Bug Report: Transcript JSONL encoding corrupted on Windows

Version

OpenClaw 2026.3.23 (ccfeecb)

Environment

• OS: Windows_NT 10.0.26200
• Node.js: v24.14.0
• Channel: QQBot

Problem

Chinese characters in chat transcripts (.jsonl files under agents/main/sessions/) are stored corrupted. Characters appear as replacement character � or garbled mojibake.

Root Cause

The transcript writer appears to encode Chinese text as GBK bytes first, then interpret those bytes as UTF-8, causing permanent corruption. The original text cannot be recovered.

Impact

• Historical chat search returns no results for Chinese queries
• Memory retrieval unreliable
• All past conversations with Chinese content permanently unreadable in transcript files
• memory_search finds matches but context is garbled

Example

User message containing "有奖原创" is stored as:

"content":[{"type":"text","text":"[QQBot] to=qqbot:c2c:6B1D6D9CF1C53871FC42C880DFD44DB5\n\n???? ?????? ????????????????"}]

Expected Behavior

Chinese text should be stored correctly as UTF-8 JSON.

Workaround

None — corrupted data cannot be recovered.

Suggested Fix

Ensure all transcript writes use UTF-8 encoding explicitly. On Windows, the default system encoding (GBK/CP936) must not be used. Force JSON.stringify() output to be written as UTF-8 bytes, or convert to a lossless encoding (e.g. escape unicode sequences) before writing.

Files Affected

• agents/main/sessions/*.jsonl — all session transcripts
• workspace/memory/*.md — daily memory files may also be affected