[MODEL] Response opens in Japanese in a Korean-only project, then reverts to Korean mid-message

Preflight Checklist

• [x] I have searched existing issues (label model) 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

My project's CLAUDE.md (via an imported AGENTS.md) contains an explicit,
unambiguous language instruction:

All responses and deliverables must be written in Korean. (Exception: log message strings in code stay English.)

In the turn where the problem occurred, I sent a short follow-up prompt asking
three questions about the side effects of a refactor. The prompt was 113
characters, 100% Korean, containing no Japanese characters and no reference to
Japanese at all. Nothing earlier in the session was in Japanese either.

What Claude Actually Did

Claude's response began with two sentences in Japanese:

確認できました。三つとも回答します。

("I have confirmed. I will answer all three.")

Immediately after those two sentences, the response reverted to Korean and the
entire remainder of the message — headings, tables, explanations — was correct
Korean. The technical content of the answer was accurate; only the opening
language was wrong.

So this was not a full language switch. It was a localized language slip
confined to the opening sentences, self-corrected from the first heading
onward.

Expected Behavior

The entire response, including the opening sentences, should have been in
Korean, per the CLAUDE.md instruction and the all-Korean conversation history.
The opening should have read:

확인했습니다. 세 가지 모두 답변하겠습니다.

Files Affected

None. No files were read or modified as part of this behavior.
This is a natural-language output issue only.

Permission Mode

Accept Edits was ON (auto-accepting changes) — the session was started with
--dangerously-skip-permissions.

Can You Reproduce This?

No, only happened once.

Steps to Reproduce

No reliable reproduction. What I can describe is the state the session was in
when it happened:

1. Start a Claude Code session in a project whose CLAUDE.md mandates

Korean-only responses.

1. Work in that session continuously for a long time (see Additional Context for

the exact figures).

1. Send a short Korean follow-up question consisting of several sub-questions.

This produced the slip exactly once out of thousands of assistant messages, so
step 3 is almost certainly not a trigger — I include it only for completeness.

Claude Model

Opus (transcript records claude-opus-5)

Relevant Conversation

Verbatim first line of the affected assistant message, taken from the session
transcript on disk (the rest of the message is redacted as it contains internal
project details):

確認できました。三つとも回答します。

## <Korean heading>
...remainder of the message is entirely in Korean...

The immediately preceding user message was entirely Korean.

Impact

Low - Minor inconvenience

Claude Code Version

2.1.222 (Claude Code)

Platform

Anthropic API

Additional Context

I went back through the JSONL transcripts on disk to quantify how isolated this
was. Scanning every assistant message for Hiragana/Katakana codepoints
([぀-ゟ゠-ヿ] — deliberately excluding Han characters, since
those overlap with Korean Hanja usage and would produce false positives):

| Session transcript | Size | Assistant messages containing kana |
|---|---|---|
| Affected session | 2.9 MB / 1,793 lines | 1 |
| 3 other recent sessions in the same project | ~5 MB combined | 0 |

Other observations about the affected session that may or may not be relevant:

• The session had been running for roughly 28 hours of wall-clock time at

the point of the slip, and had been resumed once via --resume.

• The transcript contains zero compaction summaries (isCompactSummary),

so this is not an artifact of context compaction — I checked specifically
because that was my first hypothesis and the data did not support it.

• Only this one Claude Code session was running in the working directory at the

time of the slip.

• The project is a Korean-language codebase; all user prompts in the session

were Korean.

I'm aware a one-off, non-reproducible slip is hard to act on. I'm filing it
mainly because I could not find any existing report of CJK language instructions
leaking into an adjacent language, and this may be a useful anchor if others see
the same thing.