[BUG] StreamingResponseErrorUpdate deserializing incorrectly

[stephentoub]

Describe the bug

The deserialization code is looking for top-level "code", "message", and "param" properties (which maps to how it's documented in the OpenAI API reference docs). But in the actual JSON OpenAI sends back, those properties are nested on an "error" object.

Steps to reproduce

#pragma warning disable OPENAI001
#pragma warning disable SCME0001

using OpenAI.Responses;
using System.Text;

var client = new OpenAIResponseClient("gpt-4o-mini", Environment.GetEnvironmentVariable("AI:OpenAI:ApiKey"));

string input = string.Join(". ", Enumerable.Repeat("This is a test input that is way too long", 128_000));

await foreach (var update in client.CreateResponseStreamingAsync(input))
{
    if (update is StreamingResponseErrorUpdate errorUpdate)
    {
        Console.WriteLine($"{errorUpdate.Code}: {errorUpdate.Message}");

        var data = update.Patch.GetJson("$."u8);
        Console.WriteLine(Encoding.UTF8.GetString(data));
    }
}

That outputs:

:
{"type":"error","sequence_number":2,"error":{"type":"invalid_request_error","code":"context_length_exceeded","message":"Your input exceeds the context window of this model. Please adjust your input and try again.","param":"input"}}

Code snippets

OS

Windows

.NET version

.NET 10

Library version

2.7.0

[jsquire]

Hi @stephentoub. Thanks for reaching out and we regret that you're experiencing difficulties. In this case, it appears that the OpenAI service is not honoring its public contract and is intermittently returning an incorrect error response. The platform docs (Streaming Response Error), and all of the official REST API specs agree that the format for a streaming error response is:

{
    "type": "error"
    "sequence_number": 5
    "code": ...
    "message": "..."
    "param": ...
}

In the official OpenAPI spec, this is represented by ResponseErrorEvent on line 56,363:

 ResponseErrorEvent:
      type: object
      description: Emitted when an error occurs.
      properties:
        type:
          type: string
          description: |
            The type of the event. Always `error`.
          enum:
            - error
          x-stainless-const: true
        code:
          anyOf:
            - type: string
              description: |
                The error code.
            - type: 'null'
        message:
          type: string
          description: |
            The error message.
        param:
          anyOf:
            - type: string
              description: |
                The error parameter.
            - type: 'null'
        sequence_number:
          type: integer
          description: The sequence number of this event.
      required:
        - type
        - code
        - message
        - param
        - sequence_number
      x-oaiMeta:
        name: error
        group: responses
        example: |
          {
            "type": "error",
            "code": "ERR_SOMETHING",
            "message": "Something went wrong",
            "param": null,
            "sequence_number": 1
          }

The same expected format is represented by the TypeSpec version(ResponseErrorEvent) that the client is generated from. Because the errant behavior is on the service-side, the client library has no insight nor influence over it. The maintainers of the OpenAI client library are unable to assist beyond reporting to the OpenAI service team, which we've done.

Unfortunately, OpenAI does not offer service support from this repository. The most reliable and direct way to get support is by reaching out to the OpenAI Support Team. You can either email support@openai.com or start a new chat session via the Help Center. The OpenAI support team is available 24/7 and is equipped to assist with the full range of requests.

As you noted, the easiest way to read the malformed event data is vai JsonPatch. An alternative approach is discussed in this comment on #528.

I'm going to close this out; if I've misunderstood what you're describing, please let us know in a comment and we'd be happy to assist as we're able.