[BUG] AskUserQuestion frequently produces corrupted Traditional Chinese text while normal assistant output remains unaffected

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

AskUserQuestion frequently produces corrupted or semantically meaningless Traditional Chinese text.

Important observations:

• The issue appears only in the AskUserQuestion interactive UI / structured tool path.
• Normal assistant free-text responses in the same sessions remain correct.
• Copying the corrupted text from the UI preserves the incorrect characters, so this does not appear to be only a terminal font/rendering issue.
• The problem is reproducible across multiple models.

Examples observed:

• 敷衍點 became 敷且點
• 雙月鷲油
• 肭子
• 相刃的

<img width="1162" height="396" alt="Image" src="https://github.com/user-attachments/assets/74f38218-3400-4680-b357-9b3c2e4acace" />

The latter examples consist of valid Chinese characters, but the resulting phrases are meaningless to a native Traditional Chinese speaker.

Test results:

• Sonnet 5:
• Session 1: 8/10 rounds corrupted
• Fresh session: 6/10 rounds corrupted
• Combined: 14/20 rounds corrupted (70%)
• Corruption was often severe enough to make the intended meaning impossible to infer

• Opus:
• 6/10 rounds corrupted
• Usually still interpretable from context

• Fable:
• 5/10 rounds corrupted
• Usually still interpretable from context

Environment:

• Claude Code: v2.1.236
• Language: Traditional Chinese

What Should Happen?

Traditional Chinese generated inside AskUserQuestion should remain natural, readable, and semantically correct, with text quality comparable to normal assistant free-text responses.

AskUserQuestion should not replace Chinese characters with incorrect or meaningless characters.

Error Messages/Logs

Steps to Reproduce

1. Start a fresh Claude Code session.
2. Use a model such as Sonnet 5.
3. Ask Claude to continuously chat using AskUserQuestion in Traditional Chinese.

Example prompt:

持續用 AskUserQuestion 跟我聊天

1. Answer each AskUserQuestion and continue for around 10 rounds.
2. Observe the generated questions and options.

In my testing:

• Sonnet 5 reproduced the issue in 14/20 rounds across two sessions.
• Opus reproduced it in 6/10 rounds.
• Fable reproduced it in 5/10 rounds.
• Normal assistant free-text responses remained correct.

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.236

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

_No response_