[BUG] Summarized thinking language does not follow conversation language — English/Japanese mixed within one session

Environment

• Claude Code in Claude Desktop app, Windows 11
• Model: claude-fable-5
• Conversation language: Japanese (set via language in settings.json; responses are consistently Japanese)

Description

Thinking summaries randomly switch between Japanese and English within the same session. Some turns render fully in Japanese, others fully in English, and some switch language mid-summary. The response text itself is always Japanese — only the thinking summaries are inconsistent.

Since the transcript jsonl stores the summarized text (not the raw chain of thought), the stored record is also mixed-language.

I understand from the docs that summaries are generated by a separate model ("Summarization is processed by a different model"). The issue is that this summarizer does not appear to receive or respect the conversation language, so its output language is unstable.

Steps to reproduce

1. Configure Claude Code for Japanese responses (language setting)
2. Run a session with extended thinking enabled, multiple turns
3. Expand thinking blocks across turns

Expected

Thinking summaries follow the conversation language consistently.

Actual

Summary language varies turn-by-turn (Japanese / English / mixed within one block). In one case the summary also mistranslated a referenced fact, reversing its meaning.

Impact

• Users working in non-English languages get an inconsistent reading experience
• Mistranslations in summaries can misrepresent what the model actually reasoned
• The transcript jsonl permanently records the mixed-language summaries