feat(ai-chat): ретраи finalizeAssistant + owner-write приоритет + двусторонний reconcile (#487)

Раньше finalizeAssistant ставил finalized=true ДО записи и не ретраил → один
неудачный UPDATE = строка вечно 'streaming'; свип был boot-only; run-свип
безусловный — асимметрия «run succeeded / message streaming навсегда».

- finalizeAssistant: bounded-ретраи; once-гейт закрывается ТОЛЬКО после успешной
  записи; возвращает ok. Правило owner-write: терминальная запись owner'а
  условна на status='streaming' OR metadata.finalizeFailed (repo.finalizeOwner) —
  перетирает reconcile-штамп, но не проставленный терминал. ВСЕ status-only
  штампы reconcile (stampTerminalIfStreaming, sweepStreaming) пишут строго
  onlyIfStreaming И мёржат metadata.finalizeFailed:true (иначе поздний owner-write
  не перетрёт).
- Порядок: попытка message-finalize → ран финализируется ВСЕГДА; при провале
  message onFinish помечает ран 'error' (не 'completed'). Ран не гейтится на
  message.
- Периодический reconcile-джоб (setInterval, env-tunable) клаузами по порядку:
  (a) пере-драйв зомби; (b) message streaming + ран терминален → штамп по статусу
  рана (succeeded-ран + зависшая строка → 'aborted'+finalizeFailed, НЕ
  'completed'-empty); (c) run running + НЕТ entry И НЕТ zombie + staleness →
  aborted (гейт «нет entry» первичный, staleness от last-progress updatedAt,
  X=max(2×per-call cap,15мин)+boot-warn); (d) message streaming + возраст>X + нет
  активной run-строки → aborted (двойной гейт). isInterruptResume исключает
  finalizeFailed-строки. Оппортунистический одно-чатовый reconcile при старте хода
  (best-effort, не фейлит ход). sweepStreaming boot-only → периодический.

Тесты (реальная БД): owner finalizeOwner чистит finalizeFailed; штамп не перетирает
терминал; поздний owner-write перетирает aborted-штамп; клаузы b/c/d (+живой entry
не трогать, двойной гейт d); «убить БД на finish → после восстановления ни строка,
ни ран не застряли». Юниты: finalizeFailed исключает interrupt-resume;
reconcileStaleRuns «нет entry».

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-07-11 05:13:35 +03:00
parent 307ad49324
commit 002c931c6b
9 changed files with 889 additions and 45 deletions
@@ -1,5 +1,6 @@
import { Injectable, Logger } from '@nestjs/common';
import { InjectKysely } from 'nestjs-kysely';
import { sql } from 'kysely';
import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
import { dbOrTx } from '../../utils';
import {
@@ -188,6 +189,144 @@ export class AiChatMessageRepo {
return query.returning(this.baseFields).executeTakeFirst();
}
/**
* #487 OWNER terminal write — the streamText terminal callback's finalize. Like
* `update` but CONDITIONAL on `status='streaming' OR metadata.finalizeFailed`:
* the owner writes its real content EITHER when the row is still streaming (the
* normal case) OR when a reconcile stamp already flipped it to a terminal status
* but marked `finalizeFailed:true` — the owner's real content OVERWRITES that
* placeholder stamp (owner-write priority, #487). A row that is properly terminal
* (no finalizeFailed) is left untouched (undefined) — idempotent. The `patch`
* carries the real metadata WITHOUT finalizeFailed, so a successful write CLEARS
* the flag. Returns the updated row, or undefined when nothing matched.
*/
async finalizeOwner(
id: string,
workspaceId: string,
patch: Partial<{
content: string | null;
toolCalls: unknown;
metadata: unknown;
status: string | null;
}>,
trx?: KyselyTransaction,
): Promise<AiChatMessage | undefined> {
const db = dbOrTx(this.db, trx);
return db
.updateTable('aiChatMessages')
.set({ ...(patch as Record<string, unknown>), updatedAt: new Date() })
.where('id', '=', id)
.where('workspaceId', '=', workspaceId)
.where((eb) =>
eb.or([
eb('status', '=', 'streaming'),
eb(sql<string>`(metadata->>'finalizeFailed')`, '=', 'true'),
]),
)
.returning(this.baseFields)
.executeTakeFirst();
}
/**
* #487 RECONCILE status-only stamp — settle a stuck 'streaming' row to a
* terminal status WITHOUT the owner's real content (which lived only in the
* dead process's memory — a documented loss). CONDITIONAL on `status='streaming'`
* (never touches an already-terminal row) AND it MERGES `finalizeFailed:true`
* into metadata (preserving the partial `parts` already persisted) so a LATER
* owner-write (finalizeOwner) can still OVERWRITE this placeholder with real
* content, and so `isInterruptResume` can EXCLUDE this row (a reconcile stamp is
* not a genuine user interruption). Returns the updated row, or undefined.
*/
async stampTerminalIfStreaming(
id: string,
workspaceId: string,
status: 'aborted' | 'error' | 'completed',
trx?: KyselyTransaction,
): Promise<AiChatMessage | undefined> {
const db = dbOrTx(this.db, trx);
return db
.updateTable('aiChatMessages')
.set({
status,
metadata: sql`coalesce(metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
updatedAt: new Date(),
})
.where('id', '=', id)
.where('workspaceId', '=', workspaceId)
.where('status', '=', 'streaming')
.returning(this.baseFields)
.executeTakeFirst();
}
/**
* #487 reconcile clause (b): streaming assistant rows whose linked RUN has
* already reached a terminal status — an asymmetry ("run settled / message
* streaming forever") the periodic reconcile heals by stamping the message.
* Returns the message id + its run's terminal status, bounded.
*/
async findStreamingWithTerminalRun(
limit = 200,
// #487: scope to ONE chat for the opportunistic per-turn reconcile (removes
// reconcile latency from the user-visible path); omit for the periodic sweep.
chat?: { chatId: string; workspaceId: string },
): Promise<
Array<{ messageId: string; workspaceId: string; runStatus: string }>
> {
let query = this.db
.selectFrom('aiChatMessages as m')
.innerJoin('aiChatRuns as r', 'r.assistantMessageId', 'm.id')
.select([
'm.id as messageId',
'm.workspaceId as workspaceId',
'r.status as runStatus',
])
.where('m.status', '=', 'streaming')
.where('r.status', 'in', ['succeeded', 'failed', 'aborted']);
if (chat) {
query = query
.where('m.chatId', '=', chat.chatId)
.where('m.workspaceId', '=', chat.workspaceId);
}
return query.limit(limit).execute();
}
/**
* #487 reconcile clause (d) — historical-row safety: streaming rows older than
* `staleMs` whose chat has NO active run row (double-gated). Settle them to
* 'aborted' + finalizeFailed (so a late owner-write could still overwrite).
* Returns the count. Used ONLY by the periodic reconcile, never at boot.
*/
async sweepStreamingWithoutActiveRun(
staleMs: number,
trx?: KyselyTransaction,
): Promise<number> {
const db = dbOrTx(this.db, trx);
const staleBefore = new Date(Date.now() - staleMs);
const rows = await db
.updateTable('aiChatMessages as m')
.set({
status: 'aborted',
metadata: sql`coalesce(m.metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
updatedAt: new Date(),
})
.where('m.status', '=', 'streaming')
.where('m.updatedAt', '<', staleBefore)
.where((eb) =>
eb.not(
eb.exists(
eb
.selectFrom('aiChatRuns as r')
.select('r.id')
.whereRef('r.chatId', '=', 'm.chatId')
.where('r.status', 'in', ['pending', 'running']),
),
),
)
.returning('m.id')
.execute();
return rows.length;
}
/**
* Crash-recovery sweep (#183): flip every assistant row still left in the
* 'streaming' state (a turn that died mid-write before reaching a terminal
@@ -200,13 +339,20 @@ export class AiChatMessageRepo {
* step, so an actively-streaming row never matches; this prevents a fresh
* replica's boot-sweep from aborting a turn another replica is still streaming
* in a multi-instance deploy.
*
* #487: the sweep now ALSO marks `finalizeFailed:true` so a late owner-write can
* overwrite this placeholder with real content (owner-write priority).
*/
async sweepStreaming(trx?: KyselyTransaction): Promise<number> {
const db = dbOrTx(this.db, trx);
const staleBefore = new Date(Date.now() - SWEEP_STREAMING_STALE_MS);
const rows = await db
.updateTable('aiChatMessages')
.set({ status: 'aborted', updatedAt: new Date() })
.set({
status: 'aborted',
metadata: sql`coalesce(metadata, '{}'::jsonb) || jsonb_build_object('finalizeFailed', true)`,
updatedAt: new Date(),
})
.where('status', '=', 'streaming')
.where('updatedAt', '<', staleBefore)
.returning('id')
@@ -219,6 +219,31 @@ export class AiChatRunRepo {
* sweeps only runs UNTOUCHED past the window. Phase 1 is single-process, so the
* boot path supplies no window.
*/
/**
* #487 reconcile clause (c): active (pending|running) runs UNTOUCHED past
* `staleMs` — candidates for "no live runner" abort. Staleness is measured from
* `updated_at` (the LAST-PROGRESS timestamp — recordStep bumps it), NOT
* `started_at`, so a legitimate long-running marathon (11–25 min of steady
* progress) is never a candidate. The caller filters these against its in-memory
* `active` / zombie maps ("no entry" is the PRIMARY gate — a live entry is never
* aborted) before settling any of them. Bounded.
*/
async findStaleActive(
staleMs: number,
limit = 200,
trx?: KyselyTransaction,
): Promise<Array<{ id: string; workspaceId: string; chatId: string }>> {
const db = dbOrTx(this.db, trx);
const staleBefore = new Date(Date.now() - staleMs);
return db
.selectFrom('aiChatRuns')
.select(['id', 'workspaceId', 'chatId'])
.where('status', 'in', ACTIVE_RUN_STATUSES as unknown as string[])
.where('updatedAt', '<', staleBefore)
.limit(limit)
.execute();
}
async sweepRunning(
opts: { staleMs?: number } = {},
trx?: KyselyTransaction,