LLM 0.36 makes single-turn model limits explicit

LLM 0.36 lets model plugins declare that they cannot accept conversation history. The command-line tool then rejects chat sessions for those models and raises a specific error when assistant or tool history is supplied. That makes a model’s interaction limits part of the interface instead of leaving callers to discover them through a failed provider request.

Simon Willison · Sep 22, 2026

Why it matters

Plugin authors can represent model capabilities more accurately than treating every backend as chat.

More context

A core-tool follow-up distinct from the previously covered llm-typesafe plugin. Try the capability flag when implementing a single-turn backend. The release also folds reasoning traces in Markdown log output and adds Sol and Luna model entries.

Keep in mind

Released software; model access requirements and usage costs remain provider-specific. These release notes document the contract, not independent compatibility testing of every plugin.

Original read