Hacktoberfest 2026: những issue maintainer đã đánh dấu cho tháng Mười, đang mở và phù hợp người mới. Xem issue Hacktoberfest

withPersistence stores the durability batch hint on a stopped or failed assistant message

Đang mở
  • Có gì không ổn?
#1,802 0 bình luận 0 reaction 1 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 1 ngày

@tombeckenham đang làm issue này rồi.

Từ ngày 16/10/2026.

  • #1805 của @ozolcar — đang mở

Đánh giá

Issue này chưa được đánh giá.

Mô tả

has-pr waiting-on: maintainer
TanStack AI version

@tanstack/ai 0.69.0, @tanstack/ai-persistence 0.9.0 (the same code is on main at 85c2956)

Framework/Library version

Node 26.3.1, server only (no UI framework involved)

Describe the bug and the steps to reproduce it

With snapshotStreaming: true, withPersistence keeps a metadata.tanstack.batch: true key on the stored assistant message when a run is stopped or fails after a middleware called ctx.emitCustomEvent(name, value, { batch: true }).

This is about the message store (stores.messages, which withPersistence writes and reconstructChat reads). It is not about the durability log, the delivery stream that toServerSentEventsResponse({ durability }) appends to. They are separate stores.

The batch hint is an in-process flag for the durability producer. #1355 says it "is stripped before the log and the wire", and stripDurabilityBatchHint does that in toServerSentEventsResponse and in the strip-to-spec middleware. That covers the durability log and the wire. flush() in stream-to-response.ts strips each chunk before durability.append and before forwarding it. The emitted CUSTOM chunk also passes through the middleware chain, though, before any of that. In withPersistence, onChunk merges chunk.metadata of every parent chunk into the streaming snapshot metadata (packages/ai-persistence/src/middleware.ts, the block under "Keep the call metadata (source, response id) for the snapshot rows."). The hint goes in with it. Every later snapshot row carries it. On Stop or on an error, updateStreamingRows(ctx, 'aborted' | 'error') writes the merged metadata to the final stored row, so the hint stays in storage and comes back from reconstructChat. A run that finishes normally is clean, because the final save replaces the snapshot rows.

Steps:

  1. Install @tanstack/[email protected] and @tanstack/[email protected].
  2. Save the script below as repro.mjs in an ESM package and run node repro.mjs.

repro.mjs:

import { chat, defineChatMiddleware } from '@tanstack/ai'
import { memoryPersistence, withPersistence } from '@tanstack/ai-persistence'

const controller = new AbortController()

const adapter = {
  kind: 'text',
  name: 'fixed',
  model: 'test-model',
  '~types': {},
  async *chatStream() {
    const t = Date.now()
    yield { type: 'RUN_STARTED', threadId: 'thread-1', runId: 'run-1', timestamp: t }
    yield { type: 'TEXT_MESSAGE_START', messageId: 'msg-1', role: 'assistant', timestamp: t }
    yield { type: 'TEXT_MESSAGE_CONTENT', messageId: 'msg-1', delta: 'partial', timestamp: t }
    controller.abort() // the user presses Stop
  },
}

const progress = defineChatMiddleware({
  name: 'progress',
  onChunk(ctx, chunk) {
    if (chunk.type === 'TEXT_MESSAGE_CONTENT') {
      ctx.emitCustomEvent('progress', { step: 1 }, { batch: true })
    }
  },
})

const persistence = memoryPersistence()
const stream = chat({
  adapter,
  messages: [{ role: 'user', content: 'hi' }],
  threadId: 'thread-1',
  runId: 'run-1',
  abortController: controller,
  middleware: [
    withPersistence(persistence, { snapshotStreaming: true, snapshotIntervalMs: 0 }),
    progress,
  ],
})
for await (const _ of stream) void _

const stored = await persistence.stores.messages.loadThread('thread-1')
console.log(JSON.stringify(stored.at(-1).metadata))

Expected

{"tanstack":{"source":{"provider":"fixed","api":"text","model":"test-model"},"stopReason":"aborted"}}

Actual

{"tanstack":{"source":{"provider":"fixed","api":"text","model":"test-model"},"batch":true,"stopReason":"aborted"}}

If the adapter throws instead of the abort, the row ends with "batch":true,"stopReason":"error". A snapshot row read in the middle of the run (after a second TEXT_MESSAGE_CONTENT) is {"tanstack":{"source":{...},"batch":true}}.

I found nothing that reads tanstack.batch on a stored message, so the visible effect is a stray internal key in the message store and in the hydrated messages. Any store that validates message metadata will see an unknown key. The hint is meant to stay in process, as the durability log and the wire already show, but the message store keeps it.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

Inline above. Plain Node script, no network, no model.

Screenshots or Videos (Optional)

None.

Do you intend to try to help solve this bug with your own PR?

Yes, I am also opening a PR that solves the problem along side this issue

Terms & Code of Conduct
  • I agree to follow this project's Code of Conduct
  • I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.
Ngôn ngữ chính
TypeScript
Star
3.2k
Fork
365
Merge trung bình
1 ngày 9 giờ
Pull request đã merge (30 ngày)
249

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của TanStack/ai

Tất cả issue của TanStack/ai

Issue tương tự

Thêm issue về TypeScript

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.