[BUG] Summarized thinking is shown in English in a fully Japanese conversation (re-report of #82785)

Summary

In a conversation conducted entirely in Japanese, the summarized thinking displayed in the Claude Desktop UI appears in English. The language of the summary does not follow the conversation language.

This is a re-report of #82785, which was closed as not planned by github-actions[bot] on 2026-09-20 as inactive. The behavior still reproduces on 2026-09-21.

Environment

• Claude Code on Claude Desktop, Windows 11
• Conversation language: Japanese (both user input and assistant output)
• Models confirmed: Opus 5, Fable 5.1
• Observed: 2026-09-21

Behavior

The user writes in Japanese and the assistant responds in Japanese, yet the thinking summary shown in the UI is frequently rendered in English. In practice English is the dominant language of the summary.

The language also varies at finer granularities:

1. Per turn — one turn's summary is entirely English, the next is entirely Japanese.
2. Per block — within a single turn, some thinking blocks are English and others are Japanese.
3. Within a single block — a block begins with an English token and continues in Japanese.

The screenshot below shows (2): five Japanese blocks followed by two English blocks, all within a single turn. The conversation and the subject matter were both entirely Japanese.

Example of (3), from a summary block (middle of the sentence elided):

So [...] 言語はブロック単位でバラついていて、実際は英語のほうが多い。

Only the leading "So" is English; the rest of the sentence is Japanese.

What appears to influence the language

While preparing this report I ran several turns to characterise the behaviour. The summary language appears loosely correlated with the language of the material being discussed, but the correlation is not consistent:

Condition	Summary language
Japanese subject matter (translating a term into Japanese)	All Japanese
English subject matter (explaining regex terminology)	All English
Explicit instruction to think in Japanese, with English subject matter	All Japanese
Japanese subject matter (translating an English draft into Japanese)	Mixed — five Japanese blocks followed by two English blocks

The last row is the screenshot above: the subject matter was Japanese throughout, yet the final two blocks came out in English.

These are observations, not a diagnosis.

Not model-specific

Reproduced on both Opus 5 and Fable 5.1 within the same session.

Not language-specific

#82785 received a comment on 2026-08-17 reporting the same behavior in Korean.

Additional observation

In one summary, a word written in hiragana in the original thinking appeared in kanji in the summary. The summary is not a verbatim excerpt of the thinking — it is regenerated text. The language selection may diverge at that step.

No source file, test, or implementation entry point is named. Start at the Claude Desktop UI's summarized-thinking display and reproduce the Japanese conversation cases across the listed models; done means summaries consistently follow the conversation language without per-turn, per-block, or within-block mixing.