llm 0.33 released with embedding key support, template chaining, and reasoning_summary options

Version 0.33 of the open-source command-line tool llm introduces several developer-focused changes. The project upgraded its OpenAI Python library dependency to version 3.x and replaced the HTTP client from httpx to httpx2, addressing underlying compatibility and performance considerations.

Embedding workflows now support per-call API keys. Both the command-line commands llm embed and llm embed-multi, as well as the Python APIs EmbeddingModel.embed(), EmbeddingModel.embed_multi(), Collection.embed(), and Collection.embed_multi(), accept a key= parameter. This passes the resolved per-call key to embedding plugins without altering shared model state, and existing plugins that read self.key continue to work via a compatibility fallback.

Template chaining is now supported by repeating the --template/-t flag. This allows users to combine templates in sequence, merging model configurations and options from one template with a prompt defined in another. For example, a user can save a model configuration with high reasoning effort and a prompt as separate templates, then chain them when invoking llm.

Reasoning-capable models accessed through the Responses API now support a reasoning_summary option with three modes: auto, concise, and detailed. This option can be used with the llm openai endpoint --responses command, enabling users to control the verbosity of reasoning summaries when exercising different models that emulate the OpenAI Responses API.