[Fix] Follow redirects when a custom httpx client is supplied to the OpenAI SDK

[CastilloLuis]

Problem

OpenAISDK and OpenAISDKRollout build their own httpx.Client and hand it to OpenAI(http_client=...). That silently disables redirect following:

• httpx.Client() defaults follow_redirects=False.
• openai-python only applies its own follow_redirects=True when it constructs the client — _base_client.py does self._client = http_client or SyncHttpxClientWrapper(...), and only that wrapper calls kwargs.setdefault("follow_redirects", True).

So supplying http_client bypasses the SDK default, and any endpoint that answers with a redirect fails. The failure is unhelpful: the raw redirect response surfaces as an APIStatusError containing HTML, not a connection error, so it looks like a broken endpoint rather than a client setting.

This affects any OpenAI-compatible endpoint that redirects — hosted gateways commonly do — not one specific provider.

Fix

http_client_cfg.setdefault('follow_redirects', True) at both construction sites, so the SDK's own default is preserved. Because it's setdefault, a user who explicitly passes http_client_cfg=dict(follow_redirects=False) still wins.

9 added lines, 6 of which are the explanatory comments.

Reproduction

Against a real endpoint that answers 307, running the exact constructor from _create_fresh_client:

before (http_client_cfg = {})      APIStatusError: redirect not followed
after  (setdefault applied)        OK -> openai/gpt-oss-120b
user override preserved            follow_redirects=False (not clobbered)

For contrast, the same request through a client that openai-python builds itself succeeds — which is what makes this specific to the supplied-client path.

Notes

• OpenAISDKRollout passes None when http_client_cfg is empty, so the SDK builds its own client and that path was already correct; the setdefault only applies where a client is actually constructed.
• The legacy OpenAI class uses requests, which follows redirects already, so it is unaffected.
• There is a user-side workaround today (http_client_cfg=dict(follow_redirects=True) in the model config), but it is undocumented and the failure mode gives no hint that it's needed.

Disclosure: I work on The Grid, an OpenAI-compatible provider whose endpoint redirects — that's how I hit this. The fix isn't provider-specific and I've deliberately not added any Grid config in this PR.