Generation

A reply used to be written by the page you were looking at. That works right up until the page goes away, and pages go away for ordinary reasons: you reload, you follow a link, you switch apps and your phone reclaims the tab. The model kept writing, nobody was listening, and nothing was saved. The answer was gone.

Llooma now runs the turn in the server instead, and the page watches it.

What changes for you

Send a message and then reload. The conversation comes back with the reply still arriving, from wherever it had got to. Close the tab and open it again, in a minute or tomorrow, and the finished answer is there: the turn writes what it produces as it produces it, so nothing is waiting on a browser to come back and store it. Switch to another conversation while one is generating, and come back to it.

The confirmation dialog that used to ask whether you really wanted to leave a conversation mid-reply is gone with it, because leaving no longer costs anything.

Titles and automatic compaction moved with the turn, for the same reason. A first exchange that landed while the tab was closed used to come back untitled; now the conversation is named whether or not anyone was watching.

Naming it again

The first title is written from a single message, before anything has been answered, so it names the question rather than the conversation. Settings → Chat → Naming can give it a second name a few exchanges in, once there is something to name. Off by default.

Two rules make it useful rather than distracting:

• Once. A conversation whose name keeps changing stops being something you recognise in the sidebar, which is the only thing a title is for.
• Never over a name you typed. Renaming a conversation marks it as yours, and nothing writes over it after that.

In server mode it travels with the rest of the naming configuration: an admin sets it for themselves and shares it under Settings → Admin → Title generation, like the model that writes the titles.

Stop still stops

The stop button reaches the run wherever it is. What the model had already written is kept, as it always was: a reply cut short is more useful than an empty one.

Where your conversation goes

This is worth stating plainly, because it is a real change.

Nothing moves. Your conversation already passes through the instance: it holds the accounts, the connections and the provider keys, and the proxy is the path between the browser and the model. A turn that runs in the same process reaches the provider the same way, under the same admin rules about which models are shared and which instruction is locked.

What it does not survive

Honest limits, since a promise half-kept is worse than none.

A server restart, and only that. Leaving for a day is fine: the reply was written to the database as it was produced, not held in memory waiting for a tab to come and collect it. What a restart loses is the turn that was in flight when it happened: an answer half written is not in the conversation yet, because nothing has finished it. Everything that landed before is.

Several replicas. If you run llooma behind a load balancer with more than one instance, each one keeps its own runs, and a browser that reconnects to a different replica will not find the turn to watch. It will still find the answer, since both write to the same database.