We've also seen this issue, it happens when an SSE event is larger than bufio.MaxScanTokenSize (64k) . Incidently the newly released gpt-4.1 seems to generate tokens fast enough to go over that maximum fairly easily.

This is the construction site for the Scanner instance being overflowed:

Would you be able to add an option to pass a desried max size as a RequestOption?

Experiencing same issue here with GPT-4.1, specifically with the response.output_item.done event.

For now, just doing:

// Check for any stream errors
if err := stream.Err(); err != nil {
	if stream.Current().Type == string(constant.ResponseOutputItemDone("").Default()) {
		return
	}
	logger.Error(ctx, "Streaming error: %v", err)
	errorChannel <- fmt.Errorf("streaming error: %w", err)
}