context.md contains garbled text (mojibake) when prompts include CJK or other multi-byte characters

[wasabeef]

What happened?

When a session prompt contains CJK characters (Japanese, Chinese, Korean) or other non-ASCII multi-byte characters, the generated context.md on the
entire/checkpoints/v1 branch contains garbled text (mojibake).

Root cause: generateContextFromPrompts in manual_commit_condensation.go truncates long prompts with byte slicing (displayPrompt[:500]). Because CJK characters
are 3 bytes each in UTF-8, slicing at byte offset 500 splits multi-byte sequences at arbitrary byte boundaries, corrupting the output.

The corruption can be verified locally with:

prompt := strings.Repeat("あ", 200) // 200 CJK chars = 600 bytes
truncated := prompt[:500]           // splits mid-character → invalid UTF-8

The prompt.txt and full.jsonl files in the same checkpoint are unaffected — the bug is isolated to the truncation logic in generateContextFromPrompts.

Steps to reproduce

1. entire enable
2. Start a Claude Code session and send a prompt longer than ~167 CJK characters (≥ 500 bytes)
3. Make a git commit to trigger condense
4. Inspect entire/checkpoints/v1//0/context.md — the truncated prompt will contain garbled characters

Entire CLI version

v0.4.5

OS and architecture

Darwin 24.5.0 arm64

Agent

Claude Code

Strategy

manual-commit (default)

Terminal

WezTerm

Logs / debug output

Additional context

No response

[linear]

from evis:

Thanks for submitting this, we'll prioritize this and come back with an update when ready.