OpenAIServerModel falls through to OPENAI_BASE_URL — which makes smolagents runs recordable and replayable with no code change

[xizhuomengcontin]

Small finding plus the tool it enables, verified against smolagents 1.26.0 today.

Disclosure: I work on the tool (OrcaReplay, Apache-2.0).

The finding

OpenAIServerModel constructed with no api_base does not silently fall back to OpenAI's hosted endpoint — it falls through to the OpenAI SDK's own OPENAI_BASE_URL:

model = OpenAIServerModel(model_id="qwen3.5:0.8b", api_key=...)   # no api_base
print(model.client.base_url)
# → http://localhost:11434/v1/

That is worth knowing in both directions. It means you can point smolagents at a local model, a gateway or a proxy purely through the environment — no code change. It also means if you have OPENAI_BASE_URL exported for something else, smolagents will quietly use it.

What that enables

Because the origin comes from the environment, a recording proxy can capture a run without anything being installed into smolagents — no callback, no instrumentation, no wrapper:

$ orca record generic-openai -- python smol_agent.py
smolagents said: SMOLAGENTS
info recorded run=run_83ab64f3e91b events=6 exit=0

$ taskkill /F /IM ollama.exe            # model server gone
$ orca replay run_83ab64f3e91b
smolagents said: SMOLAGENTS
info replay.done reused=1/1 exact=1 divergences=0 unmatched=0 exit=0

The program runs again against the recorded bytes with no model server anywhere on the machine. Useful for re-running an agent while you change the prompt or the tools, for CI, or for chasing one bad answer without paying for it each time.

Two honest limits

• It sees the HTTP going to the model, not smolagents' internals — no step-level spans, no tool-call tree beyond what appears on the wire. Not a substitute for tracing; a different question.
• egress=blocked on replay means model-provider egress. Recorded tool calls still execute for real, so a recorded shell or search tool runs again. It is not a sandbox.

If anyone knows a case where OpenAIServerModel resolves its origin some other way, I would like to hear it — I would rather fix my assumption than ship it.