Release v12.3.5 - Isolated Agent & Scratchpad · skalesapp/skales

Multi-agent jobs show every worker and finish on their own, and finished work stops being sent twice.

Highlights

A living map of what Skales knows. The Knowledge Graph on the Memory page is now a real, explorable network. It settles into shape on its own, you zoom, pan and drag single nodes, and busier ideas sit larger with a soft glow. Labels fade in as you zoom so a dense graph stays readable, and an Expand button opens the whole thing full-screen.
Watch every agent in a team job. When Skales fans a task out to several workers, the chat shows one live chip per worker moving from queued to working to done or failed, with a running count and elapsed time. Click it to open the Tasks view.
Drop a video into chat and ask about it. Attach a video and Skales actually watches it: it samples frames across the clip and reads them together, so you can ask what happens, why a short is working, or how a clip was made, its hook, pacing, cuts and on-screen text.
Look through a camera. Skales can take a single frame from an IP, WiFi or Home Assistant camera on demand and tell you what it sees. Ask "is anyone at the door" or pair it with a schedule to check every few minutes and only ping you when something changes. It asks before looking.

A private scratchpad for each conversation. Skales keeps its own working notes and a private work folder per session, so files from one chat never collide with another, and a long task keeps its train of thought.
Isolated agents with a life of their own. A custom agent can run with its own memory, its own workspace and only the tools you hand it, never seeing your identity, facts or saved memory. Pair it with a pinned model, a bound FTP profile and a scheduled goal and it runs a project end to end.
Publish a folder over FTP. A new publish tool uploads a workspace folder through a saved FTP profile. FTPS (explicit TLS) switches on automatically when the server requires it, so hosts like Hetzner work out of the box.
Some Discover posts invite your agent along. A post can carry a "Let your agent explore this space" button: one tap prepares a visit, your agent reads the site's welcome file, looks around and leaves a short guestbook comment, with your usual confirmations. You review the request first, and each post runs once.

Notifications carry their real button. Announcements show the exact action label they were written with instead of a generic "Learn More", and update notices get a one-tap jump to the in-app update page.
Signing in by hand works everywhere. The visible sign-in window no longer announces itself as an automated browser, so Google and other providers accept the login instead of blocking it.
Toast messages stay put. Stacked corner notifications no longer leave an invisible gap that drifted the stack toward the middle of the screen. Old toasts are always cleared and strays are swept up.

A team job finishes on its own. After splitting a goal across workers, Skales could stall waiting for a nudge. It now picks the goal back up the moment the workers are done and carries it to the finish, whether or not the chat is open.
Finished work is never sent twice. Re-checking a completed goal could re-run steps, sending a file or message two or three times. A step that genuinely succeeded now counts as proof and is never repeated.
Ask for an image, get an image. Skales now creates the picture instead of handing back a description of what it would draw. With no image provider set up it says so and points you to Settings.

Skales updates itself. Signed and notarized for macOS; Windows and Linux builds attached below.