Perhaps the issue stems from the LLM Model. Have you experimented with GPT 4.1? It's a capable model for code and tasks (consider the 4.1 mini if cost...