[Desktop app] Assistant messages shown to the user are replaced by paraphrased (often English) "thinking" text, and replies drift into English for a Japanese-speaking user

Environment: Claude Code desktop app (Code tab), Windows 11, model claude-opus-5-5 (also seen with claude-opus-5 and claude-fable-5-1), effort "high"/"max", auto mode.

Summary
The user writes only in Japanese and the project instructions say all user-facing text must be Japanese. Since mid-September, more and more of what the user sees from Claude is English. Comparing the screen with the session transcript (~/.claude/projects/<project>/<session>.jsonl) shows that part of this is not the model's reply at all:

1. Some assistant messages never reach the user as written. In the transcript, the text block is missing and a thinking block holds a paraphrase instead; the app displays that paraphrase as the message.

• The model's reply (Japanese, bullet list about posting this very report) appears on screen as a shorter, reworded Japanese paragraph stored as thinking at 2026-10-05T11:04:47Z.
• The next reply appears on screen only as an English paraphrase ("Given the numbering feature's edge lies in a larger free tier, duplicate-free numbering, and continuation from existing numbers, I'll focus on those points...") stored as thinking at 11:04:59Z; no text block with the reply exists.
• Same at 11:05:14Z ("I'll implement number assignment using optimistic concurrency ...").

The model, seeing its own context, believes it wrote something else, and told the user "my messages were Japanese" when the screen showed English.

1. In addition, many text blocks themselves are English. Share of assistant text blocks that are English, per day, from local transcripts:

• 2026-09-12 to 09-15: 0-3% (heavy tool use every day)
• from 2026-09-16: 8-45% on most days
• 2026-10-05: 85% (one long session; 428 English vs 55 Japanese text blocks)

Expected

• The user sees the assistant's actual reply, not a paraphrased thinking summary in its place.
• Replies stay in the user's language even when tool output, UI text or compaction summaries are English.

Notes

• No language or output-style setting is configured; CLAUDE.md and a saved memory both require Japanese.
• Re-stating the rule mid-session does not help; the drift resumes within a few tool calls.