[kosong] OpenAI-compatible stream parser journals an empty think part per chunk carrying reasoning_content: ""

What version of Kimi Code is running?

0.31.1

Which open platform/subscription were you using?

Third-party OpenAI-compatible provider (custom base_url), not a Kimi subscription.

Which model were you using?

qwen3.8-max-preview — again provider-agnostic; depends on the endpoint's streaming habit, not the model.

What platform is your computer?

Linux 7.0.0-28-generic x86_64 x86_64

What issue are you seeing?

Some OpenAI-compatible endpoints (e.g. Alibaba DashScope compatible-mode) keep the reasoning_content key present with an empty string in stream chunks after the thinking phase ends. kosong's stream parser yields a think part for any string value — including '':

• packages/kosong/src/providers/reasoning-key.ts:44 — extractReasoning accepts any string (typeof value === 'string'), empty or not.
• packages/kosong/src/providers/openai-legacy.ts:445 — if (reasoning !== undefined) yield { type: 'think', think: reasoning } (same at :391 for the non-stream path).

Result: thousands of no-op content.part records get journaled into wire.jsonl. In one real session: 3,041 empty think parts vs 44 non-empty ones. This bloats session files and transcripts, and amplifies per-part rendering issues in the web UI.

What steps can reproduce the bug?

1. Configure an OpenAI-compatible endpoint that keeps reasoning_content: "" in its stream chunks (e.g. DashScope compatible-mode).
2. Run any session with a few model turns.
3. Inspect the session's wire.jsonl: most think-type content.part records are empty strings.

What is the expected behavior?

Empty reasoning strings should not produce think parts — e.g. if (reasoning) instead of if (reasoning !== undefined) at openai-legacy.ts:391/:445 (and the equivalent spots in other providers).

Additional information

Harmless for model requests (the context projector only strips blank text parts; empty think parts never affect the prompt sent to the model), so this is a journal-size/noise issue rather than a correctness one — but on chunk-heavy endpoints it inflates wire.jsonl by thousands of records per session.

Start with packages/kosong/src/providers/reasoning-key.ts:44 and packages/kosong/src/providers/openai-legacy.ts at lines 391 and 445, then inspect the equivalent spots in other providers. Reproduce the stream behavior with an endpoint that sends reasoning_content as an empty string and verify that empty values no longer create think parts in wire.jsonl while non-empty reasoning remains recorded.