functionCallResult never get's called in the frontend with ChatCompletionStream

[cosbgn]

Confirm this is a Node library issue and not an underlying OpenAI API issue

• This is an issue with the Node library

Describe the bug

In my backend I do:

const steam = client.beta.chat.completions.runTools({})
return steam.toReadableStream()

And in my frontend I do:

const runner = ChatCompletionStream.fromReadableStream( response.body )
runner.on('message', (message) => {
    messages.value.push(message)
})

This works but after the role:assistant with tool_calls it doesn't push the tool_call results to the messages array. In the frontend also runner.on('functionCallResult') never gets called.

This means that next calls fail because after a tool call we must pass the tool call results.

How can I get the tool call function response to show up in messages?

To Reproduce

1. return a stream to the frontend:

const steam = client.beta.chat.completions.runTools({})
return steam.toReadableStream()

2. read the stream with ChatCompletionStream

3. fail to get the function response as message

P.s I've also tested with ChatCompletionStreamingRunner and it's the same

Code snippets

No response

OS

mac latest

Node version

22

Library version

latest

[cosbgn]

This is the same error: https://community.openai.com/t/beta-completion-runtools/531973

[cosbgn]

After some investigation, when role=="tool" the message never gets added to the stream.

in other words this will log only in the backend and never in the frontend.

	runner.on("message", (message) => {
		if (message.role === "tool"){
			console.log(message) // Works on backend but not on frontend
		}
	})

Is there a way to manually push it as a chunk? Is there anything I can do to at least send it as extra data and then somehow add it back in the correct index in my messages in the frontend? I really need this data and openai will anyways crush on my next message as it doesn't have the toolCall respose.

[RobertCraigie]

@cosbgn thanks for the report, I think you might be using the wrong stream class on the frontend, can you try using ChatCompletionStreamingRunner.fromReadableStream() instead?

[cosbgn]

Hi, I tried all possible classes and also to simply read the chucks from the stream without any class. The tool response never gets sent to the frontend, I think it's an issue with the runTool implementation.

[blechatellier]

Running into the same issue, I can create a custom ReadableStream and manually enqueue the function result but not sure about the shape of the chunk so that it gets read correctly on the frontend. @RobertCraigie any pointers please?

[RobertCraigie]

Thanks for reporting, I can reproduce the issue.

@blechatellier unfortunately I'm not currently aware of a full workaround, I think you'd have to manually enqueue chunks in your own format and then call runner._addMessage({ role: 'tool', name, content }); manually to have the functionCallResult event emitted.

We'll look into fixing this.

[cosbgn]

@RobertCraigie I'm trying something like this in the backend:

.on("message", async (message) => {
    if (message.role === "tool" && !added_tool_call_ids.includes(message.tool_call_id)) {
      added_tool_call_ids.push(message.tool_call_id)
      runner._addMessage(message)
    }

However it fails because I believe it has already the role=tool in it:

400 Invalid parameter: messages with role 'tool' must be a response to a preceeding message with 'tool_calls'.

Do you have an example on how I could get this to work?