Intermittent Korean (CJK) character corruption in model output

Description

Intermittent UTF-8 corruption of Korean (CJK) characters in Claude Code model output. Characters are partially rendered or replaced with replacement characters (U+FFFD).

Example

Expected output:

#N개의 상품을 찾았습니다.

Actual output:

#N개의 상품을 ��았습니다.

The corrupted text was generated by the model and passed through to a tool call (curl/API push), causing corrupted data to be stored in an external system.

Environment

• Claude Code CLI (macOS, Darwin 24.6.0)
• Model: claude-sonnet-4-6
• Started occurring approximately April 9-10, 2026
• Intermittent — not every output, but happens a few times per long session

Impact

When corrupted output is used in tool calls (e.g., writing to files, API calls), the corruption propagates to external systems without the user noticing until later.

Steps to Reproduce

1. Use Claude Code in a long session with heavy Korean text output
2. Generate text containing Korean characters through tool calls
3. Occasionally, some characters will be corrupted

No reliable minimal reproduction yet — it appears to be intermittent and may be related to session length or output volume.