fix(core): turn off Workers AI thinking through the chat template - #52017

Issue for this PR

No linked issue.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Workers AI used the generic Chat variant builder, which sends every catalog effort as reasoning_effort. That caused two problems:

• The none variant does not turn thinking off. The Chat endpoint accepts reasoning_effort: "none" but ignores it on deepseek-v4-flash, deepseek-v4-pro, gemma-4 and glm-5.2; only kimi-k2.6 honors it. (Cloudflare's model metadata lists none as supported, which is where the catalog value comes from.)
• Models that only offer thinking on/off get no off switch. glm-4.7-flash had no variants at all; qwen3.8 and nemotron had no way to turn thinking off.

Workers AI now gets its own builder, like NVIDIA, Baseten and the other hosting providers already have:

• none and toggle-off send chat_template_kwargs: { enable_thinking: false, thinking: false }, and toggle-on sends both as true. Cloudflare's per-model input schemas document chat_template_kwargs.enable_thinking as the reasoning switch. Kimi's template reads thinking instead, so both keys are sent.
• Other effort levels still send reasoning_effort, which works.

How did you verify your code works?

• New case in packages/core/test/variant.test.ts (fails without the change). bun typecheck passes in packages/core.
• Live, opencode run --standalone -m cloudflare-workers-ai/<model>#<variant> through a logging proxy, counting reasoning in the response and in the stored message:

• Multi-step bash tool loops with deepseek-v4-flash #none, kimi-k2.6 #none, qwen3.8 #none and glm-4.7-flash #thinking all complete, and prompt cache reads still appear on later steps.

Screenshots / recordings

Not applicable.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR