[Model] Repeatedly switches to English despite CLAUDE.md + memory + repeated in-chat instructions to reply only in Japanese

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

All replies must be in Japanese. This is stated in my user-level CLAUDE.md ("成果物と返答は日本語") and in a persistent memory file that is loaded every session ("途中の一言も含めて報告はすべて日本語" — all messages, including short progress notes between tool calls, must be Japanese). I also repeated it explicitly in chat many times during the session.

What Claude Actually Did

In a single long session (Claude Code VS Code extension, Opus 5.5 1M context, auto mode, Windows 11), the model repeatedly switched to English:

• Progress notes between tool calls and final reports were written in English, even right after I had corrected it.
• I had to say "reply in Japanese" roughly 8–10 times in one session. Each time it apologized, re-wrote the same content in Japanese, and then drifted back to English a few turns later.
• The relapse was most frequent (a) right after receiving a subagent's report (the subagent's report itself was in Japanese), and (b) after long runs of tool calls (builds, measurements, installs).
• Result: much of the output was produced twice (English, then a Japanese rewrite), wasting time and usage.

Expected Behavior

Once the language is specified in CLAUDE.md / memory and reinforced in chat, every message — including short notes between tool calls — stays in Japanese for the whole session without needing repeated corrections.

Files Affected

None (behavior issue only).

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes — it appears in long sessions with many tool calls and subagent hand-backs.

Claude Code Version

2.1.282 (Claude Code), VS Code extension, Windows 11

Additional Context

The model also kept the memory file updated with each recurrence ("this happened again"), but writing it down did not prevent the next relapse within the same session.