Cured in v0.9.0: streaming chunks used to cut UTF-8 multi-byte characters in half. Upgrade and it is gone.
While an AI reply streams in (typewriter effect), Chinese characters occasionally turn into mojibake mid-stream; the next generation looks fine again. The problem appears randomly and seems length-related.
AI replies stream as SSE chunks, and each UTF-8 Chinese character takes 3 bytes. The old implementation decoded every network chunk as it arrived; when a chunk boundary fell in the middle of a multi-byte character the truncated sequence decoded into garbage — hence random, content-length-dependent occurrences.
Q: Is garbled Chinese in AI streaming output a virus or data corruption?
A: No. It is a decoding display issue caused by stream chunk boundaries falling inside a UTF-8 multi-byte character. The mail data itself is intact, and v0.9.0 cures it completely.
Q: Why does the garbling appear randomly?
A: Network chunk boundaries differ each time; garbling only happens when a boundary lands mid-character, so it looks random.
Q: Do I need to regenerate previously garbled replies after upgrading?
A: Previously generated text is unaffected (it was a display-layer issue). If a specific old reply still displays oddly, simply regenerate it.