v12.3.0 - Flying Gecko

Flow checks its own work before handing it to you, long conversations keep their task, the calendar tells you what is actually wrong, and Skales stops getting in its own way mid-work.

Highlights

• Sign in to a website, once. When a task needs you logged in (posting on X, checking an order), Skales opens a visible browser window so you sign in by hand. The login is saved to the Skales browser profile and reused next time, so Skales works inside your accounts without ever seeing your password. Open a login window any time from Settings › Browser Control.
• Flow reviews its own output. After Flow builds or edits a deck, prototype, page or visual, it takes a picture of the result and looks at it the way you would, checking for text running off the edge, unreadable contrast, cut-off pieces or a collapsed layout. It quietly fixes one clear problem and shows you the corrected version. At most one extra pass, only when a Vision Provider is set up. Toggle under Settings › Vision Provider.
• Browsing is faster, cheaper and more reliable. Skales now reads and acts on web pages far more efficiently: clicks and typing land precisely, long browsing tasks cost noticeably less, and even smaller local models can drive the browser now.
• Long conversations no longer lose the task. The model now always works from the full conversation, sensibly condensed, and the original request stays anchored no matter how long the session gets.

Added

• Watch the browser work. A "Show browser window" switch under Settings › Browser Control runs the browser visibly instead of hidden, so you see each step.
• Multi-line scripts run without shell traps. Skales writes a script to a file and runs it with the right interpreter (Python, Node, Bash, PowerShell) instead of squeezing code through a shell one-liner. On Windows especially, that whole class of quoting failures is gone.
• Every file change has an undo point. Before overwriting, editing, appending to or deleting a file, Skales keeps a copy of the previous version, so a working script can no longer be lost to one bad change.
• Self-built tools carry a freshness mark. Each tool Skales built for itself remembers when it last ran successfully; a stale tool is tested before it is relied on and read before it is changed.
• Seed your memory graph. With learning enabled (Settings › Memory), the Knowledge Graph can be seeded in one pass from your saved long-term facts and a slice of recent chats, instead of only filling up reply by reply. Choose how much history to pull in, watch it count through, stop any time. Re-running never creates duplicates.
• Your desktop makes Studio visuals for a paired phone. A phone in remote mode with no provider key of its own can ask the desktop to build a Studio visual; the desktop generates it on the side and sends back the finished design, without touching your chat history.
• Friend Mode can reach your phone. Turn on "Mobile App" in Friend Mode channels and proactive check-ins land in your Buddy thread on the paired phone, with the same cooldowns and quiet hours, encrypted end to end.

Changed

• The Proactive operator says what it is waiting for. Instead of a bare "No initiatives yet", the operator pane now names what Skales is watching (upcoming meetings, failed tasks, blocked work, unread email) and when it last checked, so quiet reads as good news.
• Skales reads before it edits. Changing a file it has not looked at this session is refused once, with an instruction to read it first. Blind patches on working scripts were the main way tools got destroyed.
• Failing approaches end in an honest question, not a seventh attempt. When the same tool keeps failing, Skales is told exactly what failed and pushed to test one hypothesis at a time; if it still cannot move on, it stops, summarizes, names its best guess at the cause, and asks you for the one thing it needs.
• Unrestricted mode keeps its word. A few command rules used to stay blocked even in Unrestricted. They now run, with a clearly visible amber warning strip on the result card naming what was flagged, pinned to the exact command. Safe and Advanced are unchanged.
• The working glow is easier to notice. The ambient background that breathes while Skales works drifts further, on a slightly quicker rhythm, and sits a touch more present at its peak. Idle stays still, Reduce Motion switches it off entirely.

Fixed

• Work already done is not redone. After an interrupted long task, Skales keeps a live list of every file it wrote, verifies each is on disk, and continues from there instead of rewriting from scratch.
• Skales changes course instead of giving up. When a step keeps failing the same way, Skales names the failing call, blocks it, and tries a different approach; only if it still cannot move on does it stop, and the stop message names the step that jammed.
• The calendar tells you why it is empty. If none of your Google calendars could be read (usually a display name entered where the Calendar ID belongs), Skales now says which entries failed and how to fix them, instead of showing an empty calendar as if you had no events.
• The Memory page tells the truth about Knowledge Graph learning. The card now says plainly whether learning is on, and the switch that controls it lives right there with an honest name.
• Skales knows which Safety Mode it is running in, and answers directly instead of guessing or telling you to check Settings.
• Legitimate commands are no longer refused as dangerous. Command safety is now decided in one place, honours your Safety Mode everywhere, and only genuinely destructive commands stay blocked.
• Windows: scripts no longer crash on special characters. Script output uses UTF-8 by default, so an emoji or non-Latin character no longer dies with an encoding error.

---

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