[BUG] Unicode corruption in Edit/Write payloads and written file content (U+FFFD / broken UTF-8)

Claude Code intermittently corrupts Unicode in the mutation-tool path.

This report is specifically about Edit / Write / related mutation payloads and written file content, not about assistant response text.

I have observed cases where structured tool payloads already contain broken Unicode, including U+FFFD and visibly broken multibyte sequences, and cases where the corrupted text is actually written into a file and later repaired by a follow-up Edit.

So this does not look like a terminal-display-only issue. It appears that Unicode is being corrupted somewhere in the tool input / serialization / file write path.

What Should Happen?

If the model emits valid Unicode, Claude Code should preserve it exactly through:

• tool input construction
• serialization / transport
• mutation-tool execution
• file writeback

It should never introduce:

• U+FFFD
• broken multibyte sequences
• silent Unicode degradation in written files

Error Messages/Logs

Observed examples from saved session logs:

Write/Edit payload examples:

• Japanese heading text became corrupted with U+FFFD
• Japanese prose in Markdown became corrupted with broken UTF-8 sequences
• A short CJK UI string was written as �� and later corrected to 次

Concrete corrupted fragments observed:

• 技術���タック
• ア��セス
• タスクワー��ー

Written-file readback example:

• Grep returned: <span ...>��</span>
• The very next Edit changed it to: 次

This suggests the corrupted text was actually written into the file.

Steps to Reproduce

This is intermittent, so I cannot provide a deterministic 100% repro yet, but the pattern is:

1. Start Claude Code in a repo with files containing Japanese/CJK text.
2. Ask Claude to create or edit files with Japanese text via Write or Edit.
3. Prefer longer sessions or sessions involving multiple edits/refinements.
4. Inspect:

• the resulting file content
• any follow-up grep/readback output
• saved session logs under ~/.claude/projects

1. In affected sessions, some tool payloads or written file content contain U+FFFD / � or visibly broken UTF-8 sequences.

Examples of affected content:

• Japanese Markdown headings
• Japanese prose in docs
• short CJK UI strings such as 次

Additional Information

This issue is intentionally limited to the mutation-tool / file-write path.

I am separating it from the assistant-response-text corruption issue:

• #40396

Related issues:

• #26141
• #18422
• #6870

My current interpretation is:

• response-text corruption and mutation-tool corruption may be related
• but this report is specifically about corruption that already appears in Edit / Write payloads or in the resulting file content

I checked two local environments.

Environment A:

• observed in logs for 2.1.77, 2.1.91, 2.1.92

Environment B:

• scanned 4148 jsonl files under ~/.claude/projects
• files containing U+FFFD: 52
• strongest source candidates were Edit, Write, Agent, and ExitPlanMode input payloads
• Read often appeared downstream, but did not look like the primary source in the checked cases

Temporary mitigation:
For the mutation-tool path only, a PostToolUse hook for Write / Edit / MultiEdit can scan written files for UTF-8 EF BF BD and reject the tool result if found. This is useful as a workaround, but it does not mitigate assistant-response corruption.