Третий класс петли (инцидент 2026-07-10): ран упёрся в 20-шаговый кап, спалив
все шаги на чтение; на 20-м шаге final-step lockdown отнял инструменты
(toolChoice:'none') посреди незаконченной работы → модель выродилась в
текст-повтор («loadTools.» ×20416, 255КБ). Пакет защит по дизайну владельца:
- MAX_AGENT_STEPS 20→50; спеки выводятся из константы (нет захардкоженных 19/20).
- Final-step lockdown под env-тогглом AI_CHAT_FINAL_STEP_LOCKDOWN (дефолт OFF,
по образцу AI_CHAT_DEFERRED_TOOLS): OFF — инструменты доступны на всех шагах +
мягкий финальный нудж; ON — легаси toolChoice:'none'+FINAL_STEP_INSTRUCTION.
Спеки параметризованы по тогглу. .env.example задокументирован.
- Пустой ход (все шаги без текста, шаги исчерпаны) получает синтетический
маркер-текст — виден в UI и реплее.
- Детектор токен-деградации в onChunk (единственная защита от болтовни, БЕЗ
maxOutputTokens — tool-аргументы это выходные токены): чистые правила
(≥25 одинаковых строк ИЛИ периодический хвост), при срабатывании abort через
внутренний AbortController ∪ effectiveSignal (AbortSignal.any), финализация в
onAbort: усечение хвоста, ai_chat_runs.error=Output degeneration detected,
лизы MCP/снапшоты освобождаются (существующий lifecycle).
- Предупреждение о бюджете шагов на MAX-6…MAX-2 с убывающим N.
- loadTools-описание и преамбула каталога явно говорят, что CORE-тулы всегда
активны (список из CORE_TOOL_KEYS динамически).
Внутренний цикл: 2 прохода. Ревью нашло data-loss-риск: правило периодичности
детектора ложно срабатывало на markdown-разделителях/setext-подчёркиваниях/
хвостовых пробелах (монохар-хвост p-периодичен при ЛЮБОМ p → ложный abort с
пометкой error и усечением). Починка: отдельная монохар-проверка (порог 60,
выше любого реального разделителя) + требование ≥2 различных символов в
периодическом блоке при p≥2. Реальный loadTools-цикл (период ~10) ловится.
Мутационно: 59 одинаковых — не флаг, 60 — флаг; loadTools×20416 — флаг.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
resolve/unresolve enqueue — fire-and-forget (void ...catch(warn)): смысл #399
в том, что недоступность очереди НЕ должна ронять HTTP-запрос. Delete-путь уже
покрыт (enqueue awaited перед hard-delete), а reject resolve-пути — нет. Тест:
generalQueue.add реджектит -> resolveComment всё равно resolves (не throws) +
warn залогирован (ошибка проглочена на микротаске после возврата, поэтому
flushMicrotasks перед ассертом). Мутационно: сделать enqueue awaited без catch
-> тест краснеет.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
resolve/unresolve/delete comment-mark синхронно дёргали collab-gateway
(handleYjsEvent) прямо в HTTP-запросе — сетевой раунд-трип к collab на
горячем пути. Теперь:
- DB-строка пишется синхронно (источник истины) с общим таймстампом;
- сама mark-операция уходит идемпотентным COMMENT_MARK_UPDATE-джобом на
GENERAL_QUEUE, воркер проигрывает тот же handleYjsEvent;
- resolve/unresolve — fire-and-forget (best-effort), delete — await энкью
ДО необратимого hard-delete (durability split);
- race-guard: устаревшее ПРОТИВОПОЛОЖНОЕ событие (ts <= updatedAt строки и
состояние расходится) пропускается, а не флипает mark в устаревшее;
- DI-цикл обойдён ленивым moduleRef.get(CollaborationGateway, strict:false).
Внутренний цикл: 2 прохода. Правки по внутреннему ревью: `<` → `<=` в
race-guard (безопасная обработка суб-миллисекундной ничьи двух
противоположных тогглов); задокументирован сознательный компромисс —
транзиентный page.updated-broadcast из воркера несёт только {id}, теряя
name/avatarUrl «кто редактировал» (lastUpdatedById выставляется верно,
косметика, самочинится на следующем реальном редактировании).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
invalidateOnUpdatePage для sidebar-pages кеша спредил сыро {...sidebarPage, title,
icon} — при title-only событии icon приходит undefined и затирался (и наоборот).
Embed-tree путь 20 строками выше уже гардит undefined; sidebar-ветка пропустила
тот же гард. Применён тот же паттерн: ...(title!==undefined?{title}:{}) +
...(icon!==undefined?{icon}:{}). +2 теста (sidebar title-only/icon-only:
непереданное поле не затирается); мутационно (вернуть сырой спред -> тесты краснеют).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
PR #434 (drawio stage 1) added drawioGet/drawioCreate/drawioUpdate to the
shared tool-spec registry with in-app metadata (inAppKey, deferred tier,
catalogLine) but wired them only in the standalone MCP server, breaking the
contract-parity and phantom-catalog unit tests on develop CI.
- expose drawioGet/drawioCreate/drawioUpdate in forUser() via sharedTool(),
mirroring the MCP transport's argument mapping (format ?? 'xml' default,
flat schema regrouped into the client's `where` object, positional
baseHash pass-through)
- extend the DocmostClientLike hand-mirror with the three client methods
- append the three names to the HOST_CONTRACT_METHODS drift-guard whitelist
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- F1 [blocking]: share-modal.test.tsx + comment-content-view.test.tsx mocked
page-query without usePageMetaQuery → 3 tests threw (ShareModal uses it
directly, comment-content-view via MentionContent). Added usePageMetaQuery to
both mocks (the space-tree mocks were already fixed; these two were missed).
- F2: restored refetchOnMount:true on useGetSpacesQuery — ["spaces"] is
invalidated only by same-tab mutations (no socket path), so a cross-actor
change (an admin adding/removing THIS user from a space) left the list stale
until a hard reload. The other refetchOnMount removals (favorites/watched —
per-user, same-tab-only gap) stay removed.
- F3: corrected the trash-list + recent-changes KEEP comments — both keys ARE
invalidated (trash-list by 3 mutations, recent-changes by page CRUD), but
invalidateQueries only marks an UNMOUNTED query stale without refetching, so the
mount refetch closes the gap. The old "never invalidated" wording was wrong and
risked a maintainer deleting a live invalidation as dead code.
- F4: tests for the two load-bearing pure paths — invalidate-on-update-page (the
undefined-guard: a title-only event keeps the icon; sibling/unrelated subtrees
untouched) and breadcrumb-path-equal (equal chain → true; any id/slugId/name/
icon change or length diff → false; both-null → true). Exported
breadcrumbPathEqual for the test.
Gate: client tsc 0; the 4 affected/new test files 33 passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Outside the editor the UI did background work on every tree event, socket
reconnect, and navigation. Tree infra (virtualization/memo/O(N) utils) was
already good — the cost was in the subscriptions and duplicates around it.
Client-only; behavior 1:1.
- Setter-only atom subscriptions → useSetAtom: space-tree-row, use-tree-mutation,
use-tree-socket no longer subscribe every visible row to the WHOLE treeDataAtom
value (a tree event re-rendered all ~20-30 rows, bypassing the DocTreeRow memo).
space-tree-node-menu / mention-list read the tree imperatively (store.get) in
their handlers only. breadcrumb.tsx uses a selectAtom slice (ancestor chain +
field equality) instead of the whole-tree subscription.
- Socket handler cleanup (BUG): use-tree-socket + use-query-subscription now
socket.off() their named handlers on cleanup (were accumulating listeners on
every reconnect → duplicated invalidations/tree-walks). Mirrors
use-notification-socket.
- Field-update tree path: invalidateOnUpdatePage does a pointwise patch of the
cached embed subtrees instead of a blanket invalidatePageTree() (refetch storm);
structural events keep the blanket invalidate.
- usePageMetaQuery: a content-less select slice for the 13 peripheral subscribers
that read only title/permissions/id, so they stop re-rendering every ~3s while
typing / on every collab page.updated (page.tsx keeps the full query for content).
- page.tsx: skeleton + placeholderData keepPreviousData (no blank flash on nav).
- Removed refetchOnMount:true where socket/mutation invalidation already keeps the
cache fresh (favorite/space/space-watcher/workspace). KEPT it on the 3 queries
with NO other freshness path (trash-list, created-by, recent-changes) — the
global default is refetchOnMount:false, so those overrides are load-bearing.
- Small: resize mousemove/up attached only while dragging; per-row emoji-picker
keydown gated on `opened`; AiChatWindow queries enabled only when the window is
open.
Gate: client tsc 0, client vitest page+websocket 200 passed (+editor suites),
build ok.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both are direct consequences of the round-1 F1 fix (uncaching
hasRestrictedPagesInWorkspace):
- F5: that EXISTS(SELECT 1 FROM page_access WHERE workspace_id=?) now runs
per-request on every whole-workspace list endpoint (global search + suggest,
favorites, notifications, recent, created-by), and page_access only had a
space_id index → a seq scan in the common zero-restriction case. Added
idx_page_access_workspace_id to the perf migration (up + down) so it's an
index-only existence probe.
- F6: the DomainMiddleware workspace cache invalidation was untested — the
int-spec passed `{}` for cacheManager, so bustWorkspaceCache's `del` threw into
its own try/catch and never ran. Added a Map-backed cache double with a working
del and two tests: updateSetting busts WORKSPACE_SELF_HOSTED; updateSharingSettings
busts WORKSPACE_SELF_HOSTED + WORKSPACE_BY_HOST(hostname). A missed/mismatched
bust key now fails the suite instead of letting a stale security-relevant
workspace row (enforceSso/status) outlive the mutation.
Gate: server tsc 0; workspace-repo-update-setting + page-permission-workspace-filter
int-specs pass on real Postgres (the new index applies via global-setup).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- F1 [medium — the substantive one]: hasRestrictedPagesInWorkspace is now UNCACHED
(a plain EXISTS per call, like its sibling hasRestrictedPagesInSpace). Caching it
(even 5s) reintroduced an access-control leak the space path never had: a
concurrent whole-workspace read in the insert->commit window of the FIRST
restricted page could re-populate `false` under withCache (read-then-set, no
del-during-read guard) and override the insert-time bust, leaking that page to
unauthorized users for up to the TTL. Uncaching removes both the DB/cache
asymmetry and the TOCTOU race; the space path already accepts this per-call cost.
Reverted the now-unnecessary insertPageAccess cache-bust and removed the dead
HAS_RESTRICTED_PAGES_IN_WORKSPACE cache key.
- F2 [test]: page-permission-workspace-filter.int-spec.ts (real PG) — the
short-circuit returns the full input set with zero restrictions AND filters out
the page the user can't reach when a restriction is present (proving the authz
behavior is unchanged), the 0->1 transition flips immediately, and the flag is
per-workspace scoped.
- F3 [doc]: documented the deploy-time write-lock in the migration header — the
non-CONCURRENT GIN trigram builds take a SHARE lock that blocks writes on
pages/users/… for minutes on a large tenant; run in a maintenance window or
build CONCURRENTLY out-of-band for big installs.
- F4 [doc]: corrected the jwt.strategy comment — the reused req.raw.workspace is
the middleware's selectAll superset (not "the exact row this query returns"),
harmless because AuthWorkspace already preferred that object.
Gate: server tsc 0; the new int-spec 3/3 on real Postgres.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
One migration + targeted hot-path fixes. API behavior 1:1 (schema change = added
indexes + a byte-identical f_unaccent function-body swap, see below).
- Trigram + composite indexes (20260705T120000-perf-indexes.ts): GIN trigram on
LOWER(f_unaccent(title/name)) for pages/users/groups (the /search/suggest
leading-wildcard LIKE did a seq scan per keystroke — EXPLAIN now confirms
Bitmap Index Scan on idx_pages_title_trgm), + page_history(page_id,id DESC),
comments(page_id,id). DEVIATION (verified byte-identical): PG18 cannot inline
the two-arg f_unaccent body during index creation, so up() swaps it to the
schema-qualified single-arg `SELECT public.unaccent($1)` — same dictionary,
identical output for all inputs, so the tsvector trigger + main @@ search stay
consistent with NO reindex; down() restores the exact two-arg body.
- Auth path: jwt.strategy reuses req.raw.workspace when workspaceId matches (the
middleware already validated it) instead of re-querying; domain.middleware
caches the workspace lookup (withCache 15s, invalidated in all 8 WorkspaceRepo
mutators, with a Date reviver for the JSON-serialized cache). USER + SESSION
caching DEFERRED — the invalidation surface (role change doesn't revoke
sessions; revocation includes background jobs) can't be safely covered, and a
missed hook on a security path is worse than the win.
- AI re-embed coalescing: aiQueue.add gets {jobId: embed-<id>, delay: 30s} so
active editing collapses to one job (worker reads current page state).
- filterAccessiblePageIds: hasRestrictedPagesInWorkspace short-circuit skips the
recursive-ancestor CTE when a workspace has zero restricted pages (wired from
search/favorites/notifications/recent/created-by). EXISTS on the same pageAccess
table the CTE anti-joins → no false-positive / no access leak. Busts the cache
on insertPageAccess so a 0->1 restricted transition takes effect immediately
(review F1).
- Small: syncTransclusion guarded by a family-node probe (both old+new content, so
the removal path is preserved); mention notifications enqueue only when the set
gained a member; redis maintainLock clears a prior interval (leak fix).
Skipped as risky (flagged): global ValidationPipe transform change; a pool-wide
statement_timeout (would kill long CREATE INDEX migrations on the same pool).
NOTE: kept the trash query's `content` select — the trash UI reads page.content
for its preview modal (review F3, would have regressed).
Gate: server tsc 0; jest page-permission/auth/search/persistence 15 suites pass;
migration up+down+idempotency verified on real PG18 with EXPLAIN confirming index
use. No new deps.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Побочные находки инцидента #400 (правка большой таблицы через MCP подвешивает всё).
1. Гонка connect-vs-unload в @hocuspocus/server 3.4.4 (вероятный источник 25s
connect-таймаутов): createDocument проверяет loadingDocuments/documents, но НЕ
ждёт unloadingDocuments -> новое соединение может захендшейкаться на умирающий
Document -> redis-sync идёт по пути 'doc не загружен', провайдер висит до
таймаута. Апстрим (main) не починен. pnpm-патч (инфра как у yjs-патча): в начале
createDocument await in-flight unload (обёрнут в try/catch — отклонённый unload
не отравляет открытие, поведение как до патча), в ОБОИХ рантаймах (cjs+esm).
Тест hocuspocus-unload-race: реальный createDocument с засеянным in-flight
unload -> не грузит пока unload не осел; при откате патча тест краснеет.
2. Двойная перекодировка в onLoadDocument (persistence.extension.ts): хук строил
НОВЫЙ Y.Doc и возвращал его -> hocuspocus делал applyUpdate(encodeStateAsUpdate)
ВТОРОЙ раз (315КБ на каждую холодную загрузку); в JSON-ветке результат encode
выбрасывался (мёртвый вызов). Теперь стейт применяется прямо в data.document,
возврат undefined (hocuspocus мержит только при возврате Doc); мёртвый encode
убран. Содержимое документа не меняется — только меньше encode/alloc.
3. isDeepStrictEqual по 84КБ JSON на каждом store: замерил — 1.32мс на 90КБ
(immaterial, <50мс порога; доминируют fromYdoc+encodeStateAsUpdate). Изменений
кода НЕТ по правилу задачи (dirty-флаг только при material).
closes#401
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Аудит при подготовке #413 нашёл дрейфующие дубли между packages/mcp и
packages/prosemirror-markdown. Четыре дедупа (поведение тулов не меняется):
1. node-ops: форк ~960 строк сведён в ОДНУ копию в prosemirror-markdown (живая
mcp-версия — строгое надмножество замороженного #293-seed'а пакета; сверено по
git-истории, новая пакет-копия байт-в-байт == прежней mcp-копии). Barrel-экспорт
полной поверхности; mcp/client.ts/page-search.ts/transforms.ts/collaboration.ts
импортируют из пакета; тесты переехали. node-ops тянет stripInlineMarkdown ->
пакет-локальная text-normalize.ts несёт только этот примитив (mcp-версия —
домен #408; заголовок документирует дубликацию + источник истины).
2. footnote-lex/footnote-analyze (vestigial legacy [^id]: диагностика): сведены к
одному fence-aware предупреждению 'reference-style footnotes -> use ^[...]'
(полезно для класса #410); footnote-lex удалён.
3. footnote-authoring -> примитивы (footnoteContentKey/makeFootnoteDefinition/
generateFootnoteId) перенесены в пакетный footnote.ts, одна реализация конвенции.
4. parse-node-arg -> перенесён в prosemirror-markdown (не mcp: сервер CommonJS не
импортирует ESM-only @docmost/mcp, но нативно импортирует пакет), обе копии
удалены, консьюмеры перенаправлены.
canonicalizeFootnotes/ENFORCEMENT RULE #228 и comment-anchor/json-edit/text-normalize
(mcp) не тронуты. API-поверхность node-ops оставлена чистой для #409/#413.
closes#414
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
wrapToolsWithCommentSignal всегда ставил свой toModelOutput, молча выбрасывая
собственный toModelOutput инструмента (латентная ловушка — будущий тул со своим
toModelOutput тихо сломался бы). Теперь база = origToModelOutput(info) при наличии,
иначе воспроизведённый дефолт SDK; no-signal путь возвращает базу дословно, signal-
путь = части базы (modelOutputToParts: text/json/content) + элемент сигнала последним.
execute по-прежнему возвращает СЫРОЙ результат -> part.output/цитаты байт-идентичны.
Дефолтный путь (единственный исполняемый сегодня) байт-идентичен и SDK-дефолту, и
до-фиксовому signal-пути (проверено повторным ревью). json-ветку загардил ?? null
для симметрии с fallback. +2 теста: тул со своим text/content toModelOutput —
база честно сохраняется и в no-signal, и в signal (сигнал добавлен последним).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Комментарии человека находят агента сами: короткая эфемерная строка
'new comments: N on page … — call listComments(pageId)' в результате ЛЮБОГО
tool-вызова, mid-turn. Общий хелпер packages/mcp/src/comment-signal.ts
(createCommentSignalTracker: watermark + per-page debounce + working-set;
buildCommentSignalLine; defangCommentSignalTitle).
- Standalone MCP (index.ts): второй wrapper в choke point registerTool (паттерн
метрик #402) — отдельный {type:'text'} content-элемент, форма результата не
меняется. Источник: rate-limited listComments по working-set, title через
getPageRaw только на hit. State per-session.
- In-app (ai-chat-tools.service.ts): execute ВСЕГДА возвращает сырой результат
(part.output/цитаты не трогаются), сигнал доставляется модели через отдельный
toModelOutput ({type:'content', value:[raw, signal]}) — зеркало MCP; no-signal
ветка точно воспроизводит дефолт SDK. Источник: REST-probe (осознанный форк от
DB-count из ТЗ — чтобы не менять конструктор сервиса и не ломать спеки).
- Инъекционная защита: в сигнал идут только count+pageId+defanged-title, НИКОГДА
текст комментария (untrusted). Per-page watermark (не глобальный) — комментарии
на второй странице не теряются.
closes#417
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
В ai@6 упавший тул — это tool-error часть в step.content ({type,toolCallId,
toolName,input,error}), а не элемент toolResults. Раньше serializeSteps писал
только toolCalls+toolResults (ошибка терялась, orphan tool-call без результата),
а assistantParts эмитил заглушку 'Tool call did not complete.' (реальный текст
ошибки терялся для мультиходового реплея — модель не знала, почему упало, и
повторяла ошибку).
- StepLike расширен полем content; новый хелпер normalizeToolError (Error/string/
object -> строка, обрезка через существующий compactValue/лимиты).
- serializeSteps: на каждый tool-error пушит парный {toolName, error} тем же
паттерном, что успешный {toolName, output} -> колонка tool_calls фиксирует сбой.
- assistantParts: при наличии tool-error эмитит output-error с РЕАЛЬНЫМ текстом;
заглушка остаётся только для по-настоящему непарных вызовов (прерванных).
- docs/reading-ai-logs.md обновлён под новую форму + cutover-оговорка.
Обратно совместимо: старые строки читаются как раньше, error-элемент аддитивен.
closes#407
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Убран отдельный столбец-маркер .definitionMarker (order:-1, min-width:1.5em),
дававший висячий отступ. Номер сноски теперь рисуется инлайн в начале первого
параграфа через .definitionContent > :first-child::before из CSS-переменной
--footnote-number (в модель документа не попадает, экспорт не затрагивает).
Кегль сносок уменьшен до var(--mantine-font-size-sm). Инвариант #146 (contentDOM
первый в DOM) и логика мульти-бэклинков #168 сохранены.
closes#420
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Read-tool outputs were compacted at a 4000-byte gate before being stored
in metadata.parts, which is re-sent to the model every later turn. Whole-page
reads (tens of KB) got shrunk to a 500-char preview plus a "[truncated N chars]"
marker; on the next turn the model read that marker as a source truncation and
re-read the page, wasting tokens and producing wrong behavior.
- Raise MAX_TOOL_OUTPUT_BYTES 4000 -> 200_000 so normal reads are stored and
replayed verbatim; only a single >200 KB output is compacted as a backstop.
- Reword the inline string marker so it reads as a replay-history elision, not
a source truncation, and tells the model it can re-call the tool.
- Update doc comments and enlarge the compactToolOutput unit-test inputs above
the new 200 KB gate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Researcher role wrote 40 literal `^[...]` and zero real footnotes: its
incremental write path (insertNode/editPageText) doesn't parse markdown, and
the footnote-capable tool was MCP-only. Promote three tools from inline
MCP-only to the shared registry so the in-app agent gets them too.
- tool-specs.ts: insertFootnote/insertImage/replaceImage added to
SHARED_TOOL_SPECS (mcpName/schema/description moved VERBATIM from the inline
registrations — MCP names + behaviour unchanged for external clients).
- index.ts: the 3 inline registerTool calls become registerShared; drop the
"MCP-only by design" comments.
- ai-chat-tools.service.ts: register the 3 in-app via sharedTool ->
client.insertFootnote/insertImage/replaceImage (imageUrl->url,
attachmentId->oldAttachmentId mapping).
- tool-tiers.ts: insertFootnote -> core (else the original asymmetry recurs —
footnote tool hidden while editPageText is core); images -> deferred.
- research/{en,ru}.yaml FOOTNOTES: `^[...]` parses ONLY on a whole-markdown
write (create/update/import); for a pinpoint citation to existing text use
insertFootnote; via editPageText/insertNode it stays literal.
- json-edit.ts guardrail: an edit_page_text `replace` containing a `^[...]`
token is refused into failed[] with an insert_footnote hint, mirroring the
existing formatting-marker refusal. (Slightly broader net than that mirror —
a literal `^[a-z]` regex class in a replace is also refused; accepted
defense-in-depth, has a no-false-positive test.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Follow-up to the merged MCP-hang fix (16b476a2); post-hoc review DO. The
fix itself is unchanged and prod-working — this adds the coverage + one
coherence fix the review asked for.
Tests (ai-chat.service.setup-abort.spec.ts):
- onLateResolve late-close: a toolset that resolves AFTER the setup race
was lost has its leased clients released (close spy asserted).
- pure 60s deadline (signal NOT aborted): the turn proceeds Docmost-only
(reaches streamText, run NOT finalized 'aborted') — the defense-in-depth
backstop, previously untested.
- legacy no-runId: a setup abort does NOT re-throw (the `runId &&` guard);
together with the deadline test this locks both halves of the catch guard.
Coherence (external-mcp/mcp-clients.service.ts buildEntry):
- The per-server connects now run via Promise.all instead of a sequential
for-await, so total build time is bounded by the slowest single server
(~2×CONNECT_TIMEOUT_MS) instead of the sum. At >6 all-timing-out servers
the sequential build exceeded the outer 60s deadline, inverting the
"per-server bound primary, outer deadline is backstop" invariant. Merge
is done in a sequential post-Promise.all loop in server order, so tool-key
precedence/disambiguation, outcomes and client ordering are byte-identical
to the sequential build; each server keeps its own timeout + failure-close.
Comments: the onLateResolve note now says it releases the lease (refcount),
not force-closes transports (the cache owns them); the invariant comment
reflects the parallel build.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A single NUL (U+0000) in model/tool output (e.g. a truncated multibyte
read of a web page) is rejected by Postgres in BOTH the content (text) and
toolCalls/metadata (jsonb) columns, so it failed EVERY write of the
streaming assistant row ("invalid input syntax for type json") and silently
dropped the turn's content from the DB while the live stream still showed it.
- add stripNulChars: deep-strips NUL from all strings, returns the same
reference when there is nothing to strip (no needless clone)
- apply it at the flushAssistant choke point (covers content + toolCalls +
metadata for the seed, per-step and terminal writes)
- tests: deep-strip, same-reference, end-to-end via flushAssistant
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reduce the default external-MCP silence timeout from 5 min to 1 min and the
overall call timeout from 15 min to 2 min. Update related tests and comments.
Increase Fastify's JSON body limit to 25 MiB (configurable via
HTTP_JSON_BODY_LIMIT) to accommodate large AI‑chat payloads.
BREAKING CHANGE: shorter MCP timeouts may abort long‑running tool calls that
previously succeeded with the older defaults.
DO-1: add an integration test (test/mock/tool-timing-server.test.mjs) that
constructs a real createDocmostMcpServer with a spy onMetric, links a Client
over InMemoryTransport, invokes get_workspace (no input schema, so the wrapped
handler always runs) and asserts onMetric fired with
("mcp_tool_duration_seconds", <number>, { tool: "get_workspace" }). This locks
that the monkeypatch wraps every tool AND labels with the registration name —
which the isolated timeToolHandler unit test does not. Mutation-verified
(mislabel -> test fails). The tool's expected backend failure (ECONNREFUSED)
is tolerated: the wrapper times in a finally on throw too, so the metric fires.
DO-2: reword the mcp.service.ts "zero overhead when disabled" comment to
"negligible overhead" — the registerTool wrapper still runs performance.now()
+ an async try/finally per tool call when onMetric is undefined (the
`onMetric?.()` short-circuits the label build; cost is immaterial at
tool-call rate), so "zero" was literally inaccurate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
packages/mcp stays free of prom-client/server: it only calls an optional
DocmostMcpConfig.onMetric(name, value, labels?) sink the host provides.
- client.ts: onMetric on the config; fires collab_connect_timeouts_total
once, only inside the 25s connect-timeout callback (the connect-vs-unload
signal). cleanup() clears the timer on every other finish path, so no
double-count.
- index.ts: createDocmostMcpServer monkeypatches server.registerTool (before
registerShared + inline tools are registered) to wrap every handler with
timeToolHandler — times in a finally on success AND throw, re-throws
unchanged, emits mcp_tool_duration_seconds{tool=<registered name>} (bounded
cardinality). Single choke point catches all tools.
- mcp.service.ts: the per-request config resolver injects onMetric ONLY when
isMetricsEnabled(), routing mcp_tool_duration_seconds -> observeMcpTool and
collab_connect_timeouts_total -> incConnectTimeout. Disabled / standalone
(stdio, no onMetric) -> undefined -> zero-overhead no-op.
New node:test unit (tool-timing.test.mjs) covers the wrapper's value/throw
preservation and the standalone no-op. packages/mcp/build/ is gitignored,
not committed (CI rebuilds it).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extends the #355 perf-metrics registry (all behind the METRICS_PORT hard
gate — nullable instruments, no-op helpers when unset). New families:
- collab_doc_load_duration_seconds{size_bucket} — onLoadDocument timed on
the real DB-load paths only (the already-loaded early-return is skipped).
- size_bucket added to collab_store_duration_seconds; storeDocument returns
the ydoc byte length (reusing its single Y.encodeStateAsUpdate, no second
encode) so onStoreDocument observes size without extra cost.
- collab_docs_open (gauge, read-on-scrape via collect() from
hocuspocus.getDocumentsCount() — never inc/dec'd, so it can't drift) +
collab_doc_loads_total / collab_doc_unloads_total (afterLoad/afterUnload).
- collab_connect_duration_seconds — onConnect->connected, correlated by a
WeakMap keyed on the shared request (leak-free; the current client uses
one socket per document).
- collab_auth_duration_seconds — wraps onAuthenticate (success and failure).
- sizeBucket() shared helper (lt64k|lt256k|lt1m|ge1m, 4 bounded values).
The mcp_tool_duration_seconds histogram + collab_connect_timeouts_total
counter helpers are registered here but wired in pass 2 (MCP callback).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A transient network blip during an external-MCP handshake left createMCPClient
pending forever (@ai-sdk/mcp does not settle on abort). getOrBuildEntry caches the
per-workspace build PROMISE, so the never-settling connect poisoned the cache and
every later turn hung at step_count=0 before streamText — the run never finalized,
the row stayed 'running', and the chat was permanently blocked with
A_RUN_ALREADY_ACTIVE (an explicit Stop could not interrupt the un-signalled setup).
- mcp-clients: wrap connect() in a settling timeout (connectWithTimeout) that
closes a late-arriving client, so a hung handshake rejects instead of poisoning
the build cache; the bad server is skipped and the build completes.
- mcp-clients: close a connected-but-unregistered client when tools() fails,
fixing a pre-existing transport leak in buildEntry.
- ai-chat.service: bound the toolsFor build with the run's abort signal AND a
deadline (raceAgainstAbortAndTimeout); re-throw an explicit Stop only when a run
exists (runId) so the run finalizes as 'aborted', and keep the legacy
socket-bound path unchanged; settle the run 'aborted' vs 'error' accordingly.
- tests: cache-not-poisoned + orphan-client-close-once + Stop-during-setup
finalizes the run once.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Review round: add the two missing test-locks the reviewer asked for —
(1) priority order of PRIMARY_INPUT_FIELDS is now pinned (`{query,title}`
-> "Q", so a reordering breaks the test); (2) the clamp boundary is pinned
exactly (140 chars -> unchanged, no ellipsis; 141 -> 140 + "…"), catching
an off-by-one / `>` vs `>=` regression. Test-only, no production change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
For tools without a friendly label (esp. external MCP tools like
Search_web_search) the card showed a generic "Ran tool <name>" with no
sense of what was searched. Add a compact one-line, dimmed summary of the
call's arguments under the label, pulled from part.input.
New pure toolInputSummary(part): picks the first present "primary" field
(query/q/searchQuery/url/urls/title/name/text/prompt), collapses
whitespace, clamps to ~140 chars; arrays render "first (+N)". It returns
undefined during input-streaming (the input grows while state is fixed and
messageSignature doesn't track input, so a live summary would freeze) —
the state flip to input-available re-renders the row with the final value,
so message-signature.ts is left untouched. The value renders ONLY through
Mantine <Text> (React-escaped) — no markdown/HTML, no XSS.
A showInput prop (default true) is threaded MessageList -> MessageItem
(+memo) -> ToolCallCard; the public share widget passes showInput={false}
so an anonymous reader never sees the agent's raw query text (mirrors
showCitations). No JSON fallback when no primary field is present.
closes#392
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Snapshot the editor selection at send time (same live-ref pattern as
openPageRef in prepareSendMessagesRequest), carry it nested inside
openPage so it dies with the page on a fail-closed resolve, and surface
it to the model only through the existing core getCurrentPage tool.
The selection TEXT is returned exclusively in the tool result (untrusted
collaborative-page content, treated as data by SAFETY_FRAMEWORK); the
system prompt gets only a fixed one-line flag, never the text/before/
after. sanitizeSelection caps text (4000), before/after (200), blockIds
(<=64 chars, <=20). Selection is a hint, not ground truth — the tool
description tells the agent to localize the fragment before editing.
closes#388
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Under a fully parallel `pnpm -r test` run the suite computed five separate
bcrypt cost-12 hashes (one per test, ~300ms idle, multi-second with all cores
saturated) and tripped jest's default 5s per-test timeout ("DISABLED user"
test, suite 31s under load vs 6s isolated).
- compute the hash ONCE in a top-level beforeAll (covers both describe
blocks) and share the read-only string across the five former call sites
- jest.setTimeout(30_000) at module scope so the per-test bcrypt compares
inside verifyUserCredentials get headroom under load too
No change to test names, assertions, order, or the CREDENTIALS_MISMATCH
contract semantics. Suite: 8/8 green, 4.8s isolated (was ~6s).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Do 1 [F7 regression]: транзиентный сбой attach больше не роняет строку и не
теряет ран. В transport fetch-wrapper `204 || !response.ok` оба зовут
onNoActiveStream (восстановить stripped-строку + invalidate + арм poll), и catch
зовёт его перед rethrow — раньше !ok/throw только сбрасывали флаг, и на 5xx/502/
network-blip in-progress ассистент-турн исчезал, durable-ран не отслеживался.
onNoActiveStream — суперсет (его часть-г всё ещё чистит флаг), идемпотентен.
Расширяет литеральный block-3 спеки (там был только сброс флага) — по ревью и
в согласии с интенцией окна «poll must survive a server restart».
Do 2 [stability]: attach-GET абортится при unmount + mount-гейтинг сайд-эффектов.
mountedRef: mount-эффект ре-армит true и в cleanup ставит false + abort
attachAbortRef; onNoActiveStream рано выходит на !mounted, onFinish-recovery
гейтится `wasResumed && mountedRef.current`. Снимает до-10-мин спурьёзный поллинг
+ чужую invalidateQueries + утёкший fetch на новооткрытом чате (и StrictMode
double-resume).
Do 3 [coherence]: anchor-mismatch не оставляет вечную dots-строку. Reconcile
после мержа хвоста мержит fresh-history версию stripped-строки, если её id !=
id хвоста — settl'ит осиротевшую streaming-A над раном B.
Тесты: F7 500 → restore+арм; F7 network-throw → restore+арм; unmount при pending
attach → abort + поздние колбэки не летят. vitest src/features/ai-chat 304
зелёных, grep-guard пуст.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Do 1 [test-coverage]: контроллерный тест на флаг-гейт begin-hook open() — при
OFF beginRun зовётся (durable-ран независим), а streamRegistry.open НЕ зовётся
(закрывает регресс: пустая entry → non-null paused attach → зависший SSE вместо
204); при ON — open зовётся с (chatId, runId).
Do 2 [stability]: байт-кап очереди pending paused-подписчика. pendingBytes +
overflowed на Subscriber; в paused-ветке ingestFrame при превышении
SUBSCRIBER_MAX_BUFFERED_BYTES (8MB) подписчик помечается overflowed, pending
чистится, он выбрасывается из entry.subscribers (как overflowed-entry). start()
на overflowed → onEnd (чистый 204-эквивалент, без частичного реплея). Контракт
«start() в том же тике, что attach()» задокументирован в коде — кап это
структурный бэкстоп для phase-2 Redis-await шва. Юнит-тест: paused A + live B,
9×1MB > cap → A выброшен (0 доставок), B получает все 9 живьём, поздний start(A)
→ один onEnd без реплея.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
PR 1 of 2 (#381, фаза 1.5 #184): серверный реестр SSE-стримов агентских ранов,
чтобы любая вкладка могла подключиться к живому рану с реплеем кадров + живым
хвостом. «Спящий» — весь провод за флагом AI_CHAT_RESUMABLE_STREAM (off по
умолчанию); клиент (PR 2) ещё не написан.
- ai-chat-stream-registry.service.ts: in-memory реестр (open/bind/abortEntry/
attach). attach — снапшот+подписка в ОДНОМ синхронном блоке (инвариант 4: нет
await между `subscribers.add` и `frames.slice()`), paused-подписчик, overflow,
retention с identity-guard (инвариант 2), open поверх live entry даёт ровно
один onEnd (инвариант 3), anchor против кросс-ранового реплея (инвариант 6).
- ai-chat.controller.ts: begin-хук open(chatId, runId) + GET-attach эндпоинт
(403 чужой чат; 204 нет-entry/finished/anchor-мисматч; cleanup до первой
записи + recheck req.raw.destroyed; cap→destroy).
- ai-chat.service.ts: tee SSE-кадров в реестр (consumeSseStream + generateMessageId,
гейт на runId && flag) + abortEntry из внешнего catch.
- environment.service.ts: флаг isAiChatResumableStreamEnabled().
Флаг OFF ⇒ байт-в-байт legacy И #184-фаза-1 (нет start.messageId, нет tee).
Инжектируемые провайдеры НЕ @Optional() → поломка вайринга роняет старт, а не
тихо выключает фичу.
Тесты: registry unit (16), controller.attach (9), service pipe-options (4, вкл.
flag-off-with-runId негатив), int-spec ai-chat-attach (6, реальный MockLanguageModelV3).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- F5 [stability/regression]: the round-1 F2 fix re-rendered the root with
<PostHogProvider><App/></PostHogProvider> after the analytics chunk loaded. In
the ChunkLoadErrorBoundary child slot the element TYPE changes App ->
PostHogProvider, so React does NOT reconcile in place — it REMOUNTS the whole
App: every mount effect runs twice (websocket connect/disconnect, origin
tracking, subscriptions) and local state / focus / scroll / in-progress input is
lost on cloud cold-load (e.g. typing in /login before analytics loads). And it
was USELESS: the app has ZERO consumers of the PostHog React context (no
usePostHog / useFeatureFlag* / PostHogFeature), and PostHogProvider given an
initialized client is a no-op — all capture goes through the posthog singleton.
Fix: initAnalytics now inits the posthog SINGLETON only (no posthog-js/react
import, no second render); renderApp() renders <App/> once. First paint stays
instant, cloud analytics behavior unchanged, no remount.
- F6 [test]: exported isChunkLoadError + chunk-load-error-boundary.test.ts —
pins the detector (ChunkLoadError name + the 3 dynamic-import failure messages,
case-insensitive → true; null/undefined/ordinary errors → false) so a
false-negative that re-blanks the app on a real chunk-404 is caught.
Gate: client tsc 0, chunk-load + sanitize tests 14 passed. Entry chunk unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- F1 [HIGH]: added a root ChunkLoadErrorBoundary (react-error-boundary) wrapping
the routed app in main.tsx, ABOVE all the route-level/Aside/AiChatWindow
Suspense boundaries. A stale-deploy chunk 404 (React.lazy reject) is caught and
auto-reloads once (sessionStorage-guarded against a reload loop), else shows a
manual "new version available" reload UI — instead of unmounting the whole tree
to a white screen. Existing per-feature ErrorBoundaries untouched.
- F2 [MED-HIGH]: posthog no longer blocks/blanks the cloud first paint. main.tsx
now renders <App/> immediately for everyone, then `void initAnalytics()` — which
keeps the exact cloud gate, dynamically imports posthog, and RE-RENDERS the same
React root wrapped in PostHogProvider (React reconciles onto the painted DOM, so
cloud ends up wrapped exactly as before). The import+init is try/catch'd: a
failed analytics chunk (network / stale-404 / ad-blocker on a "posthog" chunk)
degrades to no-analytics instead of a permanently blank page.
- F3: sanitize-url.test.ts mirroring editor-ext's security contract (javascript:/
data:/vbscript:/obfuscated → ""; https/relative/mailto preserved).
- F4: the idle-warm `void import(...)` prefetch in layout.tsx gets `.catch(()=>{})`
so a failed best-effort prefetch can't surface as an unhandledrejection.
No new deps (lockfile unchanged). Gate: client tsc 0, sanitize test 3/3, client
build succeeds (entry chunk still 556K, posthog in separate dynamic chunks).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
agent_coder
vvzvlad
agent_vscode
agent_coder