LLM-powered health coaching and model updates arrive in llm 0.36

WHOOP delivers personalized fitness and health coaching with GPT-4, applying a large language model to tailor guidance to users. Alongside product uses of LLMs, the llm 0.36 release adds new OpenAI model entries: gpt-6-sol for GPT-6 Sol and gpt-6-luna for GPT-6 Luna, referenced as #1702. The release also introduces a clearer way for model plugins to declare limitations: plugins can now set supports_conversation = False for models that only accept single-turn prompts. If such a model receives assistant or tool-related input that assumes conversation support, LLM raises llm.ConversationNotSupported.

In a product setting, a large language model can be useful when guidance must be translated into plain language, adapted to goals, and framed around patterns rather than isolated readings. Personalized fitness and health coaching is a natural area for this kind of interaction because the model can help users understand what the available signals may mean, suggest next steps, and keep the tone of the advice consistent across many users. At the same time, such applications depend on clear boundaries: the model should support the user, not replace medical judgment, and any coaching output should be understood as part of a broader product experience.

The llm 0.36 update also broadens the set of OpenAI models that can be referenced through the package. Adding entries for GPT-6 Sol and GPT-6 Luna gives users a more explicit way to select these model variants when working with the library. Because model names can become part of workflows, a stable naming convention helps scripts, configuration files, and examples remain readable. The reference #1702 is included as a pointer to the change in the project history.

The conversation flag is another important piece of infrastructure. In plugin systems, model providers often differ in how they accept messages. Some models are designed for a single prompt and response, while others can process a longer exchange containing prior user messages, assistant replies, or tool output. Without a clear declaration, a plugin may receive input that it cannot handle, which can lead to confusing errors or degraded behavior. Setting supports_conversation = False makes the limitation visible before a request is made, and raising llm.ConversationNotSupported when a single-turn model is given conversation-shaped input creates a predictable failure mode.

For developers, the practical effect is easier integration and clearer error handling. A system can choose a model based on whether it needs multi-turn memory, or it can detect an unsupported request and adjust the prompt, switch models, or inform the user. In health-related experiences, that predictability matters because automated advice should fail transparently when the underlying model cannot support the requested interaction. Overall, llm 0.36 combines product-facing momentum around LLM-assisted guidance with technical improvements that make model selection and plugin behavior more explicit.