Skales - AI Agent & Tasks

Skales opens the screen instead of telling you where it is. Ask to be shown something and the app goes there: the Studio, the Cockpit, a team, a document, an older conversation. It only moves for a turn you started yourself; a job running in the background or a turn handed over from your computer names the place instead of pulling the screen out from under you. Four more places the conversation can reach. The Discover feed can be read and searched by author or by word, and a thread says how far back it looked. Your own widgets can be listed and read without being run. A group discussion can be started from the chat, behind the usual permission. And a team run can be asked what it is doing. Every chat says what your provider charged for it. A pill beside the context meter carries the running total for the conversation, and each answer keeps its own figure. The provider's own report comes first; where a provider reports tokens but no figure, the same published list your computer uses fills it in; and where neither exists the number is a floor with a question mark rather than one you might plan against. Tapping the pill opens a sheet with the session total, the ceiling you set, how many calls there were, and how much of the input came back out of the provider's cache. With a session ceiling set the pill turns amber at half and red at the ceiling, so the warning arrives before the stop does. The same total is named on that session's Analyze page. The ceiling of your computer reaches the phone. When a conversation on the computer hits the limit you set there, the question arrives here with its three answers - carry on, switch to a smaller model, or stop - and answering it costs nothing. The phone keeps its own ceiling under Settings > Advanced > Session budget. Work a provider interrupted comes back. If a turn dies because the provider fell over or hit its limit, the unfinished part is offered back with one tap, carrying what was already planned. Errors that would only fail again in the same place - a bad key, a message too large, a model that does not exist - are deliberately not offered. "Not now" stays answered after a restart. The plan belongs to the conversation it was interrupted in. The chat can look up your plugins and the state of the app it is in: what diagnostics say, what is running, what is parked, whether background work is allowed. Studio is six kinds again: Visuals, Image, Video, Animate, Music, Voice. Design, Scenes and Space have left the composer, because Flow already does all three and does them better. Everything you ever made with them stays in your feed - it still opens, plays, copies, exports, and can be run again or deleted. Only the way to start a new one is gone, and it is one screen over. Opening the app no longer starts a conversation. It begins with the first thing you send, as it does on your computer. Organization is back in the menu. Only Skales Local is called Beta now. Sharing a rendered clip to Discover starts with an empty caption. One language rule instead of two: a language you chose wins, otherwise your last message leads. A goal running on the phone stops and asks instead of spending its whole allowance, and a goal that stops on your computer now asks here too. Every picked voice is stored as its engine plus its own id. Fixed: a reply that failed names what failed instead of the same three words every time. A question in your language is answered in your language, read from the phone itself. A Studio visual keeps the background it was written with, an animated visual keeps animating, and every Animate card shows its film in the format you picked. Stop reaches a run on your computer. A model that sends several tool calls at once has them all run. A run on the phone keeps going when you leave the screen, and where the system stops it the card says so. A plan belongs to the conversation that wrote it. The menu shows the version you actually installed. Speech through OpenRouter plays.