0.37.1

• llm openai endpoint add now works when extra-openai-models.yaml holds a list written in square brackets, such as the [] left behind by removing the last saved endpoint. Previously it wrote invalid YAML that stopped LLM from loading any models until the file was fixed by hand. Thanks, Steven Huang. #1845
• Passing both system= and messages= to prompt() or chain() in the Python API no longer drops the system prompt for models that build their requests from prompt.messages, including the default OpenAI Chat models. #1849
• Non-streaming prompts, such as those run with --no-stream, no longer crash with an AttributeError when an OpenAI-compatible server leaves usage out of its response. Thanks, Tao An. #1697, #1698
• llm embed --format blob now outputs exactly the bytes of the embedding. It previously added a newline at the end, which made the output one byte too long to decode. Thanks, Arjan Mossel. #1848
• llm logs no longer crashes with JSON cannot hold BLOB values when a tool result logged by an LLM version before 0.32 has an attachment stored in the database. Thanks, XQ. #1775
• llm logs -q no longer misses words at the boundary between two text parts, for example in a response that interleaves text with reasoning or tool calls.
• When a prompt passes several user messages using messages=, llm logs -q now searches all of them rather than only the last one.