Hanja/foreign character substitution recurs in normal chat responses, not just suggestion box
Description
Follow-up to #86371. That issue reported Hanja substituting for Hangul in the auto-filled "next step suggestion" text box. The user then asked me to fix/prevent the recurrence of this issue (and I saved a memory note committing to avoid it). In the very next assistant chat response — a normal conversational reply, not the suggestion box — the same substitution occurred again.
Example
In an actual assistant response (plain chat text, Korean), the word "내용" (content) was rendered as "内容" (Hanja) instead of Hangul:
内容은 스크린샷에서 본 "進行" 오출력 현상과...
Expected:
내용은 스크린샷에서 본 "進行" 오출력 현상과...
Why this matters more than #86371
This shows the Hanja/Hangul substitution is not confined to the lightweight next-step-suggestion generation path — it also occurs in normal assistant chat message generation, immediately after the model had explicitly committed (in the same session, via its own persistent memory note) to avoid exactly this behavior. The instruction/memory did not prevent recurrence.
Environment
- OS: Windows 11 Home 10.0.26200
- Claude Code running in VS Code integrated terminal
- Occurred in the main conversational response text, generated by the primary model for this session (not the suggestion-box feature from #86371)
Expected behavior
Korean text generated in normal assistant responses should consistently use Hangul for Sino-Korean vocabulary, not substitute homophonous Hanja characters, and should not regress even when the user has just asked for this to be fixed within the same session.
4 Comments
Another recurrence, same session pattern as the one described above (assistant had already committed via its own memory note to avoid this).
In a normal chat response (Korean), the assistant wrote a Japanese word instead of the Korean word:
Expected:
This time it's not a Hanja/Hangul homophone substitution (as in the original report) but a full Japanese word ("新しい", "new") inserted in place of the Korean equivalent ("새로운") — same broad failure mode (non-Hangul text leaking into a Korean response) but a different source language, suggesting the underlying issue isn't limited to Sino-Korean homophone confusion.
Cross-referencing a third-party report that matches the Japanese-leak variant described above (comment https://github.com/anthropics/claude-code/issues/86373#issuecomment-5280947324): #84052.
There, a different user in a completely different Korean-only project got a response opening in Japanese ("確認できました。三つとも回答します。") before self-correcting to Korean from the first heading onward, despite a 100%-Korean prompt and no prior Japanese in the session. Same broad failure mode (non-Hangul language leaking into a Korean-only response) reproduced independently across different users/projects, which points toward a model-level issue rather than something specific to this session or repo.
Another recurrence, this time not in a chat response but in the CLI's end-of-session "recap:" summary line (the one shown after "Brewed for Xm Ys" in the terminal).
Session was entirely in Korean (project instructions force Korean-only replies, no foreign-word mixing). The recap line it auto-generated was fully in Japanese:
Not a single Korean character in the line — kanji/okurigana/particles are all Japanese grammar, not just a stray foreign word mixed into an otherwise-Korean sentence like the earlier reports here. Same as the pattern in #84052 (response starting in Japanese despite a Korean-only project), but here it's the recap/summary generation path specifically, not the main chat turn — suggesting this lightweight generation path is a separate, also-affected code path from the main response generation discussed above.
4차 재현 (2026-08-18, 다른 세션): 이번엔 recap이나 자동 제안창이 아니라 메인 채팅 응답 본문에서 발생했습니다. "자동 시작 등록을 원하시면..." 문장 중 "자동"을 써야 할 자리에 한자 "自動"이 그대로 섞여 출력됐습니다("自動 시작 등록을 원하시면..."). 이전 세 차례(제안창, 채팅 응답 초반 일본어 시작, recap 전체 일본어)와 달리 이번엔 정상적인 한국어 문장 중간에 한자 단어 하나만 삽입된 형태입니다.
세션에는
settings.json에"language": "한국어"가 설정돼 있었고,CLAUDE.md에도 "순수 한글 표기, 한자/외국어 혼용 금지" 규칙이 이미 명시돼 있었음에도 재현됐습니다.