This is a common pattern unfortunately. Big LLMs are benchmaxxing coding and one shot answers, and multi turn conversation is taking a nosedive.

https://arxiv.org/abs/2504.04717

Restructure your prompts, or better yet try non-OpenAI LLMs. I’d suggest z.ai, Jamba, and Gemini Pro for multi turn. Maybe Qwen Code, though it’s pretty deep fried too.