llm 0.32a1 Released: Fixes Tool-Calling Conversation Restoration Bug, Rapid Alpha Iteration Continues

Simon Willison's llm command-line tool just shipped version 0.32a1, primarily fixing a bug in the previous alpha (0.32a0) where tool-calling conversations couldn't be properly restored from the SQLite database. While the bug itself isn't huge, it touches on a core capability of the llm tool — conversation persistence and the integrity of tool-calling chains.

A Bug That Gave AI "Amnesia"

The bug fixed this time (issue #1426) is rather ironic: you ask AI to call a bunch of tools and do a bunch of work, then the next time you open the conversation log, AI stares at you blankly and says "I don't remember."

Specifically, llm stores all conversation history in a local SQLite database, making it easy to review and continue previous conversations at any time. But in version 0.32a0, conversations involving tool-calling failed when being reinflated from the database — the serialized tool-calling context couldn't be correctly restored into usable conversation objects.

It's like hiring an assistant who's great at getting things done, but gets amnesia every time they clock out.

What Is llm?

In case you're not familiar, llm is an open-source command-line tool and Python library developed by Django co-creator Simon Willison. It lets you chat with various large language models directly from the terminal — OpenAI, Claude, local models are all supported — and all conversation logs are automatically saved to a local SQLite database, so you never have to worry about losing chat history.

The biggest highlight of the 0.32 release series is the introduction of tool-calling support. Tool-calling is one of the core capabilities of modern large language models: the model can proactively invoke external tools during a conversation (such as search engines, calculators, various APIs), get the results, and continue reasoning. This is also the foundational capability for building AI Agents.

The Significance of Alpha Versions: Falling Gracefully in Public

The "a1" in the version number means this is an alpha version — an early test release. The subtext of alpha is basically: "I know there might be pitfalls, but let me put it out there for you to find them."

From 0.32a0 to 0.32a1, one issue, one fix — that's the daily life of open-source development. Simon Willison has consistently maintained this rapid iteration cadence: discover a problem, file an issue, fix it, release — the entire process transparent and public. Rather than calling this "fixing bugs," it's more like stomping out all the pitfalls before the official release.

For developers currently using the llm tool with tool-calling enabled, it's recommended to upgrade to 0.32a1 as soon as possible to avoid losing tool-calling context when restoring conversations.

Final Thoughts

Databases remember everything — except the things they forget — and that's probably why we need alpha testing. When your AI tool can't even remember what it said, at least there's someone willing to fall gracefully in public, then get back up even more gracefully and fix it.