v12.6.51 - Name The Failure

A patch about failures that did not say their name. A browser that would not start, a sign-in that stopped with three words, a settings page that took the whole app down with it, and a picture that came back as an error nobody could read. None of these were new. What was missing in every case was a sentence.

Fixed

• The settings page no longer takes Skales down with it. On some Windows installs, opening Settings ended the engine outright: the page went blank or the window closed, Skales quietly restarted itself, and it happened again on the next visit. The cause was the browser automation library, which had moved on to a newer version than this build can run and stops the program the moment it is loaded. Skales now uses one exact version, checks before loading anything, and refuses to load something that would end the session. The "Install Chromium" button can no longer pull in a version that breaks the app it is meant to fix.
• A browser that cannot start says why. Missing Chromium, no system Chrome, a check the machine would not let Skales run: each of these is its own sentence now, in your language, with the two ways out named. Browser actions that fail can no longer end a reply in silence.
• Signing in with ChatGPT explains itself. "OAuth flow failed." is gone. The port being taken, a firewall refusing the local listener, a browser that would not open, a sign-in left too long: each says what happened and what to try, in your language. The port is checked before your browser opens, so you are not sent to a sign-in page whose answer could never arrive.
• A window that dies is written down. If a window or the engine stops unexpectedly, Skales records what happened and what it was doing at the time, and the window comes back with the reason on screen instead of vanishing.
• Images reach local models again. Pictures sent to a local model through chat, Telegram or WhatsApp could come back as an unreadable error, depending on where the picture came from. All three now prepare images the same way.
• An older channel chat shows the model it is actually using. Telegram and WhatsApp chats always followed your current model choice; the label just kept showing the one from the day the chat started. A model you picked for a specific chat is still yours and is left alone.
• The model search finds an endpoint by the name you gave it. Typing "LM Studio" found nothing, because the search only knew the internal id. Your endpoints now appear under their names everywhere, and the internal id is gone from the screen.
• Local models get their own settings again. A newer Gemma release was quietly running on settings meant for a much older one, and being handed a shortened prompt it did not need. Measured against the real model and corrected. The shared profile library and the copy built into Skales can no longer drift apart, which matters most where the library cannot be reached at all.

Added

• A "Free" filter in the model picker. Shows only models nobody is billed for: free variants and anything running on your own machine. Worked out from the model lists themselves, so it does not go stale.
• Custom endpoints remember their model list. Ask once, with a Refresh button and the age of the list next to it, instead of asking your endpoint every time you open the picker.
• You can see how much of a request was cached. The token tooltip now shows the share the provider served from its cache, when the provider reports it.
• The model list tells the truth about what it can do right now. When a smaller local model is given a reduced set of tools, it is told so, instead of promising an ability it was not handed.