStart: s.t, tEnd: s.t, isAgent: s.isAgent });
+  }
+  flush();
+  return segments;
+}
+
+function gapThreshold(
+  a: Segment,
+  b: Segment,
+  config: WorkTimeConfig,
+): number {
+  return a.isAgent && b.isAgent ? config.agentTGap : config.tGap;
+}
+
+/** Merge intervals; overlapping OR touching intervals are unioned. */
+function unionDuration(intervals: Array<[number, number]>): number {
+  if (intervals.length === 0) return 0;
+  const sorted = [...intervals].sort((a, b) => a[0] - b[0]);
+  let total = 0;
+  let [curStart, curEnd] = sorted[0];
+  for (let i = 1; i < sorted.length; i++) {
+    const [s, e] = sorted[i];
+    if (s <= curEnd) {
+      if (e > curEnd) curEnd = e;
+    } else {
+      total += curEnd - curStart;
+      curStart = s;
+      curEnd = e;
+    }
+  }
+  total += curEnd - curStart;
+  return total;
+}
+
+/**
+ * #395 core — estimate time worked on a page from its history timeline (§5).
+ * Pure and deterministic: no DB, no clock, no I/O.
+ *
+ * Pipeline: normalize+dedup → collapse agent bursts → ONE sessionization pass
+ * over all segments (threshold depends on the pair: both-agent → agentTGap, else
+ * tGap; the last session is ALWAYS closed after the loop) → class per finished
+ * session (all-agent → agent_only, else work) → pad each session (multi-sample
+ * → [first−P_in, last+P_out]; lone scalar → [t−P_single, t]) → clip padding of
+ * adjacent DIFFERENT-class sessions at the raw-gap midpoint (so work/agent_only
+ * never overlap) → metrics are the union wall-clock within each class (union, not
+ * Σ, so overlaps never double, and cross-class-disjoint by the clip above).
+ */
+export function computeWorkTime(
+  rows: TimelineSample[],
+  config?: Partial<WorkTimeConfig>,
+): WorkTimeResult {
+  const cfg = resolveWorkTimeConfig(config);
+  const samples = normalize(rows, cfg);
+  const segments = collapse(samples, cfg);
+
+  // Sessionize — one pass over ALL segments.
+  const rawSessions: Segment[][] = [];
+  let cur: Segment[] | null = null;
+  for (const seg of segments) {
+    if (cur == null) {
+      cur = [seg];
+    } else {
+      const last = cur[cur.length - 1];
+      if (seg.tStart - last.tEnd <= gapThreshold(last, seg, cfg)) {
+        cur.push(seg);
+      } else {
+        rawSessions.push(cur);
+        cur = [seg];
+      }
+    }
+  }
+  if (cur != null) rawSessions.push(cur); // MUST close the last session (§5, §9#1)
+
+  // A finished session with BOTH its raw (unpadded) span and its padded bounds.
+  // `rawSessions` are already in ascending time order, so `built` is too.
+  interface BuiltSession {
+    rawStart: number;
+    rawEnd: number;
+    padStart: number;
+    padEnd: number;
+    cls: WorkSession['class'];
+  }
+  const built: BuiltSession[] = [];
+
+  for (const segs of rawSessions) {
+    const first = segs[0];
+    const last = segs[segs.length - 1];
+    const cls = segs.every((s) => s.isAgent) ? 'agent_only' : 'work';
+
+    let padStart: number;
+    let padEnd: number;
+    if (segs.length === 1 && first.tStart === first.tEnd) {
+      // Lone single-instant session (one scalar, or a one-snapshot agent run):
+      // pre-roll only, no invented "future" work (§5).
+      padStart = first.tStart - cfg.pSingle;
+      padEnd = first.tStart;
+    } else {
+      padStart = first.tStart - cfg.pIn;
+      padEnd = last.tEnd + cfg.pOut;
+    }
+
+    built.push({
+      rawStart: first.tStart,
+      rawEnd: last.tEnd,
+      padStart,
+      padEnd,
+      cls,
+    });
+  }
+
+  // Clip cross-class padding so a `work` and an `agent_only` session that abut
+  // never claim the same wall-clock. For each ADJACENT pair of DIFFERENT classes,
+  // cap the earlier session's trailing pad and the later session's leading pad at
+  // the MIDPOINT of the raw (unpadded) inactivity gap between them: the earlier
+  // padded interval then ends ≤ midpoint and the later one starts ≥ midpoint, so
+  // the two are disjoint (they touch at most at the midpoint). This makes the
+  // per-class unions (workMs / agentOnlyMs) cross-class-disjoint BY CONSTRUCTION
+  // — closing the double-count where a work session ending in an agent segment
+  // and a nearby agent_only session (gap in (agentTGap, pIn+pOut]) overlapped and
+  // were counted into both metrics (§5, §9). Within-class adjacency is left
+  // untouched: `unionDuration` already dedups it, and clipping there could perturb
+  // the per-class metric value.
+  for (let i = 1; i < built.length; i++) {
+    const a = built[i - 1];
+    const b = built[i];
+    if (a.cls === b.cls) continue;
+    const midpoint = (a.rawEnd + b.rawStart) / 2;
+    if (a.padEnd > midpoint) a.padEnd = midpoint;
+    if (b.padStart < midpoint) b.padStart = midpoint;
+  }
+
+  const sessions: WorkSession[] = [];
+  const workIvs: Array<[number, number]> = [];
+  const agentIvs: Array<[number, number]> = [];
+  for (const s of built) {
+    sessions.push({ start: s.padStart, end: s.padEnd, class: s.cls });
+    (s.cls === 'work' ? workIvs : agentIvs).push([s.padStart, s.padEnd]);
+  }
+
+  sessions.sort((a, b) => a.start - b.start);
+
+  return {
+    workMs: unionDuration(workIvs),
+    agentOnlyMs: unionDuration(agentIvs),
+    sessions,
+  };
+}

apps/server/src/core/page/work-time/index.ts

@@ -0,0 +1,15 @@
+export { computeWorkTime } from './compute-work-time';
+export { bucketByDay, zonedDayStart } from './bucket-by-day';
+export {
+  DEFAULT_WORK_TIME_CONFIG,
+  resolveWorkTimeConfig,
+} from './work-time.config';
+export type { WorkTimeConfig } from './work-time.config';
+export type {
+  TimelineSample,
+  WorkSession,
+  WorkTimeResult,
+  SessionClass,
+  DayWindow,
+  PerDay,
+} from './work-time.types';

apps/server/src/core/page/work-time/work-time.config.ts

@@ -0,0 +1,102 @@
+import {
+  IDLE_MAX_WAIT_USER,
+  IDLE_MAX_WAIT_AGENT,
+} from '../../../collaboration/constants';
+
+/**
+ * #395 — tunables for the work-time estimate (§10). Defaults are calibrated off
+ * #374's idle-pulse ceilings: after #374 a continuous editing session leaves a
+ * history row at least every ~IDLE_MAX_WAIT (10m user / 5m agent), so a gap
+ * WIDER than that ceiling contains un-pulsed idle time = (partial) inactivity.
+ * `tGap` therefore sits a little above the user ceiling, `agentTGap` a little
+ * above the agent ceiling — a gap within the threshold is pulse-backed and
+ * counts as work.
+ */
+export interface WorkTimeConfig {
+  /** user inactivity timeout: gap ≤ tGap between samples = continuous work. */
+  tGap: number;
+  /** timeout for a pair of consecutive agent samples (tighter than tGap). */
+  agentTGap: number;
+  /** pre-roll padding for a multi-sample session (work began before sample 1). */
+  pIn: number;
+  /** post-roll padding for a multi-sample session (work continued after last). */
+  pOut: number;
+  /** block for a lone single-sample session (pre-roll only, no invented future). */
+  pSingle: number;
+  /** drop `git`-source samples (they are not human/agent article work). */
+  excludeGit: boolean;
+  /** optional cap on one collapsed agent-burst segment's wall-clock (§9#3). */
+  burstCapMs?: number;
+  /** samples whose createdAt round to the same bucket dedup to one (§9#7). */
+  dedupRoundMs: number;
+}
+
+export const DEFAULT_WORK_TIME_CONFIG: WorkTimeConfig = {
+  // ~15m: IDLE_MAX_WAIT_USER (10m) + headroom. Empirically backcast on a real
+  // 307-snapshot article (≈24h at 15m matched the owner's estimate; 30/45m
+  // over-counted). See #395 §10.
+  tGap: 15 * 60 * 1000,
+  // ~7m: IDLE_MAX_WAIT_AGENT (5m) + headroom.
+  agentTGap: 7 * 60 * 1000,
+  pIn: 5 * 60 * 1000,
+  pOut: 5 * 60 * 1000,
+  pSingle: 2 * 60 * 1000,
+  excludeGit: true,
+  burstCapMs: undefined,
+  dedupRoundMs: 1000,
+};
+
+// Compile-time cross-check that the defaults really are pulse-anchored — if a
+// future edit moves the #374 ceilings, this reminds us to re-calibrate.
+void IDLE_MAX_WAIT_USER;
+void IDLE_MAX_WAIT_AGENT;
+
+/**
+ * Fill a partial config with defaults and validate it. Cross-class metric
+ * disjointness is guaranteed jointly by `computeWorkTime`'s adjacent-pair padding
+ * clip (it caps the padding of adjacent DIFFERENT-class sessions at the raw-gap
+ * midpoint) AND the two bounds enforced below (§5):
+ *   - `tGap ≥ pIn + pOut`: a session's own padding never exceeds its inactivity
+ *     window.
+ *   - `2·agentTGap ≥ pIn + pOut`: makes the adjacent-only clip provably COMPLETE.
+ *     A NON-adjacent (i, i+2) cross-class overlap could only arise from two
+ *     same-class sessions separated by a full intervening session of the other
+ *     class; that separation spans at least two inter-session gaps, each strictly
+ *     `> agentTGap`, so it is `> 2·agentTGap`. Requiring `2·agentTGap ≥ pIn + pOut`
+ *     means even the widest padded reach (pIn + pOut) cannot bridge it — so the
+ *     only cross-class overlaps possible are between ADJACENT sessions, which the
+ *     clip handles. `workMs`/`agentOnlyMs` are therefore disjoint by construction.
+ */
+export function resolveWorkTimeConfig(
+  partial?: Partial<WorkTimeConfig>,
+): WorkTimeConfig {
+  const config = { ...DEFAULT_WORK_TIME_CONFIG, ...(partial ?? {}) };
+
+  for (const key of [
+    'tGap',
+    'agentTGap',
+    'pIn',
+    'pOut',
+    'pSingle',
+    'dedupRoundMs',
+  ] as const) {
+    const value = config[key];
+    if (!Number.isFinite(value) || value < 0) {
+      throw new Error(`work-time config: ${key} must be a non-negative number`);
+    }
+  }
+  if (config.burstCapMs != null && config.burstCapMs <= 0) {
+    throw new Error('work-time config: burstCapMs must be > 0 when set');
+  }
+  if (config.tGap < config.pIn + config.pOut) {
+    throw new Error(
+      "work-time config: tGap must be ≥ pIn + pOut (a session's padding may not exceed its inactivity window)",
+    );
+  }
+  if (2 * config.agentTGap < config.pIn + config.pOut) {
+    throw new Error(
+      'work-time config: 2·agentTGap must be ≥ pIn + pOut (so non-adjacent cross-class padding cannot overlap)',
+    );
+  }
+  return config;
+}

apps/server/src/core/page/work-time/work-time.types.ts

@@ -0,0 +1,73 @@
+/**
+ * #395 — "time worked on an article" domain types.
+ *
+ * The estimate is built by sessionizing a page's `page_history` timeline on
+ * inactivity gaps (WakaTime-style), NOT by taking the span between the first and
+ * last edit (which over-counts sleep / lunch / idle days). See the design doc in
+ * issue #395 §5–§6.3 for the normative algorithm.
+ */
+
+/**
+ * A single `page_history` row projected for the work-time computation — the
+ * cheap columns only (no `content`). Produced by
+ * `PageHistoryRepo.findTimelineByPageId`. `createdAt` is whatever the DB driver
+ * hands back (Date); the pure core normalizes it to epoch-ms itself so it stays
+ * deterministic and DB-free.
+ */
+export interface TimelineSample {
+  createdAt: Date | string | number;
+  lastUpdatedById: string | null;
+  /** 'user' | 'agent' | 'git' | null (legacy autosave = human). */
+  lastUpdatedSource: string | null;
+  lastUpdatedAiChatId: string | null;
+  /** #370 tier: 'manual' | 'agent' | 'idle' | 'boundary' | null (legacy). */
+  kind: string | null;
+}
+
+/** A finished session's class (§5.1). */
+export type SessionClass = 'work' | 'agent_only';
+
+/**
+ * A finished session: absolute wall-clock bounds already padded with P_in/P_out
+ * (multi-sample) or P_single (single scalar), plus its class. This is enough for
+ * both the metrics and the per-day punch-card colouring.
+ */
+export interface WorkSession {
+  /** epoch-ms, inclusive lower bound (already P-padded). */
+  start: number;
+  /** epoch-ms, exclusive upper bound (already P-padded). */
+  end: number;
+  class: SessionClass;
+}
+
+/** Output of {@link computeWorkTime}. */
+export interface WorkTimeResult {
+  /** union wall-clock of `work` sessions, ms (the headline metric). */
+  workMs: number;
+  /** union wall-clock of `agent_only` sessions, ms (secondary). */
+  agentOnlyMs: number;
+  sessions: WorkSession[];
+}
+
+/** One activity window inside a calendar day (already clipped to the day). */
+export interface DayWindow {
+  /** epoch-ms. */
+  start: number;
+  /** epoch-ms. */
+  end: number;
+  class: SessionClass;
+}
+
+/** One calendar day of the punch-card (§6.3). */
+export interface PerDay {
+  /** epoch-ms of the local-midnight day start in the requested tz. */
+  day: number;
+  /** 'YYYY-MM-DD' in the requested tz — stable, tz-independent label. */
+  dayISO: string;
+  /** Σ of `work` windows this day, ms. Σ over days == workMs (invariant §6.3). */
+  activeMs: number;
+  /** Σ of `agent_only` windows this day, ms (drawn, NOT in activeMs). */
+  agentMs: number;
+  /** both classes, clipped to the day, sorted by start (for drawing). */
+  windows: DayWindow[];
+}

apps/server/src/core/search/dto/search-response.dto.ts

@@ -1,33 +1,76 @@
 import { Space } from '@docmost/db/types/entity.types';
 
-export class SearchResponseDto {
+// #529 A7 — the single per-hit SUPERSET returned by the unified search engine.
+// The web-UI reads id/highlight/icon/space/title/…; the MCP agent maps id→pageId
+// and reads snippet/score/path. `rank`/`highlight` are null for substring-only
+// hits (the web already falls back). Nothing the legacy web response carried is
+// dropped.
+export class SearchResultDto {
   id: string;
-  title: string;
+  // Alias of `id` for the MCP layer (it addresses pages by pageId).
+  pageId: string;
+  slugId: string;
   icon: string;
-  parentPageId: string;
+  title: string;
+  space?: Partial<Space>;
   creatorId: string;
-  rank: number;
-  highlight: string;
   createdAt: Date;
   updatedAt: Date;
-  space: Partial<Space>;
+  // ts_rank_cd of the FTS branch; null for substring-only hits.
+  rank: number | null;
+  // ts_headline marked HTML; null for substring-only hits.
+  highlight: string | null;
+  // Plain windowed snippet around the match (empty for titleOnly).
+  snippet: string;
+  // Ancestor titles root → direct parent ([] for a root page).
+  path: string[];
+  // Per-response ordering proxy (falls back to rank).
+  score: number;
+  // Which fields matched: 'title' and/or 'text'.
+  matchedFields: string[];
+  // Which parsed positive/required terms this hit matched.
+  matchedTerms: string[];
 }
 
-// Response shape for the opt-in agent-lookup mode (#443, `substring: true`).
-// Additive to the FTS response: carries the location (`path`), a windowed
-// `snippet` around the first match and a per-response sort `score`. The MCP
-// layer maps `id → pageId`; `slugId` is never exposed.
-export class SearchLookupResponseDto {
-  id: string;
-  slugId: string;
-  title: string;
-  parentPageId: string | null;
-  // Ancestor titles from the space root down to the direct parent; [] for a
-  // root page.
-  path: string[];
-  // ~300–500 chars around the first match (or a leading text window / extended
-  // ts_headline fallback).
-  snippet: string;
-  // 0..1 float, meaningful ONLY for sorting within one response.
-  score: number;
+// #530 Phase B — the semantic (vector) layer's per-request status.
+//  - `state`: 'full' when a provider resolved and the vector arm ran this
+//    request; 'off' otherwise (kill-switch, no provider, or a degrade).
+//  - `available`: whether the vector arm actually contributed this request.
+//  - `reason`: why the arm did NOT run — 'no-provider' (no embedding provider
+//    resolved) or 'degraded' (sidecar down / query embed timed out).
+//  - `indexed`/`total`: reserved for PR-2 coverage reporting (unused in PR-1).
+export class SearchSemanticDto {
+  state: 'full' | 'off';
+  available: boolean;
+  reason?: 'no-provider' | 'degraded';
+  indexed?: number;
+  total?: number;
+}
+
+// The paginated envelope (A5). `total` is the permission-filtered count of the
+// candidate UNION (lexical ∪ vector top-N, fail-closed) — a deliberate, #530
+// documented change from Phase A's exact-lexical count (on a semantic degrade it
+// falls back to exactly that lexical count). `hasMore` is true when more results
+// exist WITHIN the fusion window; `truncatedAtCap` signals the match set exceeded
+// CANDIDATE_CAP and the tail is unreachable by pagination.
+export class SearchResponseDto {
+  items: SearchResultDto[];
+  total: number;
+  hasMore: boolean;
+  truncatedAtCap: boolean;
+  offset: number;
+  query: {
+    raw: string;
+    parsed: {
+      positive: string[];
+      required: string[];
+      excluded: string[];
+      reason?: string;
+    };
+    mode: 'or' | 'and';
+    match: string;
+  };
+  // Absent on the early-exit paths (empty/garbage query); present once search
+  // actually runs. Optional so those short-circuit responses stay valid.
+  semantic?: SearchSemanticDto;
 }

apps/server/src/core/search/dto/search.dto.ts

@@ -1,16 +1,30 @@
 import {
   IsBoolean,
+  IsIn,
   IsNotEmpty,
   IsNumber,
   IsOptional,
   IsString,
+  MaxLength,
 } from 'class-validator';
 
 export class SearchDTO {
+  // Defense-in-depth cap on the raw query length. The real stack-depth bound is
+  // the parser's MAX_PARSED_TERMS term cap (see search-query-parser.ts); this
+  // just rejects absurd payloads early. 10k chars still comfortably holds any
+  // legitimate query.
   @IsNotEmpty()
   @IsString()
+  @MaxLength(10000)
   query: string;
 
+  // #529 A3 — match mode. `auto` (default) routes identifier-like terms
+  // (10.31.41, esp32, WB-MGE-30D86B) to the substring/trigram branch and words
+  // to full-text; `word`/`prefix`/`substring` are explicit overrides.
+  @IsOptional()
+  @IsIn(['auto', 'word', 'prefix', 'substring'])
+  match?: 'auto' | 'word' | 'prefix' | 'substring';
+
   @IsOptional()
   @IsString()
   spaceId: string;
@@ -33,15 +47,19 @@ export class SearchDTO {
 
   // --- Opt-in agent-lookup mode (#443). ------------------------------------
   // These fields are ADDITIVE and default-off: a web client that sends none of
-  // them gets byte-identical FTS behaviour and result shape. They are only read
-  // by the substring/path/snippet code path in SearchService.searchPage.
+  // them gets byte-identical FTS behaviour and result shape. In the unified #529
+  // engine, `parentPageId` and `titleOnly` are read by SearchService.searchPage
+  // (subtree scoping and title-only matching, respectively). `substring` is NOT
+  // read by the native driver — it is accepted-but-ignored, kept only for
+  // back-compat with the upstream lookup request shape.
   //
   // NOTE (standalone stdio vs stock upstream): stock upstream validates this DTO
   // with `whitelist: true`, so an older server silently strips these unknown
   // fields and the request degrades gracefully to the plain FTS behaviour.
 
-  // Enables the hybrid substring branch (title + text_content LIKE) merged with
-  // the existing FTS branch, plus tiered ranking, path and windowed snippet.
+  // Accepted-but-ignored by the #529 native driver (kept for upstream lookup
+  // back-compat). The unified engine ALWAYS runs the hybrid FTS + substring/
+  // trigram branches with tiered ranking, so this flag no longer toggles anything.
   @IsOptional()
   @IsBoolean()
   substring?: boolean;

apps/server/src/core/search/search-query-parser.spec.ts

@@ -0,0 +1,168 @@
+import {
+  parseSearchQuery,
+  hasPositiveRecall,
+  MAX_PARSED_TERMS,
+} from './search-query-parser';
+
+describe('parseSearchQuery — tokenization & operators (A2)', () => {
+  it('splits on whitespace into positive terms (OR recall)', () => {
+    const p = parseSearchQuery('стамбул роснефть');
+    expect(p.positive.map((t) => t.text)).toEqual(['стамбул', 'роснефть']);
+    expect(p.required).toEqual([]);
+    expect(p.excluded).toEqual([]);
+    expect(p.mode).toBe('or');
+  });
+
+  it('treats +term as required and -term as excluded', () => {
+    const p = parseSearchQuery('+кофейня -архив');
+    expect(p.required.map((t) => t.text)).toEqual(['кофейня']);
+    expect(p.excluded.map((t) => t.text)).toEqual(['архив']);
+    expect(p.positive).toEqual([]);
+  });
+
+  it('keeps a leading-operator-free hyphen/dot/colon token as ONE literal term', () => {
+    // WB-MGE-30D86B, 10.0.12.5, a:b — internal -,.,: are literal, one term each.
+    expect(parseSearchQuery('WB-MGE-30D86B').positive[0].text).toBe(
+      'WB-MGE-30D86B',
+    );
+    expect(parseSearchQuery('10.0.12.5').positive[0].text).toBe('10.0.12.5');
+    expect(parseSearchQuery('host:8080').positive[0].text).toBe('host:8080');
+  });
+
+  it('only a LEADING +/- is an operator; -архив excludes архив', () => {
+    const p = parseSearchQuery('-архив');
+    expect(p.excluded.map((t) => t.text)).toEqual(['архив']);
+    expect(p.positive).toEqual([]);
+  });
+
+  it('drops a bare "-" / "+" and an all-operator remainder', () => {
+    const p = parseSearchQuery('- + foo -- ++');
+    expect(p.positive.map((t) => t.text)).toEqual(['foo']);
+    expect(p.required).toEqual([]);
+    expect(p.excluded).toEqual([]);
+  });
+
+  it('parses a quoted phrase as one adjacency term', () => {
+    const p = parseSearchQuery('"воздушный шар" кофе');
+    expect(p.positive[0]).toMatchObject({ text: 'воздушный шар', branch: 'phrase' });
+    expect(p.positive[1].text).toBe('кофе');
+  });
+
+  it('applies +/- to a phrase', () => {
+    const req = parseSearchQuery('+"воздушный шар"');
+    expect(req.required[0]).toMatchObject({ text: 'воздушный шар', branch: 'phrase' });
+    const exc = parseSearchQuery('-"воздушный шар"');
+    expect(exc.excluded[0]).toMatchObject({ text: 'воздушный шар', branch: 'phrase' });
+  });
+
+  it('drops an unbalanced quote token', () => {
+    const p = parseSearchQuery('kafka "unclosed here');
+    expect(p.positive.map((t) => t.text)).toEqual(['kafka']);
+  });
+
+  it('strips tsquery metacharacters from a bare FTS term (no 500)', () => {
+    const p = parseSearchQuery('foo|bar');
+    // `|` is not an identifier signal → FTS branch; the metachar is stripped so
+    // the term becomes the two words that survive.
+    expect(p.positive[0].branch === 'fts' || p.positive[0].branch === 'ftsPrefix').toBe(true);
+    expect(p.positive[0].text).toBe('foo bar');
+  });
+});
+
+describe('parseSearchQuery — match=auto classification (A3)', () => {
+  it('routes identifier-like terms to the substring branch', () => {
+    expect(parseSearchQuery('10.31.41').positive[0].branch).toBe('substring');
+    expect(parseSearchQuery('esp32').positive[0].branch).toBe('substring');
+    expect(parseSearchQuery('WB-MGE-30D86B').positive[0].branch).toBe('substring');
+  });
+
+  it('routes purely-alphabetic words to the FTS (prefix) branch', () => {
+    expect(parseSearchQuery('печат').positive[0].branch).toBe('ftsPrefix');
+    expect(parseSearchQuery('ресторан').positive[0].branch).toBe('ftsPrefix');
+  });
+
+  it('explicit match overrides: word / prefix / substring', () => {
+    expect(parseSearchQuery('печат', { match: 'word' }).positive[0].branch).toBe('fts');
+    expect(parseSearchQuery('печат', { match: 'prefix' }).positive[0].branch).toBe(
+      'ftsPrefix',
+    );
+    expect(parseSearchQuery('печат', { match: 'substring' }).positive[0].branch).toBe(
+      'substring',
+    );
+  });
+});
+
+describe('parseSearchQuery — reasons & recall', () => {
+  it('only-negation yields reason only-negation and no positive recall', () => {
+    const p = parseSearchQuery('-архив');
+    expect(p.reason).toBe('only-negation');
+    expect(hasPositiveRecall(p)).toBe(false);
+  });
+
+  it('empty / whitespace / garbage yields reason empty', () => {
+    expect(parseSearchQuery('').reason).toBe('empty');
+    expect(parseSearchQuery('   ').reason).toBe('empty');
+    // A bare operator drops to nothing → empty (no exclusion survived).
+    expect(parseSearchQuery('+ -').reason).toBe('empty');
+  });
+
+  it('a required term alone IS positive recall (no reason)', () => {
+    const p = parseSearchQuery('+кофейня');
+    expect(p.reason).toBeUndefined();
+    expect(hasPositiveRecall(p)).toBe(true);
+  });
+
+  it('mode flag flows through', () => {
+    expect(parseSearchQuery('a b', { mode: 'and' }).mode).toBe('and');
+    expect(parseSearchQuery('a b').mode).toBe('or');
+  });
+});
+
+describe('parseSearchQuery — term cap (stack-depth guard)', () => {
+  it('caps the total parsed terms at MAX_PARSED_TERMS without throwing', () => {
+    // A pasted text block: far more words than the cap. The parser must bound the
+    // SQL tsquery nesting depth (else Postgres blows its stack → HTTP 500).
+    const words = Array.from({ length: 5000 }, (_, i) => `w${i}`);
+    let p!: ReturnType<typeof parseSearchQuery>;
+    expect(() => {
+      p = parseSearchQuery(words.join(' '));
+    }).not.toThrow();
+    const total = p.positive.length + p.required.length + p.excluded.length;
+    expect(total).toBe(MAX_PARSED_TERMS);
+    // Stable order: the FIRST cap terms are kept.
+    expect(p.positive.slice(0, 3).map((t) => t.text)).toEqual(['w0', 'w1', 'w2']);
+    expect(p.positive[MAX_PARSED_TERMS - 1].text).toBe(`w${MAX_PARSED_TERMS - 1}`);
+  });
+
+  it('counts positive + required + excluded together toward the cap', () => {
+    // Interleave operators so overflow can fall on any bucket. Total must still
+    // never exceed the cap, and the first cap terms (in stable order) win.
+    const tokens: string[] = [];
+    for (let i = 0; i < 200; i++) {
+      const op = i % 3 === 0 ? '+' : i % 3 === 1 ? '-' : '';
+      tokens.push(`${op}t${i}`);
+    }
+    const p = parseSearchQuery(tokens.join(' '));
+    const total = p.positive.length + p.required.length + p.excluded.length;
+    expect(total).toBe(MAX_PARSED_TERMS);
+    // t0 (+, required) and t1 (-, excluded) and t2 (bare, positive) are all within
+    // the first cap tokens → each bucket got its leading terms.
+    expect(p.required[0].text).toBe('t0');
+    expect(p.excluded[0].text).toBe('t1');
+    expect(p.positive[0].text).toBe('t2');
+  });
+
+  it('leaves a normal (<= cap) query completely unchanged', () => {
+    const p = parseSearchQuery('+кофейня -архив "воздушный шар" ресторан 10.31.41');
+    expect(p.required.map((t) => t.text)).toEqual(['кофейня']);
+    expect(p.excluded.map((t) => t.text)).toEqual(['архив']);
+    expect(p.positive.map((t) => t.text)).toEqual([
+      'воздушный шар',
+      'ресторан',
+      '10.31.41',
+    ]);
+    // Phrase / substring branch handling survives under the cap.
+    expect(p.positive[0].branch).toBe('phrase');
+    expect(p.positive[2].branch).toBe('substring');
+  });
+});

apps/server/src/core/search/search-query-parser.ts

@@ -0,0 +1,242 @@
+// #529 Phase A — server-side query parser (the SINGLE source of query semantics).
+//
+// Clients (web-UI, MCP agent, public share) send a RAW query string plus flags
+// (`match`, `mode`); ALL query interpretation happens HERE, so every consumer
+// gets identical operator/phrase/morphology behaviour. The parser is PURE (no
+// SQL, no DB) so it is exhaustively unit-testable; SearchService turns the parsed
+// AST into a parameterized tsquery/predicate tree (never string-concatenated SQL).
+//
+// Grammar (A2):
+//   - Whitespace splits tokens, but a double-quoted run is ONE token ("a b" is a
+//     phrase). An unbalanced quote is dropped.
+//   - A token is an OPERATOR token only when it STARTS with `+` or `-`, the
+//     remainder is non-empty, and the remainder is not itself only operators.
+//     `+`/`-` inside a token (`WB-MGE-30D86B`, `10.0.12.5`, `a:b`) is a LITERAL —
+//     the token is a single term. A bare `-`/`+` is dropped.
+//   - `"phrase"` → phrase term; `+"phrase"`/`-"phrase"` apply the operator to it.
+//   - Positive terms (bare + phrase, no operator) form the OR recall set.
+//     `+term` is a REQUIRED predicate, `-term` an EXCLUDED predicate (A2): both
+//     are applied in SQL WHERE against the whole candidate set, not folded into
+//     the positive tsquery.
+//   - Only-negation (no positive term) short-circuits to an empty result with
+//     reason `only-negation` (never runs a costly NOT-scan).
+
+export type SearchMatchMode = 'auto' | 'word' | 'prefix' | 'substring';
+export type SearchBooleanMode = 'or' | 'and';
+
+// Hard cap on the total number of parsed terms (positive + required + excluded).
+// SearchService folds each FTS term into a LEFT-NESTED SQL tsquery expression
+// `(((t1)||(t2))||(t3))…`, embedded several times per query — so nesting depth
+// grows with the term count. A pasted text block (thousands of words) would nest
+// deep enough to blow Postgres' `stack depth limit` → an ERROR → HTTP 500 for the
+// caller. Capping in the parser bounds that depth for EVERY consumer (web / MCP /
+// share), since they all route through here. 64 comfortably covers any real query
+// while keeping the SQL nesting shallow. Overflow terms (beyond the first 64, in
+// stable order) are dropped rather than throwing.
+export const MAX_PARSED_TERMS = 64;
+
+// How a single term is matched against the index.
+//   - 'fts'       : full-text lexeme, exact (no trailing prefix).
+//   - 'ftsPrefix' : full-text lexeme with a `:*` prefix match.
+//   - 'phrase'    : an adjacency phrase (phraseto_tsquery).
+//   - 'substring' : a literal LOWER(f_unaccent(col)) LIKE '%needle%' branch
+//                   (identifiers the tokenizer mangles: IPs, hostnames, IDs).
+export type SearchTermBranch = 'fts' | 'ftsPrefix' | 'phrase' | 'substring';
+
+export interface ParsedTerm {
+  // The user-visible term text, operator stripped, quotes removed. This is what
+  // `matchedTerms` echoes back per hit.
+  text: string;
+  branch: SearchTermBranch;
+}
+
+export interface ParsedQuery {
+  raw: string;
+  // OR-recall set (bare + phrase terms with no operator).
+  positive: ParsedTerm[];
+  // AND predicates (`+term`) — the candidate MUST match each of these.
+  required: ParsedTerm[];
+  // NOT predicates (`-term`) — the candidate must match NONE of these.
+  excluded: ParsedTerm[];
+  mode: SearchBooleanMode;
+  // Set only when the query yields no positive recall: 'empty' (nothing usable)
+  // or 'only-negation' (there were exclusions but no positive term).
+  reason?: 'empty' | 'only-negation';
+}
+
+interface RawToken {
+  op: '' | '+' | '-';
+  kind: 'word' | 'phrase';
+  text: string;
+}
+
+// tsquery metacharacters that must never reach to_tsquery from a bare term — they
+// are what turned adversarial input into a 500 before (#139). Stripped for the FTS
+// branch; the substring branch keeps them (they are literal there).
+const TSQUERY_META = /[:&|!()*<>\\]+/g;
+
+// A term is "identifier-like" when it carries a digit or one of . _ : / - AND is
+// not purely alphabetic (letters only). Such tokens (10.31.41, esp32,
+// WB-MGE-30D86B) are mangled by the FTS tokenizer, so `match: auto` routes them
+// to the substring branch. A purely-alphabetic word (печат, ресторан) stays FTS.
+const IDENTIFIER_SIGNAL = /[0-9._:/\\-]/;
+const PURELY_ALPHA = /^\p{L}+$/u;
+
+function isIdentifierLike(text: string): boolean {
+  return IDENTIFIER_SIGNAL.test(text) && !PURELY_ALPHA.test(text);
+}
+
+// Clean a bare term for the FTS branch: NFC-normalize, drop tsquery metacharacters
+// and collapse whitespace. Returns '' when nothing usable remains.
+export function cleanFtsLexeme(raw: string): string {
+  return (raw ?? '')
+    .normalize('NFC')
+    .replace(TSQUERY_META, ' ')
+    .replace(/\s+/g, ' ')
+    .trim();
+}
+
+// Split a raw query into tokens, honouring double quotes and a single leading
+// +/- operator. Unbalanced quotes and bare operators are dropped.
+function tokenize(raw: string): RawToken[] {
+  const tokens: RawToken[] = [];
+  const s = raw ?? '';
+  let i = 0;
+  const n = s.length;
+
+  const isSpace = (c: string) => /\s/.test(c);
+
+  while (i < n) {
+    // Skip leading whitespace.
+    while (i < n && isSpace(s[i])) i++;
+    if (i >= n) break;
+
+    let op: '' | '+' | '-' = '';
+    // A single leading +/- is a tentative operator. Only ONE leading operator is
+    // consumed; a second (`--x`) leaves `-x` as the remainder (a literal dash).
+    if (s[i] === '+' || s[i] === '-') {
+      op = s[i] as '+' | '-';
+      i++;
+    }
+
+    if (i < n && s[i] === '"') {
+      // Quoted phrase: read until the closing quote.
+      const close = s.indexOf('"', i + 1);
+      if (close === -1) {
+        // Unbalanced quote → drop this token and everything the open quote would
+        // have consumed (the rest of the string).
+        break;
+      }
+      const phrase = s.slice(i + 1, close);
+      i = close + 1;
+      if (phrase.trim().length > 0) {
+        tokens.push({ op, kind: 'phrase', text: phrase.trim() });
+      }
+      continue;
+    }
+
+    // Bare word: read until the next whitespace.
+    let j = i;
+    while (j < n && !isSpace(s[j])) j++;
+    const word = s.slice(i, j);
+    i = j;
+
+    // A bare operator (`-`/`+` with no remainder) or an all-operator remainder is
+    // dropped.
+    if (word.length === 0) continue;
+    if (op && /^[+-]+$/.test(word)) continue;
+
+    tokens.push({ op, kind: 'word', text: word });
+  }
+
+  return tokens;
+}
+
+// Resolve the match branch for a single term given the global match mode.
+function branchForTerm(
+  text: string,
+  kind: 'word' | 'phrase',
+  mode: SearchMatchMode,
+): SearchTermBranch {
+  if (kind === 'phrase') return 'phrase';
+  switch (mode) {
+    case 'word':
+      return 'fts';
+    case 'prefix':
+      return 'ftsPrefix';
+    case 'substring':
+      return 'substring';
+    case 'auto':
+    default:
+      // Identifiers the tokenizer mangles go to substring; words get a prefix
+      // FTS match (so `печат` still finds `печатать`, but `печат` no longer drags
+      // in `впечатления` because the russian stemmer anchors the stem).
+      return isIdentifierLike(text) ? 'substring' : 'ftsPrefix';
+  }
+}
+
+/**
+ * Parse a raw user query + flags into a structured, SQL-agnostic AST.
+ * Pure and total: never throws, always returns a ParsedQuery.
+ */
+export function parseSearchQuery(
+  raw: string,
+  opts: { match?: SearchMatchMode; mode?: SearchBooleanMode } = {},
+): ParsedQuery {
+  const match: SearchMatchMode = opts.match ?? 'auto';
+  const mode: SearchBooleanMode = opts.mode ?? 'or';
+
+  const positive: ParsedTerm[] = [];
+  const required: ParsedTerm[] = [];
+  const excluded: ParsedTerm[] = [];
+
+  for (const tok of tokenize(raw)) {
+    // For an FTS branch, the token must survive metacharacter cleaning; for the
+    // substring/phrase branch the literal text is used. A term that cleans to
+    // nothing AND is not usable as a substring is dropped.
+    const branch = branchForTerm(tok.text, tok.kind, match);
+
+    let usableText: string;
+    if (branch === 'fts' || branch === 'ftsPrefix') {
+      usableText = cleanFtsLexeme(tok.text);
+    } else {
+      // phrase / substring keep the literal (trimmed) text.
+      usableText = tok.text.trim();
+    }
+    if (!usableText) continue;
+
+    // Stack-depth guard: stop after MAX_PARSED_TERMS surviving terms (positive +
+    // required + excluded, combined) so the SQL tsquery nesting stays shallow.
+    // The first 64 terms are kept in stable order; the rest are dropped.
+    if (positive.length + required.length + excluded.length >= MAX_PARSED_TERMS) {
+      break;
+    }
+
+    const term: ParsedTerm = { text: usableText, branch };
+
+    if (tok.op === '+') required.push(term);
+    else if (tok.op === '-') excluded.push(term);
+    else positive.push(term);
+  }
+
+  const parsed: ParsedQuery = { raw: raw ?? '', positive, required, excluded, mode };
+
+  if (positive.length === 0) {
+    // Required terms with no positive recall still form a valid positive set (the
+    // required predicates ARE the recall). Only when there is neither a positive
+    // nor a required term is the query empty / only-negation.
+    if (required.length === 0) {
+      parsed.reason = excluded.length > 0 ? 'only-negation' : 'empty';
+    }
+  }
+
+  return parsed;
+}
+
+/**
+ * Does this parsed query have any positive recall to run? False means we must
+ * short-circuit to an empty result (with `reason`), never a costly NOT-only scan.
+ */
+export function hasPositiveRecall(parsed: ParsedQuery): boolean {
+  return parsed.positive.length > 0 || parsed.required.length > 0;
+}

apps/server/src/core/search/search.module.ts

@@ -1,8 +1,15 @@
 import { Module } from '@nestjs/common';
 import { SearchController } from './search.controller';
 import { SearchService } from './search.service';
+import { AiModule } from '../../integrations/ai/ai.module';
 
+/**
+ * #530: AiModule supplies AiService (embedQuery / resolveEmbeddingProvider).
+ * PageEmbeddingRepo is provided by the @Global DatabaseModule, so it is injected
+ * into SearchService without an explicit import here.
+ */
 @Module({
+  imports: [AiModule],
   controllers: [SearchController],
   providers: [SearchService],
   exports: [SearchService],

apps/server/src/core/search/search.service.lookup.spec.ts

@@ -1,19 +1,14 @@
-import {
-  computeLookupScore,
-  escapeLikePattern,
-  SearchLookupTier,
-} from './search.service';
+import { escapeLikePattern } from './search.service';
 
 /**
- * Pure-function coverage for the #443 agent-lookup helpers:
- *  - escapeLikePattern: LIKE-metacharacter escaping so `%`/`_`/`\` are literals
- *    (the acceptance-table requirement that a query of `%` or `_` does NOT match
- *    everything);
- *  - computeLookupScore: the tiered 0..1 ranking score, where a stronger tier
- *    always outranks a weaker one regardless of the in-tier secondary signal.
+ * Pure-function coverage for `escapeLikePattern` — LIKE-metacharacter escaping so
+ * `%`/`_`/`\` are matched literally (the acceptance requirement that a query of
+ * `%` or `_` does NOT match everything, #529 acceptance #10). The substring
+ * branch's DB behaviour is covered by the integration spec.
  *
- * The DB-touching branch (substring UNION FTS, path CTE, snippet window) is
- * covered by the integration spec against the real schema.
+ * NOTE (#529): the old tiered `computeLookupScore` was replaced by RRF rank
+ * fusion in the unified engine, so its unit coverage moved to the integration
+ * ordering tests; only the escaping helper remains a pure unit here.
  */
 describe('escapeLikePattern', () => {
   it('escapes the LIKE metacharacters % _ and \\', () => {
@@ -43,53 +38,3 @@ describe('escapeLikePattern', () => {
     expect(escapeLikePattern(null as any)).toBe('');
   });
 });
-
-describe('computeLookupScore', () => {
-  it('keeps every score within (0, 1]', () => {
-    for (const tier of [
-      SearchLookupTier.TITLE_EXACT,
-      SearchLookupTier.TITLE_SUBSTRING,
-      SearchLookupTier.TEXT,
-    ]) {
-      for (const secondary of [0, 0.001, 1, 100, 1e6]) {
-        const s = computeLookupScore({ tier, secondary });
-        expect(s).toBeGreaterThan(0);
-        expect(s).toBeLessThanOrEqual(1);
-      }
-    }
-  });
-
-  it('a stronger tier ALWAYS outranks a weaker tier, whatever the secondary', () => {
-    // Weak tier with a huge secondary must still lose to a strong tier with a
-    // tiny secondary — tiers dominate.
-    const strongLowSecondary = computeLookupScore({
-      tier: SearchLookupTier.TITLE_EXACT,
-      secondary: 0,
-    });
-    const weakHighSecondary = computeLookupScore({
-      tier: SearchLookupTier.TEXT,
-      secondary: 1e9,
-    });
-    expect(strongLowSecondary).toBeGreaterThan(weakHighSecondary);
-  });
-
-  it('within a tier a larger secondary sorts higher', () => {
-    const lo = computeLookupScore({
-      tier: SearchLookupTier.TEXT,
-      secondary: 0.1,
-    });
-    const hi = computeLookupScore({
-      tier: SearchLookupTier.TEXT,
-      secondary: 5,
-    });
-    expect(hi).toBeGreaterThan(lo);
-  });
-
-  it('treats a negative/absent secondary as 0', () => {
-    const zero = computeLookupScore({ tier: SearchLookupTier.TEXT, secondary: 0 });
-    expect(computeLookupScore({ tier: SearchLookupTier.TEXT })).toBe(zero);
-    expect(
-      computeLookupScore({ tier: SearchLookupTier.TEXT, secondary: -5 }),
-    ).toBe(zero);
-  });
-});

apps/server/src/core/search/search.service.query-mode.spec.ts

@@ -1,74 +1,45 @@
 import { SearchService } from './search.service';
 
 /**
- * Coverage for SearchService.searchPage query-mode selection (search.service.ts
- * @25). searchPage chooses HOW the result set is scoped — by explicit space, by
- * the authenticated user's member spaces, or by a share — and must return an
- * empty set (without leaking data) for every disallowed combination.
+ * Unit coverage for SearchService.searchPage SCOPE-SECURITY early returns — the
+ * branches that must yield an empty result WITHOUT ever touching the DB, so they
+ * can leak nothing. The happy-path scope SQL (explicit space / member spaces /
+ * share id set) is covered against the real schema in the integration spec.
  *
- * The kysely query builder is mocked with the same chainable pattern as the
- * existing search.service.spec.ts: every builder method returns the same builder
- * and `.execute()` resolves the supplied rows. Each `.where(...)` call is
- * recorded so we can assert exactly which scope clause was applied — that is the
- * mutation-resistant signal that distinguishes one query mode from another.
- *
- * These specs catch cross-space / cross-workspace search leakage and
- * share-scope bypass (data exposure).
+ * Every case here returns BEFORE the raw-SQL candidate query runs, so a bare `db`
+ * stub (never called) is enough — a call to it would itself be a failure signal.
  */
-describe('SearchService.searchPage — query-mode selection', () => {
-  // Build a chainable selectFrom('pages') builder that records its calls. The
-  // builder is returned from `db.selectFrom` and is the single object every
-  // chained call mutates/returns, mirroring the existing spec's pattern.
-  function makeBuilder(rows: Array<{ id: string; highlight?: string }>) {
-    const builder: any = {};
-    builder.select = jest.fn(() => builder);
-    builder.where = jest.fn(() => builder);
-    builder.$if = jest.fn(() => builder);
-    builder.orderBy = jest.fn(() => builder);
-    builder.limit = jest.fn(() => builder);
-    builder.offset = jest.fn(() => builder);
-    builder.execute = jest.fn(async () => rows);
-    return builder;
-  }
-
+describe('SearchService.searchPage — scope-security early returns', () => {
   function makeService(opts?: {
-    rows?: Array<{ id: string; highlight?: string }>;
     share?: any;
     isRestricted?: boolean;
-    descendants?: Array<{ id: string }>;
+    memberSpaceIds?: string[];
   }) {
-    const builder = makeBuilder(opts?.rows ?? []);
-
-    const db: any = {
-      selectFrom: jest.fn(() => builder),
-    };
-
-    // `getUserSpaceIdsQuery` returns a sub-query object that searchPage passes
-    // straight into `.where('spaceId', 'in', <subquery>)`. A sentinel is enough
-    // to assert the user-scoped branch was taken.
-    const userSpaceIdsQuery = { __userSpaceIdsQuery: true };
+    // A db that THROWS if touched — these branches must not reach SQL.
+    const db: any = new Proxy(
+      {},
+      {
+        get() {
+          throw new Error('db must not be touched on an empty-scope branch');
+        },
+      },
+    );
 
     const pageRepo = {
-      // `.select((eb) => this.pageRepo.withSpace(eb))` — value ignored by stub.
-      withSpace: jest.fn(() => ({ __withSpace: true })),
-      getPageAndDescendantsExcludingRestricted: jest
-        .fn()
-        .mockResolvedValue(opts?.descendants ?? []),
+      getPageAndDescendantsExcludingRestricted: jest.fn(),
+      getPageAndDescendants: jest.fn(),
     };
     const shareRepo = {
       findById: jest.fn().mockResolvedValue(opts?.share ?? null),
     };
     const spaceMemberRepo = {
-      getUserSpaceIdsQuery: jest.fn(() => userSpaceIdsQuery),
+      getUserSpaceIds: jest.fn().mockResolvedValue(opts?.memberSpaceIds ?? []),
     };
     const pagePermissionRepo = {
       hasRestrictedAncestor: jest
         .fn()
         .mockResolvedValue(opts?.isRestricted ?? false),
-      // Let everything through page-level permission filtering by default.
-      filterAccessiblePageIds: jest
-        .fn()
-        .mockImplementation(async ({ pageIds }: { pageIds: string[] }) => pageIds),
+      filterAccessiblePageIds: jest.fn(),
     };
 
     const service = new SearchService(
@@ -77,146 +48,86 @@ describe('SearchService.searchPage — query-mode selection', () => {
       shareRepo as any,
       spaceMemberRepo as any,
       pagePermissionRepo as any,
+      // #530 semantic path off for these query-mode unit tests: a stubbed
+      // embedQuery that throws makes the service degrade to the lexical path.
+      { embedQuery: jest.fn().mockRejectedValue(new Error('no embed')) } as any,
+      { vectorCandidateArm: jest.fn() } as any,
     );
-
-    return {
-      service,
-      db,
-      builder,
-      pageRepo,
-      shareRepo,
-      spaceMemberRepo,
-      pagePermissionRepo,
-      userSpaceIdsQuery,
-    };
+    return { service, pageRepo, shareRepo, spaceMemberRepo, pagePermissionRepo };
   }
 
-  const whereCallFor = (builder: any, column: any) =>
-    builder.where.mock.calls.find((c: any[]) => c[0] === column);
-
-  it('returns {items:[]} for a blank query WITHOUT touching the DB', async () => {
-    const { service, db } = makeService();
-
+  it('returns total:0 for a blank query WITHOUT touching the DB or any repo', async () => {
+    const { service, shareRepo, spaceMemberRepo } = makeService();
     const result = await service.searchPage(
       { query: '' } as any,
       { userId: 'user-1', workspaceId: 'ws-1' },
     );
-
-    expect(result).toEqual({ items: [] });
-    // Blank query is rejected before any query builder is constructed.
-    expect(db.selectFrom).not.toHaveBeenCalled();
+    expect(result.items).toEqual([]);
+    expect(result.total).toBe(0);
+    expect(shareRepo.findById).not.toHaveBeenCalled();
+    expect(spaceMemberRepo.getUserSpaceIds).not.toHaveBeenCalled();
   });
 
-  it('scopes to the explicit spaceId branch', async () => {
-    const { service, builder, db, spaceMemberRepo, shareRepo } = makeService({
-      rows: [{ id: 'p-1' }],
-    });
-
+  it('only-negation short-circuits with reason "only-negation", never scanning', async () => {
+    const { service, spaceMemberRepo } = makeService();
     const result = await service.searchPage(
-      { query: 'plan', spaceId: 'space-42' } as any,
+      { query: '-архив' } as any,
       { userId: 'user-1', workspaceId: 'ws-1' },
     );
-
-    expect(db.selectFrom).toHaveBeenCalledWith('pages');
-    // The explicit-space branch adds exactly `.where('spaceId', '=', 'space-42')`.
-    expect(whereCallFor(builder, 'spaceId')).toEqual([
-      'spaceId',
-      '=',
-      'space-42',
-    ]);
-    // It must NOT fall through to the user-member-spaces or share branch.
-    expect(spaceMemberRepo.getUserSpaceIdsQuery).not.toHaveBeenCalled();
-    expect(shareRepo.findById).not.toHaveBeenCalled();
-    expect(result.items.map((i: any) => i.id)).toEqual(['p-1']);
+    expect(result.total).toBe(0);
+    expect(result.query.parsed.reason).toBe('only-negation');
+    // Never resolves scope (returns before) — no expensive NOT-only scan.
+    expect(spaceMemberRepo.getUserSpaceIds).not.toHaveBeenCalled();
   });
 
-  it('scopes an authenticated user WITHOUT spaceId to their member spaces', async () => {
-    const { service, builder, spaceMemberRepo, userSpaceIdsQuery, shareRepo } =
-      makeService({ rows: [{ id: 'p-9' }] });
-
-    await service.searchPage(
-      { query: 'plan' } as any,
-      { userId: 'user-7', workspaceId: 'ws-1' },
-    );
-
-    // The user-scoped branch resolves the member-spaces sub-query for that user
-    // and restricts both spaceId (to that sub-query) and workspaceId.
-    expect(spaceMemberRepo.getUserSpaceIdsQuery).toHaveBeenCalledWith('user-7');
-    expect(whereCallFor(builder, 'spaceId')).toEqual([
-      'spaceId',
-      'in',
-      userSpaceIdsQuery,
-    ]);
-    expect(whereCallFor(builder, 'workspaceId')).toEqual([
-      'workspaceId',
-      '=',
-      'ws-1',
-    ]);
-    // Authenticated user path must not consult shares.
-    expect(shareRepo.findById).not.toHaveBeenCalled();
-  });
-
-  it('returns {items:[]} when the share belongs to a DIFFERENT workspace', async () => {
-    const { service, builder, shareRepo, pagePermissionRepo } = makeService({
-      share: {
-        id: 'share-1',
-        pageId: 'page-1',
-        workspaceId: 'OTHER-ws',
-        includeSubPages: false,
-      },
+  it('returns empty when the share belongs to a DIFFERENT workspace (no leak)', async () => {
+    const { service, shareRepo, pagePermissionRepo } = makeService({
+      share: { id: 's1', pageId: 'p1', workspaceId: 'OTHER', includeSubPages: false },
     });
-
     const result = await service.searchPage(
-      { query: 'plan', shareId: 'share-1' } as any,
+      { query: 'plan', shareId: 's1' } as any,
       { workspaceId: 'ws-1' },
     );
-
-    expect(shareRepo.findById).toHaveBeenCalledWith('share-1');
-    expect(result).toEqual({ items: [] });
-    // Workspace mismatch short-circuits before any restricted-ancestor / id
-    // scoping or DB execution: no leak across workspaces.
+    expect(shareRepo.findById).toHaveBeenCalledWith('s1');
+    expect(result.items).toEqual([]);
+    // Workspace mismatch short-circuits before restricted-ancestor / enumeration.
     expect(pagePermissionRepo.hasRestrictedAncestor).not.toHaveBeenCalled();
-    expect(builder.execute).not.toHaveBeenCalled();
   });
 
-  it('returns {items:[]} when the shared page has a restricted ancestor', async () => {
-    const { service, builder, pagePermissionRepo, pageRepo } = makeService({
-      share: {
-        id: 'share-1',
-        pageId: 'page-1',
-        workspaceId: 'ws-1',
-        includeSubPages: true,
-      },
+  it('returns empty when the shared page has a restricted ancestor', async () => {
+    const { service, pagePermissionRepo, pageRepo } = makeService({
+      share: { id: 's1', pageId: 'p1', workspaceId: 'ws-1', includeSubPages: true },
       isRestricted: true,
     });
-
     const result = await service.searchPage(
-      { query: 'plan', shareId: 'share-1' } as any,
+      { query: 'plan', shareId: 's1' } as any,
       { workspaceId: 'ws-1' },
     );
-
-    expect(pagePermissionRepo.hasRestrictedAncestor).toHaveBeenCalledWith(
-      'page-1',
-    );
-    expect(result).toEqual({ items: [] });
-    // Restricted ancestor must block before page enumeration and DB execution.
+    expect(pagePermissionRepo.hasRestrictedAncestor).toHaveBeenCalledWith('p1');
+    expect(result.items).toEqual([]);
     expect(
       pageRepo.getPageAndDescendantsExcludingRestricted,
     ).not.toHaveBeenCalled();
-    expect(builder.execute).not.toHaveBeenCalled();
   });
 
-  it('returns {items:[]} with no userId, no spaceId and no shareId', async () => {
-    const { service, builder, shareRepo } = makeService();
-
+  it('returns empty with no userId, no spaceId and no shareId', async () => {
+    const { service, shareRepo } = makeService();
     const result = await service.searchPage(
       { query: 'plan' } as any,
       { workspaceId: 'ws-1' },
     );
-
-    expect(result).toEqual({ items: [] });
-    // The catch-all else returns empty without scoping/executing or hitting shares.
+    expect(result.items).toEqual([]);
     expect(shareRepo.findById).not.toHaveBeenCalled();
-    expect(builder.execute).not.toHaveBeenCalled();
+  });
+
+  it('an authenticated user with NO member spaces gets an empty result', async () => {
+    const { service, spaceMemberRepo } = makeService({ memberSpaceIds: [] });
+    const result = await service.searchPage(
+      { query: 'plan' } as any,
+      { userId: 'user-1', workspaceId: 'ws-1' },
+    );
+    expect(spaceMemberRepo.getUserSpaceIds).toHaveBeenCalledWith('user-1');
+    expect(result.items).toEqual([]);
+    expect(result.total).toBe(0);
   });
 });

apps/server/src/core/search/search.service.spec.ts

@@ -1,4 +1,4 @@
-import { SearchService, buildTsQuery } from './search.service';
+import { SearchService } from './search.service';
 
 describe('SearchService', () => {
   it('should be defined', () => {
@@ -11,6 +11,8 @@ describe('SearchService', () => {
       {} as any, // shareRepo
       {} as any, // spaceMemberRepo
       {} as any, // pagePermissionRepo
+      {} as any, // aiService
+      {} as any, // pageEmbeddingRepo
     );
     expect(service).toBeDefined();
   });
@@ -61,6 +63,8 @@ describe('SearchService.searchSuggestions — onlyTemplates filter', () => {
       shareRepo as any,
       spaceMemberRepo as any,
       pagePermissionRepo as any,
+      {} as any, // aiService (searchSuggestions never embeds)
+      {} as any, // pageEmbeddingRepo
     );
 
     return { service, db, pageBuilder };
@@ -99,59 +103,3 @@ describe('SearchService.searchSuggestions — onlyTemplates filter', () => {
     expect(isTemplateWhereCall(pageBuilder)).toBeUndefined();
   });
 });
-
-// Unit tests for `buildTsQuery` (extracted from search.service.ts). It turns a raw
-// user query into a prefix tsquery string fed to `to_tsquery('english', ...)`.
-//
-// REAL BUG (Gitea #139, item 10): the previous inline `tsquery(query.trim() + '*')`
-// let to_tsquery operator characters through, so adversarial inputs could produce a
-// fragment that to_tsquery rejects -> 500. The extraction sanitizes the input
-// (strip everything but letters/numbers/whitespace) so these inputs degrade to a
-// safe, neutral query with NO throw, while normal queries keep working.
-describe('buildTsQuery', () => {
-  it('builds a prefix query for a normal single word', () => {
-    expect(buildTsQuery('hello')).toBe('hello:*');
-  });
-
-  it('joins multiple words with AND and a trailing prefix match', () => {
-    expect(buildTsQuery('foo bar')).toBe('foo&bar:*');
-  });
-
-  it('preserves accented and non-Latin words', () => {
-    expect(buildTsQuery('héllo café')).toBe('héllo&café:*');
-    expect(buildTsQuery('日本語')).toBe('日本語:*');
-  });
-
-  it('neutralizes to_tsquery operator inputs without throwing', () => {
-    // Each of these previously risked an invalid to_tsquery -> 500. They must now
-    // produce a safe (here empty) query and never throw.
-    for (const input of ['&', '!', '*', '<->', '\\']) {
-      expect(() => buildTsQuery(input)).not.toThrow();
-      expect(buildTsQuery(input)).toBe('');
-    }
-  });
-
-  it('handles stopword-only input safely', () => {
-    // pg-tsquery still tokenizes stopwords; to_tsquery reduces them to nothing.
-    // The important contract is: no throw, and a deterministic string.
-    expect(() => buildTsQuery('the a of')).not.toThrow();
-    expect(buildTsQuery('the a of')).toBe('the&a&of:*');
-  });
-
-  it('returns empty string for empty / whitespace-only / null-ish input', () => {
-    expect(buildTsQuery('')).toBe('');
-    expect(buildTsQuery('   ')).toBe('');
-    expect(buildTsQuery(undefined as unknown as string)).toBe('');
-  });
-
-  it('handles a very long input without throwing', () => {
-    const long = 'a'.repeat(10000);
-    expect(() => buildTsQuery(long)).not.toThrow();
-    expect(buildTsQuery(long)).toBe(`${long}:*`);
-  });
-
-  it('strips punctuation embedded in otherwise valid words', () => {
-    expect(buildTsQuery('c++ code')).toBe('c&code:*');
-    expect(buildTsQuery('a-b-c')).toBe('a&b&c:*');
-  });
-});

apps/server/src/database/database.module.ts

@@ -32,6 +32,7 @@ import { TemplateRepo } from '@docmost/db/repos/template/template.repo';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
 import { AiChatRunRepo } from '@docmost/db/repos/ai-chat/ai-chat-run.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { AiChatPageSnapshotRepo } from '@docmost/db/repos/ai-chat/ai-chat-page-snapshot.repo';
 import { AiProviderCredentialsRepo } from '@docmost/db/repos/ai-chat/ai-provider-credentials.repo';
 import { AiMcpServerRepo } from '@docmost/db/repos/ai-chat/ai-mcp-server.repo';
@@ -125,6 +126,7 @@ import { firstSqlToken } from '../integrations/metrics/metrics.constants';
     AiChatRepo,
     AiChatMessageRepo,
     AiChatRunRepo,
+    AiChatRunStepRepo,
     AiChatPageSnapshotRepo,
     AiProviderCredentialsRepo,
     AiMcpServerRepo,
@@ -161,6 +163,7 @@ import { firstSqlToken } from '../integrations/metrics/metrics.constants';
     AiChatRepo,
     AiChatMessageRepo,
     AiChatRunRepo,
+    AiChatRunStepRepo,
     AiChatPageSnapshotRepo,
     AiProviderCredentialsRepo,
     AiMcpServerRepo,

apps/server/src/database/migrations/20260707T130000-search-ru-en-config.ts

@@ -0,0 +1,242 @@
+import { type Kysely, sql } from 'kysely';
+
+/**
+ * #529 Phase A1 — the `ru_en` text-search configuration + the config swap.
+ *
+ * WHY: the search stack was pinned to the `english` FTS config, which stems only
+ * Latin words. On a Russian-language wiki that is a morphology black hole:
+ * «ресторанов москвы» never matched a page titled «ресторан в москве». `ru_en`
+ * layers the russian_stem over the Cyrillic token classes and english_stem over
+ * the ascii ones, so BOTH languages get proper morphology from one config.
+ *
+ *   CREATE TEXT SEARCH CONFIGURATION ru_en (COPY = simple);
+ *   ALTER ... asciiword/asciihword/hword_asciipart WITH english_stem;
+ *   ALTER ... word/hword/hword_part           WITH russian_stem;
+ *
+ * `to_tsvector('ru_en', …)` with the LITERAL config name is IMMUTABLE, so it is
+ * valid inside a trigger, a generated column and an index expression.
+ *
+ * THE INVARIANT (acceptance #13): the config of the STORED column and the config
+ * of the QUERY must change together. This migration flips BOTH stored sides
+ * (pages.tsv via its trigger + a reindex; page_embeddings.fts, the RAG lexical
+ * leg, via its generated expression); the matching QUERY-side flips
+ * (search.service.ts and page-embedding.repo.ts `hybridSearch`) ship in the SAME
+ * commit. Trigram indexes are LOWER(f_unaccent(...)) and do NOT depend on the FTS
+ * config, so they are untouched.
+ *
+ * REINDEX / LOCK MODEL (deploy-critical). Kysely runs EACH migration in its OWN
+ * transaction (see sibling 20260706T120000 "Kysely runs each migration in a
+ * transaction"; migrate.ts / migration.service.ts set no `disableTransactions`,
+ * and PostgresJSDialect has transactional DDL). A single migration file therefore
+ * cannot commit between batches, so the issue's "procedural batch job OUTSIDE the
+ * transaction + dual-config read window + migration_complete gate" is not
+ * expressible in-migration here. That machinery bridges a reindex spread over
+ * MANY committed batches, during which some rows are still `english` while others
+ * are already `ru_en`. Our pages.tsv reindex is a SINGLE `UPDATE pages SET tsv`,
+ * ATOMIC within THIS migration's own transaction: at COMMIT every row is `ru_en`
+ * at once, so no morphology-desync window exists and no dual-config read path is
+ * required — the query config flips to `ru_en` in the very same release. This is
+ * the deliberate, correct adaptation to this framework (see the PR notes).
+ *
+ *   - pages.tsv: swapping the trigger is a cheap catalog change (no table lock),
+ *     and the reindex is a single `UPDATE pages SET tsv = <ru_en expr>` — a
+ *     ROW-level-lock (RowExclusiveLock) backfill, NOT an ACCESS EXCLUSIVE rewrite
+ *     (mirrors the existing space_id backfill in 20250725T052004). On a LARGE
+ *     tenant this still writes every row + its WAL and leaves dead tuples, so it
+ *     can take MINUTES and blocks the startup migrator for that time — but it
+ *     never blocks concurrent reads. It stays inline unconditionally.
+ *
+ *   - page_embeddings.fts is a GENERATED STORED column; Postgres cannot change a
+ *     generated expression without DROP+ADD, which is a full-table ACCESS
+ *     EXCLUSIVE REWRITE of page_embeddings — it blocks ALL reads AND writes on
+ *     that table (including the RAG agent) for the rewrite's duration. That inline
+ *     rewrite is appropriate for small/typical tenants (this fork's target) and
+ *     is the DEFAULT.
+ *
+ *     LARGE TENANTS have two documented escape hatches, either of which makes the
+ *     migration genuinely no-op the rewrite (it is NOT a blind DROP+ADD):
+ *       (a) Set `SEARCH_EMBEDDINGS_FTS_INLINE_REWRITE=false`. The migration then
+ *           SKIPS the embeddings rewrite entirely and logs a WARNING. The operator
+ *           MUST perform the ru_en fts swap out-of-band; until they do, the RAG
+ *           lexical leg stays on `english` while the query config is `ru_en` — a
+ *           documented, operator-owned desync window. (pages.tsv still swaps
+ *           inline — the gate is ONLY the embeddings rewrite.)
+ *       (b) Perform the swap out-of-band BEFORE deploy — add a plain column →
+ *           batched backfill → brief-lock swap → CREATE INDEX CONCURRENTLY — so
+ *           the `fts` column's generated expression already references the TARGET
+ *           config when the migration runs. The migration detects this (it reads
+ *           the column's actual generation expression from pg_catalog) and does a
+ *           TRUE no-op — no DROP, no ADD, no rewrite. This is real idempotency,
+ *           not the old (false) "IF-EXISTS guards no-op" claim: `DROP COLUMN IF
+ *           EXISTS` guards against ABSENCE, not presence, so it would have dropped
+ *           and recreated an existing `fts` regardless. The at-target check is the
+ *           only honest no-op path.
+ *
+ * Same documented trade-off family as the #443 trgm GIN migration (20260706T120000).
+ */
+
+// pages.tsv trigger body for a given FTS config — mirrors the latest form
+// (20250729T213756): f_unaccent + a 1MB text cap on text_content, weights A/B.
+function pagesTriggerSql(config: 'ru_en' | 'english') {
+  return sql`
+    CREATE OR REPLACE FUNCTION pages_tsvector_trigger() RETURNS trigger AS $$
+    begin
+        new.tsv :=
+                  setweight(to_tsvector('${sql.raw(config)}', f_unaccent(coalesce(new.title, ''))), 'A') ||
+                  setweight(to_tsvector('${sql.raw(config)}', f_unaccent(substring(coalesce(new.text_content, ''), 1, 1000000))), 'B');
+        return new;
+    end;
+    $$ LANGUAGE plpgsql;
+  `;
+}
+
+async function swapPagesConfig(db: Kysely<any>, config: 'ru_en' | 'english') {
+  // 1. Point the trigger at the target config (new/edited rows use it going
+  //    forward). CREATE OR REPLACE FUNCTION takes only a brief catalog lock.
+  await pagesTriggerSql(config).execute(db);
+
+  // 2. Reindex existing rows: recompute tsv directly with the target config. A
+  //    plain UPDATE — row locks, no ACCESS EXCLUSIVE. Equivalent to firing the
+  //    trigger but cheaper (no self-update round trip).
+  await sql`
+    UPDATE pages
+    SET tsv =
+        setweight(to_tsvector('${sql.raw(config)}', f_unaccent(coalesce(title, ''))), 'A') ||
+        setweight(to_tsvector('${sql.raw(config)}', f_unaccent(substring(coalesce(text_content, ''), 1, 1000000))), 'B')
+  `.execute(db);
+}
+
+// The default in-migration ACCESS EXCLUSIVE rewrite of page_embeddings.fts is
+// ON unless the operator explicitly opts out with the env flag. Parsed strictly
+// (mirrors CLIENT_TELEMETRY_ENABLED / DEBUG_MODE in common/): only a literal
+// (case-insensitive) 'false' disables it; anything else — unset included —
+// keeps the default true.
+function inlineEmbeddingsRewriteEnabled(): boolean {
+  return (
+    (process.env.SEARCH_EMBEDDINGS_FTS_INLINE_REWRITE ?? 'true').toLowerCase() !==
+    'false'
+  );
+}
+
+// Read page_embeddings.fts's ACTUAL generated-column expression from pg_catalog
+// (the generation expression is stored as a column default marked generated).
+// Returns '' when the column is absent.
+async function embeddingsFtsExpr(db: Kysely<any>): Promise<string> {
+  const r = await sql<{ def: string }>`
+    SELECT pg_get_expr(d.adbin, d.adrelid) AS def
+    FROM pg_attrdef d
+    JOIN pg_attribute a ON a.attrelid = d.adrelid AND a.attnum = d.adnum
+    WHERE a.attname = 'fts'
+      AND a.attrelid = 'page_embeddings'::regclass
+      AND NOT a.attisdropped
+  `.execute(db);
+  return r.rows[0]?.def ?? '';
+}
+
+async function swapEmbeddingsFtsConfig(
+  db: Kysely<any>,
+  config: 'ru_en' | 'english',
+) {
+  // 1. TRUE no-op path (real out-of-band escape hatch): if the column's current
+  //    generation expression already references the TARGET config, there is
+  //    nothing to do. An operator who pre-swapped the column out-of-band lands
+  //    here and the migration does NOT rewrite the table. ('ru_en' and 'english'
+  //    are disjoint tokens, neither a substring of the other or of the rest of
+  //    the expression, so a plain contains-check is unambiguous.)
+  const currentExpr = await embeddingsFtsExpr(db);
+  if (currentExpr.includes(config)) return;
+
+  // 2. Env-gated opt-out: large tenants set SEARCH_EMBEDDINGS_FTS_INLINE_REWRITE
+  //    =false to skip the ACCESS EXCLUSIVE rewrite in-migration and own the swap
+  //    out-of-band. Warn loudly so the desync window is not silent.
+  if (!inlineEmbeddingsRewriteEnabled()) {
+    // eslint-disable-next-line no-console
+    console.warn(
+      `[migration 20260707T130000] SEARCH_EMBEDDINGS_FTS_INLINE_REWRITE=false: ` +
+        `SKIPPING the page_embeddings.fts rewrite to '${config}'. The operator MUST ` +
+        `perform this fts swap out-of-band. Until then the RAG lexical leg stays on ` +
+        `its current config while the query config is '${config}' (documented, ` +
+        `operator-owned desync window).`,
+    );
+    return;
+  }
+
+  // 3. Default inline path: the generated `fts` expression can only change via
+  //    DROP+ADD (a full-table ACCESS EXCLUSIVE rewrite; see the lock note in the
+  //    header). The GIN index depends on the column, so it is dropped with it and
+  //    recreated.
+  await sql`DROP INDEX IF EXISTS idx_page_embeddings_fts`.execute(db);
+  await sql`ALTER TABLE page_embeddings DROP COLUMN IF EXISTS fts`.execute(db);
+  await sql`
+    ALTER TABLE page_embeddings
+      ADD COLUMN fts tsvector
+      GENERATED ALWAYS AS (to_tsvector('${sql.raw(config)}', f_unaccent(content))) STORED
+  `.execute(db);
+  await sql`
+    CREATE INDEX IF NOT EXISTS idx_page_embeddings_fts
+      ON page_embeddings USING gin(fts)
+  `.execute(db);
+}
+
+async function ruEnConfigExists(db: Kysely<any>): Promise<boolean> {
+  const r = await sql<{ n: number }>`
+    SELECT count(*)::int AS n FROM pg_ts_config WHERE cfgname = 'ru_en'
+  `.execute(db);
+  return (r.rows[0]?.n ?? 0) > 0;
+}
+
+async function ensureRuEnConfig(db: Kysely<any>): Promise<void> {
+  // Idempotent by EXISTENCE, not by drop-recreate. The old `DROP ... IF EXISTS;
+  // CREATE` was safe only on a first run: on a re-run the page_embeddings.fts
+  // generated column already has a hard dependency on ru_en, so dropping the
+  // config would fail. Create only when it is genuinely missing.
+  if (await ruEnConfigExists(db)) return;
+  await sql`CREATE TEXT SEARCH CONFIGURATION ru_en (COPY = simple)`.execute(db);
+  // Latin token classes → english_stem.
+  await sql`
+    ALTER TEXT SEARCH CONFIGURATION ru_en
+      ALTER MAPPING FOR asciiword, asciihword, hword_asciipart
+      WITH english_stem
+  `.execute(db);
+  // Cyrillic / non-ascii token classes → russian_stem.
+  await sql`
+    ALTER TEXT SEARCH CONFIGURATION ru_en
+      ALTER MAPPING FOR word, hword, hword_part
+      WITH russian_stem
+  `.execute(db);
+}
+
+export async function up(db: Kysely<any>): Promise<void> {
+  await ensureRuEnConfig(db);
+
+  // Flip both stored sides to ru_en (query-side flips in the same commit).
+  // swapEmbeddingsFtsConfig no-ops when fts already references ru_en, so a
+  // re-run of up() is idempotent and does NOT re-rewrite the embeddings table.
+  await swapPagesConfig(db, 'ru_en');
+  await swapEmbeddingsFtsConfig(db, 'ru_en');
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  // Reverse ORDER matters: the trigger and the generated column reference the
+  // `ru_en` config by name, so they must be moved back to `english` BEFORE the
+  // config can be dropped (a generated column that still depends on `ru_en` would
+  // block the DROP with a dependency error).
+  await swapEmbeddingsFtsConfig(db, 'english');
+  await swapPagesConfig(db, 'english');
+
+  // Drop the config ONLY if nothing still references it. When the embeddings
+  // rewrite was gated off (SEARCH_EMBEDDINGS_FTS_INLINE_REWRITE=false), the fts
+  // column can still reference ru_en — dropping the config would then fail with a
+  // dependency error. Skip + warn so down() stays non-fatal; the operator drops
+  // ru_en after completing the out-of-band english swap.
+  if ((await embeddingsFtsExpr(db)).includes('ru_en')) {
+    // eslint-disable-next-line no-console
+    console.warn(
+      `[migration 20260707T130000] down(): page_embeddings.fts still references ` +
+        `ru_en (inline rewrite was gated off) — leaving the ru_en text-search ` +
+        `configuration in place. Drop it out-of-band once fts is back on 'english'.`,
+    );
+    return;
+  }
+  await sql`DROP TEXT SEARCH CONFIGURATION IF EXISTS ru_en`.execute(db);
+}

apps/server/src/database/migrations/20260708T120000-ai-chat-run-steps.ts

@@ -0,0 +1,70 @@
+import { type Kysely, sql } from 'kysely';
+
+/**
+ * `ai_chat_run_steps` — append-only per-step persistence for an assistant turn
+ * (#492 wave C). Each finished agent step's UI `parts` (its text part + a part
+ * per tool call, WITH the tool output) is INSERTed as its own lightweight row the
+ * moment the step ends, instead of REWRITING the whole assistant row's growing
+ * `metadata.parts` jsonb on every `onStepFinish`.
+ *
+ * WHY a separate table + INSERT (not a jsonb `||` append on the message row): a
+ * Postgres jsonb UPDATE rewrites the ENTIRE TOASTed row version under MVCC, so
+ * re-persisting a growing `metadata.parts` on every step is O(n²) write volume
+ * (a 50-step run with ~100 KB tool outputs wrote hundreds of MB of WAL / dead
+ * tuples per turn, hammering autovacuum). `||` would only shave the network
+ * payload — the WAL/TOAST rewrite harm remains. An INSERT into a per-step table
+ * writes ONLY that step's bytes, so the per-turn write volume is O(Σ steps).
+ *
+ * The full `metadata.parts` on the message row is assembled ONCE at finalize (the
+ * terminal completed/error/aborted write). Mid-run, a resuming client's seed is
+ * reconstructed by concatenating these step rows in `step_index` order — which
+ * reproduces exactly what the old per-step full-row rewrite persisted. Records
+ * written the OLD way (full `metadata.parts` on the row, no step rows) still
+ * reconstruct from the row unchanged; the two eras are distinguished by whether
+ * the row already carries non-empty `metadata.parts` (see reconstructRunParts /
+ * assembleStepParts in ai-chat.service.ts).
+ *
+ * ON DELETE CASCADE on `message_id`: the step rows are a derived projection of the
+ * assistant message; they must vanish with it (or with its workspace).
+ */
+export async function up(db: Kysely<any>): Promise<void> {
+  await db.schema
+    .createTable('ai_chat_run_steps')
+    .ifNotExists()
+    .addColumn('id', 'uuid', (col) =>
+      col.primaryKey().defaultTo(sql`gen_uuid_v7()`),
+    )
+    // The assistant message row this step belongs to (the #183 projection). The
+    // step rows are a derived, per-step slice of that message, so they cascade.
+    .addColumn('message_id', 'uuid', (col) =>
+      col.references('ai_chat_messages.id').onDelete('cascade').notNull(),
+    )
+    .addColumn('workspace_id', 'uuid', (col) =>
+      col.references('workspaces.id').onDelete('cascade').notNull(),
+    )
+    // 0-based index of the finished step within the turn. Ordering key for
+    // reconstruction; unique per message (idempotent step re-persist).
+    .addColumn('step_index', 'integer', (col) => col.notNull())
+    // The step's UI parts (text part + a `tool-*` part per call, WITH output).
+    // Concatenated in step order to rebuild the turn's `metadata.parts`.
+    .addColumn('parts', 'jsonb', (col) => col.notNull())
+    .addColumn('created_at', 'timestamptz', (col) =>
+      col.notNull().defaultTo(sql`now()`),
+    )
+    .execute();
+
+  // Idempotent per-step persist: a retried INSERT of the same (message, step)
+  // is a no-op (the service uses ON CONFLICT DO NOTHING). This also serves the
+  // reconstruction read (WHERE message_id ORDER BY step_index).
+  await db.schema
+    .createIndex('ai_chat_run_steps_message_step_uidx')
+    .ifNotExists()
+    .on('ai_chat_run_steps')
+    .columns(['message_id', 'step_index'])
+    .unique()
+    .execute();
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema.dropTable('ai_chat_run_steps').ifExists().execute();
+}

apps/server/src/database/migrations/20260712T120000-page-embeddings-fingerprint.ts

@@ -0,0 +1,49 @@
+import { type Kysely, sql } from 'kysely';
+
+/**
+ * #530 Search Phase B (PR-1): add the embedding FINGERPRINT column.
+ *
+ * The fingerprint is a deterministic id of the embedding configuration that
+ * produced a row — model id + revision + query/doc prefix scheme + dimensions
+ * (see AiService.computeEmbeddingFingerprint). Search filters vector candidates
+ * by the workspace's ACTIVE fingerprint so a revision bump or a prefix-scheme
+ * change never fuses incompatible vectors into results.
+ *
+ * PR-1 scope: this migration only ADDS the column (nullable — existing rows stay
+ * NULL = legacy) and the composite index the vector-candidate scan uses. The full
+ * generational swap / GC lifecycle (target-fingerprint reindex, atomic flip,
+ * old-generation GC) is deliberately deferred to PR-2.
+ *
+ * Independent migration (per the #363 crash-loop net): it creates ONLY its own
+ * objects and never touches another migration's tables/indexes, so a partial
+ * failure cannot leave a shared object half-built.
+ */
+export async function up(db: Kysely<any>): Promise<void> {
+  // Nullable text column. Existing rows keep NULL (legacy generation) and are
+  // simply not matched by the active-fingerprint filter until re-indexed.
+  await sql`
+    ALTER TABLE page_embeddings
+      ADD COLUMN IF NOT EXISTS fingerprint text
+  `.execute(db);
+
+  // Composite btree supporting the scoped, dimension- + fingerprint-filtered
+  // vector-candidate scan (workspace_id + space_id + fingerprint + model_dimensions).
+  await db.schema
+    .createIndex('idx_page_embeddings_ws_space_fp_dim')
+    .ifNotExists()
+    .on('page_embeddings')
+    .columns(['workspace_id', 'space_id', 'fingerprint', 'model_dimensions'])
+    .execute();
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema
+    .dropIndex('idx_page_embeddings_ws_space_fp_dim')
+    .ifExists()
+    .execute();
+
+  await sql`
+    ALTER TABLE page_embeddings
+      DROP COLUMN IF EXISTS fingerprint
+  `.execute(db);
+}

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

@@ -0,0 +1,95 @@
+import { Injectable } from '@nestjs/common';
+import { InjectKysely } from 'nestjs-kysely';
+import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
+import { dbOrTx } from '../../utils';
+import { AiChatRunStep } from '@docmost/db/types/entity.types';
+
+/**
+ * Append-only per-step persistence for an assistant turn (#492). Each finished
+ * agent step's UI `parts` (its text part + a `tool-*` part per call, WITH the
+ * tool output) is INSERTed as its own lightweight row the moment the step ends —
+ * instead of REWRITING the assistant row's growing `metadata.parts` jsonb on every
+ * `onStepFinish` (a Postgres jsonb UPDATE rewrites the whole TOASTed row version
+ * under MVCC, so that was O(n²) WAL/dead-tuple churn per turn).
+ *
+ * The full `metadata.parts` on the message row is assembled ONCE at finalize;
+ * mid-run, a resuming client's seed is rebuilt from these rows in `stepIndex`
+ * order (see `assembleStepParts` / the reconstruct seam in ai-chat.service.ts).
+ * Every method is workspace-scoped as defense-in-depth.
+ */
+@Injectable()
+export class AiChatRunStepRepo {
+  constructor(@InjectKysely() private readonly db: KyselyDB) {}
+
+  /**
+   * Append one finished step's parts. Idempotent: a retried persist of the SAME
+   * (message, stepIndex) is a no-op via ON CONFLICT DO NOTHING — the per-step
+   * writes are fired fire-and-forget + serialized, and a duplicate must never
+   * throw into the stream or double the parts. Returns whether a NEW row landed
+   * (false = the step was already persisted).
+   */
+  async insertStep(
+    messageId: string,
+    workspaceId: string,
+    stepIndex: number,
+    parts: unknown,
+    trx?: KyselyTransaction,
+  ): Promise<boolean> {
+    const db = dbOrTx(this.db, trx);
+    const inserted = await db
+      .insertInto('aiChatRunSteps')
+      .values({
+        messageId,
+        workspaceId,
+        stepIndex,
+        // jsonb column: cast through never (same pattern as the message repo).
+        parts: parts as never,
+      })
+      .onConflict((oc) => oc.columns(['messageId', 'stepIndex']).doNothing())
+      .returning('id')
+      .executeTakeFirst();
+    return inserted !== undefined;
+  }
+
+  /** All persisted steps for ONE assistant message, in step order. */
+  async findByMessage(
+    messageId: string,
+    workspaceId: string,
+  ): Promise<AiChatRunStep[]> {
+    return this.db
+      .selectFrom('aiChatRunSteps')
+      .selectAll('aiChatRunSteps')
+      .where('messageId', '=', messageId)
+      .where('workspaceId', '=', workspaceId)
+      .orderBy('stepIndex', 'asc')
+      .execute();
+  }
+
+  /**
+   * All persisted steps for a SET of assistant messages, grouped by messageId
+   * (each group in step order). One query for the batch — the hydration seam
+   * (getMessages / delta / export) calls this only for the rows that actually
+   * need reconstruction (an active new-style row whose `metadata.parts` is still
+   * empty), which is usually none, so this is skipped on the common path.
+   */
+  async findByMessageIds(
+    messageIds: string[],
+    workspaceId: string,
+  ): Promise<Map<string, AiChatRunStep[]>> {
+    const byMessage = new Map<string, AiChatRunStep[]>();
+    if (messageIds.length === 0) return byMessage;
+    const rows = await this.db
+      .selectFrom('aiChatRunSteps')
+      .selectAll('aiChatRunSteps')
+      .where('messageId', 'in', messageIds)
+      .where('workspaceId', '=', workspaceId)
+      .orderBy('stepIndex', 'asc')
+      .execute();
+    for (const row of rows) {
+      const list = byMessage.get(row.messageId);
+      if (list) list.push(row);
+      else byMessage.set(row.messageId, [row]);
+    }
+    return byMessage;
+  }
+}

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

@@ -1,6 +1,6 @@
 import { Injectable } from '@nestjs/common';
 import { InjectKysely } from 'nestjs-kysely';
-import { sql } from 'kysely';
+import { RawBuilder, sql } from 'kysely';
 import * as pgvector from 'pgvector';
 import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
 import { dbOrTx } from '../../utils';
@@ -35,6 +35,9 @@ export interface PageEmbeddingChunkRow {
   content: string;
   modelName: string;
   modelDimensions: number;
+  // #530 PR-1: the active embedding fingerprint (see computeEmbeddingFingerprint).
+  // null only when no provider resolves (the indexer no-ops in that case).
+  fingerprint: string | null;
   embedding: number[];
 }
 
@@ -120,6 +123,7 @@ export class PageEmbeddingRepo {
           content: row.content,
           modelName: row.modelName,
           modelDimensions: row.modelDimensions,
+          fingerprint: row.fingerprint,
           // pgvector.toSql -> '[1,2,3]'; cast the bound literal to vector.
           embedding: sql`${pgvector.toSql(row.embedding)}::vector`,
         })),
@@ -127,6 +131,82 @@ export class PageEmbeddingRepo {
       .execute();
   }
 
+  /**
+   * #530: build the VECTOR candidate arm for SearchService's fused RRF union — a
+   * page-level nearest-neighbour sub-select. Returns a `sql` fragment (not an
+   * executed query) so the caller can UNION ALL it with the lexical arm inside a
+   * single ranked-ids query.
+   *
+   * SCALING BOUNDARY: the column is dimension-agnostic, so it carries NO ANN
+   * index — this is a brute-force O(N) KNN seq scan with `<=>`. Accepted for the
+   * small-tenant / homelab fork target (ANN + a pinned-dimension column are
+   * deferred). It is NOT unbounded, though: the caller (SearchService) runs this
+   * fused query under a per-statement timeout (SEARCH_VECTOR_STATEMENT_TIMEOUT_MS)
+   * and DEGRADES to lexical-only on a 57014 cancellation, so a pathological scan
+   * can never hang the interactive search request.
+   *
+   * The arm collapses a page's chunks to its best (MIN) cosine distance:
+   *   SELECT pages.id, NULL fts_score, NULL sub_tier,
+   *          MIN(pe.embedding <=> $qvec) AS vec_distance
+   *   FROM page_embeddings pe JOIN pages ON pages.id = pe.page_id
+   *   WHERE <scope> AND pe.model_dimensions = $dim AND pe.fingerprint = $fp
+   *   GROUP BY pages.id ORDER BY vec_distance LIMIT $limit
+   *
+   * `scope` is SearchService's shared scope predicate (workspace + space/id set +
+   * creator + descendants + deleted_at), referencing the `pages` table — it must
+   * be spliced verbatim so the vector candidate set mirrors the lexical scope
+   * EXACTLY, minus the lexical text predicate (vector candidates need not match
+   * text). The vector is bound via pgvector's `toSql(...)::vector`, and both the
+   * dimension and the ACTIVE fingerprint are filtered so `<=>` only ever compares
+   * compatible, same-generation vectors (pgvector errors on a dimension
+   * mismatch; a fingerprint mismatch would fuse incomparable vectors).
+   *
+   * `workspaceId` is ALSO filtered directly on `page_embeddings` — not only via
+   * the pages join. It is IMMUTABLE (there are no cross-workspace page moves), so
+   * an embedding row's workspace_id always matches its page's, and this drops NO
+   * legitimate hit; it lets the composite index idx_page_embeddings_ws_space_fp_dim
+   * (workspace_id, space_id, fingerprint, model_dimensions) bite from its LEADING
+   * column, and workspace-scopes candidates at the embedding level (defense in
+   * depth).
+   *
+   * SPACE is deliberately NOT filtered on page_embeddings: `page_embeddings.space_id`
+   * is stamped at index time and is NOT updated when a page is MOVED between spaces
+   * (movePageToSpace does not reindex, and PAGE_MOVED_TO_SPACE has no reindex
+   * consumer), so a moved page's rows carry the OLD space until the next reindex. A
+   * `page_embeddings.space_id = ANY(scope)` predicate would then wrongly drop a
+   * legitimate vector hit for a page moved INTO the searched space. Space scoping
+   * is therefore enforced ONLY by the join to `pages` (pages.space_id ∈ scope,
+   * always current) — the same way the lexical arm scopes, so no space leak.
+   *
+   * The NULL fts_score / sub_tier columns keep the arm UNION-compatible with the
+   * lexical arm's column list.
+   */
+  vectorCandidateArm(params: {
+    queryEmbedding: number[];
+    dimensions: number;
+    fingerprint: string;
+    scope: RawBuilder<unknown>;
+    limit: number;
+    workspaceId: string;
+  }): RawBuilder<unknown> {
+    const qvec = sql`${pgvector.toSql(params.queryEmbedding)}::vector`;
+    return sql`
+      SELECT pages.id AS id,
+             NULL::float AS fts_score,
+             NULL::int AS sub_tier,
+             MIN(page_embeddings.embedding <=> ${qvec}) AS vec_distance
+      FROM page_embeddings
+      JOIN pages ON pages.id = page_embeddings.page_id
+      WHERE ${params.scope}
+        AND page_embeddings.workspace_id = ${params.workspaceId}
+        AND page_embeddings.model_dimensions = ${params.dimensions}
+        AND page_embeddings.fingerprint = ${params.fingerprint}
+      GROUP BY pages.id
+      ORDER BY vec_distance ASC
+      LIMIT ${params.limit}
+    `;
+  }
+
   /**
    * Cosine search over the embeddings, scoped to a workspace AND a set of
    * spaces the caller may read (see semanticSearch access-scoping). Orders by
@@ -200,8 +280,11 @@ export class PageEmbeddingRepo {
    *
    * The `model_dimensions = $dim` filter applies ONLY on the semantic side
    * (cosine compares same-dimension vectors; pgvector errors otherwise). The
-   * lexical side (`fts`) is dimension-independent. If `websearch_to_tsquery`
-   * yields an EMPTY query (e.g. the text is all stopwords) the `@@` matches
+   * lexical side (`fts`) is dimension-independent. Its query config is `ru_en`,
+   * matched IN LOCKSTEP with the `page_embeddings.fts` generated column's config
+   * (#529 acceptance #13): a mismatch silently breaks Cyrillic RAG retrieval. If
+   * `websearch_to_tsquery` yields an EMPTY query (e.g. the text is all stopwords)
+   * the `@@` matches
    * nothing and the lexical CTE is empty, so results degrade to pure-semantic —
    * which is correct behaviour, not an error.
    *
@@ -249,7 +332,7 @@ export class PageEmbeddingRepo {
                row_number() OVER (ORDER BY ts_rank(pe.fts, q.query) DESC) AS rank_ix
         FROM page_embeddings pe
         JOIN pages p ON p.id = pe.page_id,
-             websearch_to_tsquery('english', f_unaccent(${queryText})) AS q(query)
+             websearch_to_tsquery('ru_en', f_unaccent(${queryText})) AS q(query)
         WHERE pe.workspace_id = ${workspaceId}
           AND pe.space_id IN (${spaceList})
           AND p.deleted_at IS NULL

apps/server/src/database/repos/page/page-history.repo.ts

@@ -157,6 +157,44 @@ export class PageHistoryRepo {
     return { ...result, items: result.items.map(attachPageHistoryAgent) };
   }
 
+  /**
+   * #395 — cheap projection of a page's FULL history timeline for the work-time
+   * estimate: only the columns the sessionizer needs, no heavy `content`, sorted
+   * oldest→newest. The secondary `id` tie-break keeps rows sharing a `createdAt`
+   * (e.g. a synchronous pre-agent boundary row + the immediate agent snapshot)
+   * in a deterministic order.
+   */
+  async findTimelineByPageId(
+    pageId: string,
+    trx?: KyselyTransaction,
+  ): Promise<
+    Array<
+      Pick<
+        PageHistory,
+        | 'createdAt'
+        | 'lastUpdatedById'
+        | 'lastUpdatedSource'
+        | 'lastUpdatedAiChatId'
+        | 'kind'
+      >
+    >
+  > {
+    const db = dbOrTx(this.db, trx);
+    return db
+      .selectFrom('pageHistory')
+      .select([
+        'createdAt',
+        'lastUpdatedById',
+        'lastUpdatedSource',
+        'lastUpdatedAiChatId',
+        'kind',
+      ])
+      .where('pageId', '=', pageId)
+      .orderBy('createdAt', 'asc')
+      .orderBy('id', 'asc')
+      .execute();
+  }
+
   async findPageLastHistory(
     pageId: string,
     opts?: {

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

@@ -692,6 +692,22 @@ export interface AiChatRuns {
   updatedAt: Generated<Timestamp>;
 }
 
+// Append-only per-step persistence for an assistant turn (#492). Mirrors
+// migration 20260708T120000-ai-chat-run-steps.ts. Each finished agent step's UI
+// `parts` are INSERTed as their own row (instead of rewriting the message row's
+// growing `metadata.parts` jsonb every step — an O(n²) WAL/TOAST churn). The full
+// `metadata.parts` is assembled once at finalize; mid-run a resuming client's seed
+// is rebuilt by concatenating these rows in `stepIndex` order. Cascades with the
+// assistant message row it projects.
+export interface AiChatRunSteps {
+  id: Generated<string>;
+  messageId: string;
+  workspaceId: string;
+  stepIndex: number;
+  parts: Json;
+  createdAt: Generated<Timestamp>;
+}
+
 // Per-(chat,page) snapshot of the open page's Markdown at the END of the agent's
 // previous turn (#274). Mirrors migration 20260702T120000-ai-chat-page-snapshot.ts.
 // The next turn diffs the CURRENT Markdown against `contentMd` to surface edits a
@@ -729,6 +745,7 @@ export interface DB {
   aiChats: AiChats;
   aiChatMessages: AiChatMessages;
   aiChatRuns: AiChatRuns;
+  aiChatRunSteps: AiChatRunSteps;
   aiChatPageSnapshots: AiChatPageSnapshots;
   apiKeys: ApiKeys;
   attachments: Attachments;

apps/server/src/database/types/embeddings.types.ts

@@ -7,6 +7,11 @@ export interface PageEmbeddings {
   spaceId: string;
   modelName: string;
   modelDimensions: number;
+  // #530 PR-1: the active embedding fingerprint (model id + revision + prefix
+  // scheme + dimensions). NULL on legacy rows written before this column existed.
+  // Search filters candidates by the ACTIVE fingerprint so a revision/prefix
+  // change never mixes incompatible vectors.
+  fingerprint: string | null;
   workspaceId: string;
   // Nullable: page-body embeddings have no attachment (only attachment chunks set it).
   attachmentId: string | null;