Default enable_thinking off in the agents scaffold for vLLM-served runs

[yilunzhao]

The openai_agents scaffold builds its own OpenAI client via provider.get_openai_client() and never goes through the provider's generate path, so a provider-level chat_template_kwargs was inert here. The request body this scaffold builds is the only place the value reaches the server.

That matters because no agentic preset configures a vLLM reasoning parser. A backbone whose chat template defaults thinking on would silently run in thinking mode and leak <think> blocks straight into the scored output. The scaffold now sends chat_template_kwargs={"enable_thinking": false} by default, gated to the vLLM-server provider class — the one that exposes a chat_template_kwargs attribute — because managed APIs reject unrecognized body fields with a 400. (A vllm_server provider pointed at a managed base_url would now get the field by default; before, that only happened when the kwarg was set explicitly.) An explicitly configured value still wins; the default only fills a gap.

For a model whose chat template already defaults thinking off, the kwarg is a no-op and nothing changes. This pins an implicit template default explicitly.

Tests: nine cases in tests/core/harness/test_openai_agents_scaffold.py covering the default, an explicit enable_thinking: true winning, unrelated configured kwargs keeping the default, a managed-API provider getting no injection at all, a metrics-wrapped provider still being recognized, the built dict aliasing neither the provider config nor the module-level default, and the logged outcomes. One of them drives the SDK with a stub client to confirm the value lands in the chat completions request body.

[undfined]

Looks good, one suggestion

[undfined]

We probably want to log something here about this outcome. wdyt?

[yilunzhao]

Added — the default now logs once at INFO per worker (the resolver sits behind the agent cache, so it runs once per worker init, not per instance), with the no-op outcomes at DEBUG. Two caplog tests pin the behavior.