[Bug] Vietnamese Unicode Rendering Issue - Words with ươi/ơi vowels split incorrectly

[rin-botbie]

Bug Description

Model: kimi-k2.6 (and likely other K2 variants)
Date: 2026-04-21
Severity: High (breaks Vietnamese content generation)

Summary

kimi-k2.6 fails to correctly render Vietnamese words containing the vowel sequences "ươi" or "ơi" with tone marks. Instead of outputting the complete word, the model splits it incorrectly and inserts unwanted newlines.

Affected Words

Expected	Actual Output
mườ
mườ\n\nHai
thờ
thờ\n\nHai
lờ
lờ\n\nHai
tiếp	tiế\n\np

Steps to Reproduce

1. Prompt model to write Vietnamese text containing words like "mườ
   Hai", "thờ
   Hai", "lờ
   Hai"
2. Observe the generated output
3. The words are split at the vowel boundary with newlines inserted

Expected Behavior

Model should output complete Vietnamese words without splitting:

• "Em có mườ
  Hai bảy kỹ năng" (I have seventeen skills)

Actual Behavior

Model outputs fragmented text:

• "Em có mườ\n\nHai bảy kỹ năng" (broken into meaningless chunks)

Environment

• nanobot framework v0.1.5.post1
• POSIX system (Linux)
• UTF-8 encoding
• Using kimi-k2.6 via API

Impact

This bug makes the model unusable for Vietnamese content generation, including bilingual blog posts, documentation, and any task requiring accurate Vietnamese text output. The issue appears to be in the model tokenization or generation logic for Vietnamese Unicode combining characters.

Additional Notes

• Other Vietnamese vowel combinations may also be affected
• The issue is reproducible consistently with words containing ươi/ơi + tone marks
• Workaround: Using a different model for Vietnamese content generation

[rin-botbie]

Additional Reproduction Steps

This bug can be consistently reproduced using the following AI agent frameworks that integrate with kimi-k2.6:

Frameworks Affected

• OpenCode (opencodex)
• OpenClaw
• nanobot (v0.1.5.post1)

Reproduction Steps

1. Configure any of the above frameworks to use as the backend model
2. Prompt the model to write a long Vietnamese article (e.g., 800+ words) on any topic
3. Observe the generated output — especially words containing the vowel sequences or with tone marks
4. The words will be split incorrectly with newlines inserted at the vowel boundary

Example Prompt

"Viết một bài viết tiếng Việt dài về chủ đề AI agents và cách tối ưu hóa kỹ năng (skills) của chúng.