claude-code - 💡(How to fix) Fix [Critical] Opus 4.7 Korean output degeneration — Korean grammar itself collapses in long contexts

Severity

High. This is a regression that effectively blocks Korean-speaking users from running long sessions. The underlying work (code edits, file changes, tool calls) is performed correctly, but the natural-language output to the user degrades into meaningless strings, breaking the human-agent communication channel. In autonomous agent modes, the user keeps receiving broken reports for a long time before noticing.

Summary

In long Korean-language sessions on Opus 4.7, the assistant's output collapses in two simultaneous ways:

1. Lexical collapse — a wide range of content words is replaced by a single high-frequency Korean word (e.g., a generic demonstrative or generic noun).
2. Grammatical collapse — at the same time, Korean grammar itself breaks down. Particle attachment goes wrong, verb endings concatenate ungrammatically, subject–predicate agreement fails, and the same morpheme repeats meaninglessly. The result looks like Korean script but is not interpretable as Korean by a native speaker.

This did not happen on previous Claude models doing equivalent long Korean sessions — it is a new regression introduced in Opus 4.7.

Scope

• Observed only on Opus 4.7 (Claude Code, 1M context).
• Not reproduced on other Claude models under the same usage pattern.
• Recurs across different sessions with different trigger words (one session degenerates around one generic noun, another session around a different generic noun). This is not a specific-word bug; it is a model-specific Korean decoding regression.
• Switching to English within the same session produces normal output — this is not a generic long-context limit issue.

Reproduction

1. Run a very long Korean coding/agent session on Opus 4.7 (code + Korean prose, context exceeding ~50k tokens).
2. Once the context is large enough, the assistant's Korean prose starts substituting almost every content word with a single generic noun.
3. At the same time, grammar breaks down — particles and verb endings desync, the same morpheme repeats, sentence boundaries blur. Paragraph-level non-Korean output results.
4. Tool calls and file edits remain correct — only the user-facing Korean narration collapses.
5. Switching to English in the same session immediately restores normal output.

Expected

Natural Korean prose that maintains vocabulary diversity and correct Korean grammar (particles, endings, word order, agreement) regardless of session length. Previous Claude models behaved this way.

Actual

Opus 4.7's Korean output enters a self-reinforcing loop where both vocabulary and grammar collapse simultaneously. A single token dominates and the rest of the sentence degenerates into it, and the grammatical structure no longer parses as Korean.

Why I treat this as a regression

• Same user, same task type, same interface — previous Claude models never produced this; Opus 4.7 reproduces it across sessions.
• The trigger word differs across sessions, which points to an attractor in Korean decoding rather than a prompt-specific pattern.
• Not only lexical substitution but grammatical structure breaks at the same time — this looks deeper than a missing repetition penalty; it looks like overall Korean language-modeling quality is degraded under load.
• English output in the same session at the same point is normal, ruling out a generic long-context limit.

Impact (important)

• Korean users running long sessions cannot understand the assistant mid-task and have to start a fresh session.
• In autonomous agent modes, the user cannot follow progress or decisions if they do not intervene immediately.
• The artifacts (code, files) are still correct, but the user loses the ability to understand the meaning, reasoning, and next steps in Korean — the natural-language interface effectively stops working.
• English speakers cannot see this issue, so it is at high risk of being missed in regression testing.

Reproduction material

Full transcript withheld (contains private project code) — available privately on request.

• Early turns of the affected session are normal.
• Korean lexical and grammatical collapse begins after the context grows past ~50k tokens.
• English output remains normal in the same session at the same point.

Suggested areas to investigate

• Changes in Korean tokenization/sampling in Opus 4.7 vs. earlier models — is there a regression?
• Whether decoding parameters (repetition penalty, nucleus, etc.) for CJK are tuned the same as English, and whether they were changed in 4.7.
• Whether low-diversity attractors form among Korean grammatical tokens (particles, endings) under long-context attention dilution.
• Whether long-context CJK regression tests were part of the 4.7 release evaluation.

Request

Korean users cannot use Opus 4.7 for long sessions in its current state. Please prioritize.