fix(llm): avoid implicit reasoning-off for responses

[wey-gu]

PR Description:

• omit reasoning and include unless reasoning_effort is explicitly set on OpenAI Responses requests
• keep thinking=False from mapping to with_thinking(off) for openai_responses providers
• add regression coverage in kosong snapshot tests and create_llm tests

Reviewer Notes:

• this preserves provider-default reasoning behavior for Responses-compatible endpoints that require reasoning
• the change matches the safety invariant carried by our local AI Now and builtin-agent patches

Related Issue

None.

Description

This fixes an OpenAI Responses transport mismatch for providers and models that require reasoning by default.

Previously:

• create_llm(..., thinking=False) mapped openai_responses providers to with_thinking("off")
• OpenAIResponses.generate() always sent a reasoning block, even when no explicit reasoning level was requested

That request shape can be interpreted as disabling reasoning, which breaks Responses-compatible endpoints that require reasoning and returns a 400 error.

This change:

• omits reasoning and include unless reasoning_effort is explicitly set
• preserves provider-default reasoning behavior for openai_responses when no reasoning level is requested
• stops mapping thinking=False to with_thinking("off") for openai_responses
• adds regression tests in both kosong snapshot coverage and kimi_cli create_llm coverage

Checklist

• I have read the CONTRIBUTING document.
• I have linked the related issue, if any.
• I have added tests that prove my fix is effective or that my feature works.
• I have run make gen-changelog to update the changelog.
• I have run make gen-docs to update the user documentation.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.

[wey-gu]

gently ping but not pushy, we used this patch locally in production already and upstream here :)

@RealKai42