I encountered the same issue. Kimi 2.7 looked impressive on paper, but in practice, the code was so riddled with errors that I ended up using GPT-5.5 to fix it.