ai-harness: recovery runs a turn without the user message when the turn never started #1795

TanStack AI version

@tanstack/ai-harness on the feat/harness-core branch (not released yet).

Framework/Library version

Node 24.8.0. No UI framework: the bug is in the durable host.

Describe the bug and the steps to reproduce it

A durable host (stores.log) stores a prompt in the log, then stops before the turn starts (before it writes its run lease). The next host that opens the thread recovers the input and runs the turn. But the transcript has no user message for that input, so the model gets an empty message list.

Expected: the transcript is ['First.', 'One.']. Actual: ['One.'], and the model call has no messages.

Steps:

1. Open a durable host whose runs.createOrResume never returns, so the turn never starts.
2. Send prompt('First.') and wait for the receipt.
3. Stop the host's claim renewal, and wait until the claim expires.
4. Open the thread on a second host. It recovers the input and calls the model.
5. Read transcript() on the second host.

Found during the review of #1780. That PR is not the cause: the same test fails on feat/harness-core before #1780.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

A test in packages/ai-harness/tests/:

import { expect, it, vi } from 'vitest'
import { memoryLogStore, memoryPersistence } from '@tanstack/ai-persistence'
import { createHarnessHost, defineHarness } from '../src'
import { mockAdapter, text } from './helpers'

it('recovery keeps the user message of a turn that never started', async () => {
  const { runs, metadata, workClaims } = memoryPersistence().stores
  const persistence = {
    stores: { log: memoryLogStore(), runs, metadata, workClaims },
  }
  const { adapter, calls } = mockAdapter([() => text('One.')])
  const harness = defineHarness({ name: 'probe', adapter })
  const owner = await createHarnessHost({
    lease: { ttlMs: 200, renewMs: 50 },
    persistence: {
      stores: {
        ...persistence.stores,
        runs: Object.assign(Object.create(runs), {
          createOrResume: () => new Promise(() => {}),
        }),
      },
    },
  }).open(harness, { threadId: 't1' })
  await owner.prompt('First.').receipt
  ;(owner as unknown as { stopClaimTimer: () => void }).stopClaimTimer()
  await new Promise((resolve) => setTimeout(resolve, 300))

  const next = await createHarnessHost({ persistence }).open(harness, {
    threadId: 't1',
  })
  await vi.waitFor(() => expect(calls).toHaveLength(1), { timeout: 3_000 })
  await new Promise((resolve) => setTimeout(resolve, 200))
  expect((await next.transcript()).map((message) => message.content)).toEqual([
    'First.',
    'One.',
  ])
})

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

Maybe, I'll investigate and start debugging

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.