Simon Willison shipped LLM 0.32.1, a patch release for his command-line tool for accessing large language models, after fresh installations of LLM stopped working, he wrote on his blog.
The break traced back to OpenAI's own Python library, which dropped its use of httpx, a networking library. LLM depended on httpx but only installed it as a side effect of installing the openai package, so when OpenAI stopped requiring httpx, new LLM installs lost access to a library LLM still needed, Willison wrote.
The 0.32.1 release fixes this for now by pinning to a version of openai below 3. Willison said a more complete fix is planned for version 0.33, which will move LLM off httpx entirely in favor of a library he referred to as httpx2.
It's a small release, but a useful reminder for anyone shipping a command-line tool: a dependency you never declared, because a bigger dependency happened to pull it in for you, can vanish the moment that bigger dependency changes its own requirements. Pin what your tool actually uses, not just what currently shows up in the lockfile.