Measured: OpenAIChatCompletion honours OPENAI_BASE_URL but ignores OPENAI_API_BASE — and replaying a kernel run offline

[Xizhuo (xizhuomengcontin)]

Sharing a small measured result about the Python connector plus the workflow I built on top of it.

OpenAIChatCompletion honours OPENAI_BASE_URL, not OPENAI_API_BASE

I wanted to point a kernel at a local endpoint without changing the code that builds it, so I checked which environment variable actually redirects the call rather than assuming. A local HTTP server on 127.0.0.1:9979 that returns a canned chat-completion and logs the path it is asked for, semantic-kernel 1.44.1:

OPENAI_BASE_URL=http://127.0.0.1:9979/v1
  -> request arrived at /v1/chat/completions, content == "ok"

OPENAI_API_BASE=http://127.0.0.1:9979/v1
  -> nothing arrived; the service raised on the real endpoint

So OPENAI_BASE_URL redirects and OPENAI_API_BASE is ignored. Both spellings turn up across the ecosystem — I have probed frameworks that honour both, and one that honours neither because it pins the origin in its own configuration — and pointing the wrong one at a gateway gives you a run that quietly went to production. That is worth knowing independently of anything below.

(base_url on the constructor is of course the explicit route; this is about the no-code-change path.)

What I do with it: run the same kernel again for free

Once the call can be redirected, it can be recorded and replayed:

npm i -g orcareplay
orca record generic-openai -- python my_kernel.py   # real run, recorded
orca replay last                                    # same run, no model called, nothing spent

Nothing in the kernel changes — the recorder moves OPENAI_BASE_URL for the child process and proxies that origin. orca replay last --from N --model <other> replays the run up to step N from the recording and then continues on a different model, so when comparing models the model is the only variable rather than one of several.

The case where this has paid off most for me is function calling: a plan that goes wrong on turn 7 of 20 is expensive to reproduce live, and replay makes "run that again, but change one thing" free after the first pass.

Disclosure: OrcaReplay is mine — Apache-2.0, Node 20+, no account, replay needs no key. The measurement above is about Semantic Kernel.

Three limits worth stating

• A matching replay is not a determinism result. It proves the recorded exchange reproduces; it does not prove the service or the model is stable, and it does not prove a fresh run would plan the same way.
• Replay blocks model-provider egress only. It is not a sandbox. Your plugins still execute for real on replay — a native function that writes a file writes it again.
• Embedding calls are not captured by the default adapter, so anything built on memory or vector search replays at the chat layer and is not reproduced below it.

If the base-URL behaviour is documented somewhere in the .NET or Python docs that I missed, I would appreciate the pointer.