Conol 1.2.2 — Lighter on its feet

This is a small release, and you won't see much of it. It changes how Conol saves a reply while the agent is still writing it, and that path runs constantly.

Streaming replies put far less load on Conol

Why. As the agent writes, its reply reaches Conol a few words at a time. Until now, Conol saved the whole conversation again for each of those pieces, so a single long answer could mean a thousand or more saves. Each save rewrote more than the last as the reply grew. Most of that work protected text the agent was about to finish anyway. Because it happened on every reply in every chat, it was one of the biggest loads on our database. It was also behind our outage on July 30, when the churn from those saves filled the database's disk.

Conol now collects those pieces and saves them together a few times a second. The first words of a reply are still saved and shown the moment the agent starts. After that, the reply catches up about four times a second. The end of each thinking step, each tool call and the finished reply are still saved right away, so nothing you see as complete is ever held back. In our tests, a reply that used to take more than a thousand saves now takes five.

How to use it. There's nothing to do. Chat as usual. On a long reply you may see the text arrive in short bursts instead of word by word. That's the batching at work. Behind the scenes, Conol has much more headroom, which keeps it fast and steady when many people are chatting at once.