feat: emit thinking in stream-json (JSONL or documented stderr • block)

What feature would you like to see?

Please emit model thinking during kimi -p --output-format stream-json. Either of these is enough:

1. Preferred for machines: a JSONL field on stdout (reasoning_content, a thinking content block, or a thinking.delta meta line), or
2. Also fine: the documented print-mode • thinking block on stderr, which PromptTranscriptWriter already writes in text -p.

Today thinking.delta events exist and both print drivers forward them to PromptJsonWriter.writeThinkingDelta(), but that method is a no-op. Tool progress still goes to stderr; the resume hint is a stdout meta line. Headless hosts therefore sit silent for the entire reasoning window.

Additional information

Official docs (current docs/en/reference/kimi-command.md)

In stream-json mode, regular replies produce an Assistant message; when the model calls a tool, an Assistant message with tool_calls is emitted first, followed by the corresponding Tool message, then subsequent Assistant messages. Thinking content is not written to JSONL; tool progress and "resuming session" notices are still written to stderr.

Print-mode text (-p without stream-json) documents thinking on stderr:

Output uses a transcript style: thinking content and Assistant text are both prefixed with • , and wrapped lines are indented by two spaces. Assistant text goes to stdout; thinking, tool progress, and "resuming session" notices go to stderr.

PromptJsonWriter.writeThinkingDelta() is a no-op

apps/kimi-code/src/cli/prompt-render.ts:

export class PromptJsonWriter implements PromptTurnWriter {
  // …
  writeThinkingDelta(): void {}

Compare PromptTranscriptWriter.writeThinkingDelta, which writes the • block to stderr.

Both print drivers already call it:

• apps/kimi-code/src/cli/run-prompt.ts — case 'thinking.delta': outputWriter.writeThinkingDelta(event.delta);
• apps/kimi-code/src/cli/v2/run-v2-print.ts — same

Same files, still in stream-json:

• tool.progress writes event.update.text to stderr
• writeResumeHint(..., 'stream-json', ...) writes { "role": "meta", "type": "session.resume_hint", ... } on stdout

Verified on official 0.39.1: a live isolated-home kimi -p … --output-format stream-json capture never emitted a thinking JSONL field or a • stderr block.

Suggested shapes (pick one)

JSONL on stdout:

{"role":"assistant","reasoning_content":"…thinking text…"}

or a content block:

{"role":"assistant","content":[{"type":"thinking","thinking":"…"}]}

or streaming deltas:

{"role":"meta","type":"thinking.delta","delta":"…"}

Or, if you prefer not to change JSONL: write the same • stderr block PromptTranscriptWriter already uses, so stream-json matches the documented print-mode thinking channel.

Related: #3080 (same print path — PromptJsonWriter also drops turn.step.completed usage).

Use case: orchestration hosts that wrap kimi -p --output-format stream-json (same pattern as Claude / Cursor stream-json) have no thinking signal, so the UI shows only a spinner until the first assistant / tool_calls line.

Start in apps/kimi-code/src/cli/prompt-render.ts, where PromptJsonWriter.writeThinkingDelta() is currently a no-op, and compare it with PromptTranscriptWriter.writeThinkingDelta(). Trace the existing thinking.delta handling in apps/kimi-code/src/cli/run-prompt.ts and apps/kimi-code/src/cli/v2/run-v2-print.ts. Done means stream-json exposes thinking through one documented JSONL or stderr path without breaking the existing tool-progress and resume-hint behavior.