v12.5.0 - Skills & Stack

Added

• A built-in skill library, ready on install. Skales ships with a curated set of working skills: disciplined bug diagnosis, relentless plan interviews, research with cited sources, test-first coding, handoff documents and more. Skills marked automatic are picked up when the task fits; the others run when you name them. Every skill shows its author and license on the card, the library is searchable with category filters, built-ins cannot be deleted (only switched off), and a skill you import under the same name takes over.
• Skills in chat, on demand. Mention a skill with @ or ask for it by name, and the agent loads the full instructions itself before acting.
• Style packs for Studio. 74 aesthetic references can be pinned to a Flow project, shaping typography, color and layout mood. Your brief always wins on content, and no logos or trademarks are reproduced.
• Flow knows its skills. The Flow landing gains two controls: pin up to three skills that ride along on every step, and pick a style pack next to the Brand Kit. Your choices show as chips on the first message, so a project tells you how it was set up.
• Ready-made agent teams. Organization offers five presets, Research, Build, Analysis, Content and Ops, each with sensibly prompted roles. One click creates the team and everything stays editable.
• Images carry their origin. A picture that really came from image generation wears an origin badge; text that merely claims a picture does not. The generation prompt sits in a collapsible section under the image instead of flooding the chat.

Changed

• The agent finishes what it starts. An answer that ends mid-plan is no longer the end of the turn: with open checklist items or an announced next step, Skales continues on its own instead of waiting for you to type "continue".
• Scheduled work resumes itself. A scheduled task that hits its time limit while still making progress queues its own continuation and picks up from the checkpoint, so overnight jobs no longer stall.
• One schedule, once. Creating the same schedule twice is refused with a clear message instead of silently doubling future runs.
• Downloads tell the truth. A downloaded file is checked against what it claims to be, so an error page pretending to be an image fails loudly instead of leaving a broken file behind.
• Websites that block the built-in browser now say so and point to the visible-browser login handoff instead of failing quietly.
• Long-term memory learns only from real conversations. Automated feeds and error messages no longer qualify as facts about you.
• Quitting means quitting. The power buttons close the app and its background helpers for real instead of reloading the window.
• Knowledge-graph import shows real progress instead of sitting at zero, every provider key can be removed, and a cancelled goal says so exactly once.

Fixed

• Push notifications tapped on the phone open the exact conversation they belong to, even when the app was fully closed.
• Attached documents with embedded pictures no longer flood the conversation with image requests.
• Image messages no longer show stray formatting fragments beneath the picture.
• The Flow file panel closes properly, and long file writes no longer end with the whole page pasted into the chat.