Willison's "llm" Tool Stops Duplicating Chat Logs

Simon Willison shipped LLM 0.32rc1 on July 30, deduplicating conversation logs by hashing message parts instead of storing every near-identical array whole.

That's the difference between linear and quadratic storage growth for anyone logging agent traffic to SQLite over a long session.

OpenAI-style clients resend the entire growing message array on every turn. A logger that stores each turn's payload whole grows quadratically as sessions get longer. Hashing each part once and referencing it afterward keeps the growth closer to linear.

Willison also released a companion tool, llm-chat-completions-server, which exposes every model in an LLM install behind an OpenAI Chat Completions endpoint in three commands. He says GPT-5.6 Sol wrote the whole thing, with strong recall of OpenAI's own API spec.

The companion server is the bigger claim of the two. Willison credits GPT-5.6 Sol with writing the whole thing solo. But he cites the model's recall of OpenAI's spec, not a confirmed compliance check, and doesn't say how much he corrected before shipping it. Worth watching whether llm-chat-completions-server needs a patch in the next few releases, since that's the real signal on how much editing the output needed.

The dedup fix itself ships only as an 0.32rc1 release candidate, not yet in the "llm" tool's stable channel.