Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Durable resume through withPersistence never fills ctx.inputResponse for mcp_input interrupts

未关闭
  • 哪里不对？
#1,885 0 条评论 0 个 reaction 已指派 1 人 在 GitHub 查看

维护者通常 1 天内回复

@tombeckenham 已经在做这个了。

开始于 2026年10月23日。

评估

这个 Issue 还没有评估数据。

描述

waiting-on: maintainer
TanStack AI version

@tanstack/ai 1.0.0 (also seen in 0.69.0)

Framework/Library version

Node.js 24, server-side chat() with withPersistence; no framework involved

Describe the bug and the steps to reproduce it

A server tool that pauses on an input request gets its answer on ctx.inputResponse when the next run is resumed from the messages the client sends. When the run is resumed through withPersistence (the durable path, messages: [] and the stores rebuild the thread), the answer never reaches ctx.inputResponse. The tool is called again with inputResponse undefined, so it asks the same question again.

Steps

  1. A server tool throws an input request on its first call and reads the answer on the second:
import { toolDefinition } from '@tanstack/ai'
import { z } from 'zod'

const askCity = toolDefinition({
  name: 'ask_city',
  description: 'Ask the user which city to use',
  inputSchema: z.object({}),
}).server(async (_args, ctx) => {
  if (ctx.inputResponse?.status === 'resolved') {
    return { city: (ctx.inputResponse.payload as { value: string }).value }
  }
  throw Object.assign(new Error('Input required'), {
    name: 'MCPInputRequiredError',
    kind: 'form',
    request: {
      message: 'Which city?',
      requestedSchema: {
        type: 'object',
        properties: { value: { type: 'string' } },
        required: ['value'],
      },
    },
  })
})

  1. Run it under withPersistence. The first run ends with RUN_FINISHED and an interrupt outcome whose item id is mcp_input_<toolCallId>, and the stores park it.

  2. Resume with the answer, the way a durable continuation does:

chat({
  adapter,
  messages: [], // the stores own the history
  threadId,
  runId: nextRunId,
  parentRunId: interruptedRunId,
  resume: [
    {
      interruptId: `mcp_input_${toolCallId}`,
      status: 'resolved',
      payload: { value: 'Paris' },
    },
  ],
  tools: [askCity],
  middleware: [withPersistence({ /* stores */ })],
})

Expected: askCity is called again with ctx.inputResponse set to { status: 'resolved', payload: { value: 'Paris' } }, as it is on the ephemeral resume.

Actual: ctx.inputResponse is undefined, so the tool throws the same input request again and the run parks a second time.

Where it happens

I found this by reading the source, not by bisecting. In src/activities/chat/index.ts at 1.0.0:

  • applyEphemeralInterruptResume (around line 5253) collects the mcp_input_* ids of the pending server tool calls and copies each resolution into resumeInputResponses (around line 5427). That map is what executeToolCalls reads to set ctx.inputResponse.
  • applyResumeToolState (around line 5614), which applies the resumeToolState that middleware such as withPersistence hands back from onConfig, copies genericInterrupts into resumeGenericInterrupts but never into resumeInputResponses.

So the mcp_input_* resolution survives the durable resume as a generic interrupt resolution, but the tool it belongs to never sees it. A possible fix is for the durable path to map mcp_input_<toolCallId> resolutions to resumeInputResponses the same way the ephemeral path does.

Workaround: read the resolution out of the stores and pass it to the tool through chat({ context }), keyed by tool call id.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

No sandbox; the snippets above are the minimal shape, and the source locations show the gap.

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

No, because I don't have time to dive into it

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.
主要语言
TypeScript
星标
3.2k
派生
372
平均合并
1 天 11 小时
30 天内合并 PR
274

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

TanStack/ai 的其他 Issue

查看 TanStack/ai 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。