@theo@mattshumer_ Last night it failed to send a prompt to fable via cli so it opened chrome, logged in, and prompted it there. It forgot to change models to 4.8 haha. The 4.8 cli path still worked too
@kr0der I’ve had ultra sol grinding one prompt for 19 hours now on a 20x. Make sure you have it do orchestration. Manages context better that way and uses cheaper models when needed. I even call opus and fable to support it
@initjean I’m using sol ultra and have it doing the orchestration. It queues up 5.6 and opus agents for me via cli. Calls fables for consults. It’s about four hours in to its current goal. Usage windows keeping up with the burn rate so far
@theo@ChristianSelig Does t3 allow us to use our subscriptions, full computer use as well as control remotely from phone or a browser? Of note my dev machine is Mac
@thsottiaux Allowing the model to self report what model is being used.
Allow agents to make decisions based on usage windows. Currently no way to natively check remaining usage windows for the models
@theo I’m setting this up right now to use git as a bridge, a listener on my Mac, to drive cli to codex. Want to use fable to orchestrate more but dispatch sucks so using cloud Claude code. We’ll see how it works.
@saynedbread@davis7 I’ve seen it a few times in my Claude chat now. Basically sometimes the model uses Chinese chars because they can carry more meaning and thus be more token efficient. Sometimes they leak through into the output