LLM 0.32 adds reasoning traces and server-side tools

LLM 0.32 is the project’s largest release since launch, adding visible reasoning traces, server-side provider tools, content-addressable SQLite logging, new models, and features enabled by the OpenAI Responses API. Reasoning models now stream their traces to standard error, keeping generated output clean for piping into other tools, with an option to hide those traces.

The CLI now supports the GPT-5.6 model family and uses GPT-5.6 Luna as its new default model. OpenAI integrations gain CodeInterpreter and WebSearch tools, while llm-anthropic 0.26 adds support for the Claude 5 family along with WebSearch, WebFetch, CodeExecution, and AnthropicMCP. A new llm openai endpoint command can send one-off prompts to OpenAI-compatible endpoints without logging them.

The Python API adds a model.prompt(messages=[]) interface for passing complete message histories directly, plus a streaming events system designed for mixed outputs such as reasoning text, response text, tool calls, and image attachments. A new llm-chat-completions-server plugin provides an OpenAI chat completions-compatible server, supported by a Git-style content-addressable message store that reduces duplicate logging across multi-turn conversations.