llm 0.32rc1

Simon Willison has pushed out the release candidate for LLM 0.32, and it wraps up a redesign that's been brewing since the 0.32a0 alpha. The headline change is under the hood: a new database schema that captures prompt and response data with far more fidelity than before, especially for the newer model families that don't fit neatly into the old request-response mold.

The technical trick here is content-addressable hash IDs for stored messages. Instead of just appending rows to a log, LLM now identifies each message by a hash of its content. That sounds like a small implementation detail, but it unlocks two things at once: automatic de-duplication, so the same message doesn't get stored twice, and the ability to represent conversations as trees rather than straight lines. If you've ever forked a chat to try a different follow-up prompt, that's now something LLM can actually model properly instead of flattening into a linear history.

Because this touches the core schema, Willison is being careful about it. The change only adds new tables and shouldn't disturb any existing data, but he's still telling users to run a backup — a simple llm logs backup logs-backup.db — before installing the RC. It's the kind of caution that separates a tool built by someone who actually uses it daily from one shipped and forgotten.

Tucked into the same release is support for three new models: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. Willison doesn't dwell on them here, which suggests the schema work is the real story of this release, with model support along for the ride.