Skills single-file ZIP uploads are dropped during file extraction

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library.

Describe the bug

client.skills.create(files=("skill.zip", zip_bytes)) silently loses the file and constructs an application/json request with an empty body object. Wrapping the same file tuple in a list correctly produces multipart data.

The Skills create contract accepts either multiple files or a single zip. The SDK tries extraction paths [["files", "<array>"], ["files"]; the array path pops files before noticing that a single file is not a list. The fallback then cannot find it. The async client behaves the same way.

To Reproduce

Run the following against the current Python SDK checkout. It generates a zip in memory and uses MockTransport to inspect the actual public-client request; no API key or live service request is needed.

Code snippets

import io
import zipfile
import httpx2
from openai import OpenAI

buffer = io.BytesIO()
with zipfile.ZipFile(buffer, "w") as archive:
    archive.writestr("SKILL.md", "---\nname: example\ndescription: Test skill\n---\n# Example\n")
zip_bytes = buffer.getvalue()

def inspect_request(request):
    print(request.headers["content-type"], zip_bytes in request.content)
    return httpx2.Response(200, json={})  # Local response fixture only.

with OpenAI(
    api_key="fake-key",
    base_url="https://example.test",
    http_client=httpx2.Client(transport=httpx2.MockTransport(inspect_request)),
) as client:
    single = ("skill.zip", zip_bytes)
    client.skills.with_raw_response.create(files=single)
    client.skills.with_raw_response.create(files=[single])

Actual: the first request reports application/json False; the second reports multipart/form-data; boundary=... True.

Expected: both requests include the zip in multipart data. The single-file field is files, while the list form uses files[].

OS

macOS 26.6.2 ARM64.

Python version

Python 3.13.15.

Library version

Current main source reporting openai 3.14.1, loaded via PYTHONPATH=src; httpx2 2.12.0, Pydantic 2.13.5.

Start at the public client entry point, client.skills.with_raw_response.create, and reproduce the behavior with httpx2.MockTransport using both the single tuple and list forms. Trace the extraction paths for ["files", ""] and ["files"] in the request-building code, including the async client. Done means both forms preserve the ZIP in multipart data, with tests covering the single-file and list cases.