simonw/llm 0.36

on GitHub

New features:

• New OpenAI models: gpt-6-sol for GPT-6 Sol and gpt-6-luna for GPT-6 Luna. #1702
• Model plugins can now declare supports_conversation = False for models that only accept single-turn prompts. LLM raises llm.ConversationNotSupported when these models receive assistant or tool history, and llm chat rejects them before starting a session. See {ref}Models that do not support conversations <advanced-model-plugins-conversations>. The first plugin to use this is llm-typesafe. #1692
• Reasoning traces in the Markdown output of llm logs are now wrapped in <details><summary> tags. #1701

Bug fixes:

• Fixed llm logs -t crashing with KeyError: 'options_json' when displaying truncated logs with options or schemas. Thanks, rdslw. #1656
• Fixed AsyncResponse.log_to_db() raising sqlite3.ProgrammingError after a response had been awaited. Structured messages and provider metadata are also preserved when logging restored async responses. Thanks, Roli Bosch. #1668
• Tool parameters and nested model fields named title are now preserved in generated input schemas. Thanks, Christopher Pruijsen. #1679, #1680
• Constructing llm.Prompt directly without options now initializes the model's default Options object, fixing errors in providers that access option attributes. Thanks, Jane Li. #1028, #1663
• Closing an asynchronous response.astream_events() stream before completion now closes the underlying provider generator, allowing it to release streaming connections and other resources. Thanks, Shunsuke. #1659
• SQLite connections opened by CLI commands are now closed when the command finishes, including on errors, avoiding unclosed-connection warnings.

Documentation:

• Corrected the structured output example in the advanced model plugins documentation to use supports_schema = True. Thanks, Ethan Stoner. #1686