Negative initialStartIndex can resume in the middle of a UI chunk sequence and break AI SDK chat state

[joetifa2003]

Summary

When using resumable streams with WorkflowChatTransport and a negative initialStartIndex (for example -50), the reconnect response can start in the middle of a logical UI-message part sequence.
That causes the AI SDK client stream processor to throw errors like:
Received reasoning-delta for missing reasoning part with ID "reasoning-0". Ensure a "reasoning-start" chunk is sent before any "reasoning-delta" chunks.
This happened even though the reconnect endpoint correctly returned x-workflow-stream-tail-index as described in the docs.

Versions

• workflow: 4.2.4
• @workflow/ai: 4.1.2
• ai: 6.0.168
• @ai-sdk/react: 3.0.170

What I expected

I expected negative resume indices to be safe to use for resumable chat streams, as long as:

• the client uses initialStartIndex: -50
• the reconnect route returns x-workflow-stream-tail-index
• the route returns getRun(runId).getReadable({ startIndex })

What actually happened

The reconnect stream started with deltas for an already-open logical part, instead of the matching start chunk.
Example from Chrome SSE output:

data: {"type":"reasoning-delta","id":"reasoning-0","delta":" space in the header"}
data: {"type":"reasoning-delta","id":"reasoning-0","delta":".\n\nNow I"}
...
data: {"type":"reasoning-end","id":"reasoning-0"}
data: {"type":"tool-input-start","toolCallId":"call_exjjt7cc","toolName":"grep"}
...
data: {"type":"finish-step"}
data: {"type":"tool-output-available", ...}
data: {"type":"start-step"}
data: {"type":"reasoning-start","id":"reasoning-0"}

The AI SDK client then throws because it receives reasoning-delta before reasoning-start.

Why this seems surprising

The current docs around negative resume indices talk about transport/chunk offsets and x-workflow-stream-tail-index, but they do not mention that:

• a negative chunk index can land in the middle of a text-* sequence
• or in the middle of a reasoning-* sequence
• or in the middle of a tool-input-* sequence
  The AI SDK UI stream processor enforces those semantic boundaries, so raw chunk-index resume is not always safe for UI chunk replay.

Minimal setup

Client:

new WorkflowChatTransport({
  api: "/api/chat",
  initialStartIndex: -50,
})

Reconnect route:

const run = getRun(params.id)
const stream = run.getReadable({
  startIndex: startIndex ? parseInt(startIndex, 10) : undefined,
})
const tailIndex = await stream.getTailIndex()
return createUIMessageStreamResponse({
  stream,
  headers: {
    "x-workflow-run-id": params.id,
    "x-workflow-stream-tail-index": String(tailIndex),
  },
})

Workaround I had to implement
On the server, when startIndex < 0, I had to:

• resolve the negative index against the tail index
• scan backward a bit
• rewind to the previous safe boundary such as start or start-step
• only then return the stream
  That fixed the issue.

Request

Can the docs and/or library behavior be clarified here?

Possible directions:

• document that negative resume indices can cut into the middle of UI part sequences
• recommend rewinding to a safe boundary before returning resumed streams
• or make the transport/runtime handle this automatically for UI chunk streams

---

This was reproduced with DurableAgent streaming UI chunks directly to the chat client.

[VaguelySerious]

Back to working on this, making some changes in #2082 that should help doing partial resumptions

[joetifa2003]

This is great! nice work

But a question i have is, DurableAgent will get replaced with WorkflowAgent from ai-sdk right?
I don't know if you are going to implement the same thing there also or not.

[VaguelySerious]

Automated port to WorkflowAgent: vercel/ai#15543