fix(core): send reasoning_effort at the top level of the request

[coderbuzz]

Problem

buildThinkingRequestOptions returns { thinking, extra_body: { reasoning_effort } }, and SessionManager spreads that into the
chat.completions.create params. extra_body is a Python OpenAI SDK option; the Node SDK has no equivalent and serializes it as a
literal "extra_body" key in the JSON body. The DeepSeek API ignores that key, so the reasoningEffort setting
(low / high / max) never reaches the model.

Evidence

Direct calls to https://api.deepseek.com/chat/completions with model: deepseek-v4-flash and thinking: {type: "enabled"}:

body	response
"reasoning_effort": "bogus" (top level)	422: reasoning_effort: unknown variant 'bogus', expected one of none, minimal, low, medium, high, xhigh, ultra, max
"extra_body": {"reasoning_effort": "bogus"}	200 (field never read)

So the API reads the top-level field and never looks inside extra_body.

Fix

Send reasoning_effort at the top level, next to thinking (which is already top level). The tests in
openai-thinking.test.ts are updated to the new shape.

I could only verify the DeepSeek endpoint. The Volcengine Ark tests use the same payload as before, so please check whether Ark
also expects reasoning_effort at the top level.

Tests

npm test in packages/core: all openai-thinking tests pass. 3 unrelated tests fail with and without this change on my machine
(computeToolCallPermissions allows read tool calls under skill scan paths, ReadImage uses filesystem read permissions ...,
UnderstandImage requires network ...).

🤖 Generated with Claude Code