LSP helps massively with some models.
It sounds a lot like you are focused on using Pi as a OpenAI/Anthropic wrapper when that isn’t your audience. (Hint: raw token usage is less of a concern).