llm 0.36

What it claims

llm 0.36 adds three changes: two OpenAI model IDs, a plugin-level flag for single-turn-only models, and collapsible reasoning traces in llm logs Markdown output. The parts that matter for adopting it are the exact model IDs and how the single-turn guard behaves at runtime.

How to use it

New model IDs (#1702):

• gpt-6-sol for GPT-6 Sol
• gpt-6-luna for GPT-6 Luna

Single-turn models (#1692): a model plugin declares supports_conversation = False. The guard then applies at two points:

• Prompt time. LLM raises llm.ConversationNotSupported when the model receives assistant history or tool history. Tool-call history counts as well as prior assistant turns. Agent loops that feed tool results back to the model will fail on these models, not just multi-turn chat.
• llm chat. The command rejects such a model before the session starts, so there is no mid-session failure. Wrappers that call llm chat should expect a rejection at startup.

The page points to a "Models that do not support conversations" docs section for details. The first plugin to set the flag is llm-typesafe.

Logs (#1701): in the Markdown output of llm logs, reasoning traces are now wrapped in <details><summary> tags. Anything that parses or post-processes that Markdown will now see these HTML tags around reasoning content.

Limits

The page is a three-item changelog. It gives no install or upgrade commands. It says nothing about pricing or capabilities for the GPT-6 models, and nothing about backward compatibility for plugins that don't declare the flag. The plugin-author specifics are in the linked docs section, not on this page.