Skip to content

Personal AI agent for macOS, Windows, Linux, Android & iOS. Set a goal, it works alone: coding (Skales Code), desktop + browser automation, autonomous scheduled Tasks. Teams of agents and humans across paired desktops. Agent2Agent, MCP, computer use, voice, local LLMs (Ollama), Agent Skills (SKILL.md). 26 AI providers, BYOK. No Docker, no terminal.

The published changelog is the one the build carries

Mario Simic2ae05a46231e
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2d34b69..2050893 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,86 +8,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## v12.9.30 - DeepDive
-### Changed
-- **The licence question is asked before the update is delivered, not after it.** Automatic downloads are on by default, so a version that comes under a new agreement used to arrive downloaded and installed, and the licence screen came afterwards - the decision landed after the delivery. An update that crosses the licence boundary now holds: the notice names the change, links the text, and the Download button is the decision. An installation that already runs on this agreement updates exactly as before, and somebody who presses Download themselves is never held up. The one group this cannot reach is the one already on the old licence, because the holding happens in the build that does the holding - so the release feed itself now carries the notice, which is the only thing an older Skales shows you that we write at release time.
-- **Two lines stand between the licence and the buttons.** Private use stays free of charge, as before; commercial use needs a licence, and that was true under the previous licence too. They sit below the text rather than above it, because they are the last thing read before the press. There is nothing to click that classifies you as private or commercial: a press cannot make that true, and a promise that "nothing changes for you" would be a promise to a group the app cannot identify.
-- **Declining says what you keep, and stops handing out an old build.** The refusal screen shows the data folder, untouched, and now says the thing that is actually true and costs nothing: an installation that was already running is unaffected and stays on the licence it was supplied under. It names that licence - the Business Source Licence 1.1 - and quotes the part of it that has not changed: its Additional Use Grant covered personal use, learning and non-commercial internal testing, so commercial use needed a licence there too. What it no longer does is offer a download of a superseded release. The same screen is shown to somebody installing for the first time, a superseded build gets no security fixes, and an installed copy can never change the address it shipped with - so the one link goes to the published licence and nowhere else.
-- **The app carries no price, and it never will again.** The Business page held twelve euros a seat in twelve languages, in a block where the numeral and the unit were two type roles - and a website can be corrected every day while an installed copy keeps what it shipped with until somebody updates. In its place stands what is true today: an evaluation licence, free, the full scope, no seat limit, confirmed in writing with an end date, which is a written commercial licence at no charge rather than a stretched trial. The commercial terms come down the same live wire as the figures, so setting a price later is a change on the server and not a release, and a machine with no connection shows the honest statement instead of a number that has aged.
-- **The Business page stops saying things it is not allowed to say.** "Free, with your own key" now names which half is free where it is said: your own key decides who pays the model, the licence decides whether you may use Skales commercially. The line about nothing being locked in no longer says "stop paying and the free agent keeps working with your own key" - a sentence about the end of a payment, with the licence unmentioned, reads on a business page as a free commercial alternative. And the waiting list says at the point of consent that Skales may write about commercial licensing, with what the box asked for stored beside the privacy notice version it was agreed against.
-- **What the offer carries instead of a badge: your logo on our page, and your name to your own clients.** The fifth line of the evaluation licence used to promise a founding badge on an account and thirty minutes on the phone. From v13 it is a company logo under "Trusted by" on skales.app, linked back to your own site, and your name listed there as a qualified contact for setting up your clients - a link somebody can follow and a mention somebody can act on, instead of a mark inside an account only you can see.
-- **The v13 section says what is coming, not what is missing.** The band over it read "not built yet", and the paragraph under it listed what the download you can get today does NOT contain. That it comes with v13 already says it is not here now. Nobody reading a business page should have to learn our release plan to understand an offer.
-
-- **FFmpeg is in the third-party notices, with its licence and where to get its source.** Skales uses it to merge audio and render video, and it ships in the package - but only the small JavaScript wrapper around it was ever attributed. The build that ships carries the GPL, version 3 or later; FFmpeg runs as its own program and is never built into Skales, so those terms cover FFmpeg and not the application around it. Its full licence text and FFmpeg's own notice now travel with every download, and the notice says where the matching source can be had.
-- **Everything in the download is attributed, and a new piece cannot slip in unattributed.** The notices used to be assembled from a list that could not see the programs added to a package last. Whatever ends up in a download now has to be named and described before the build will finish.
-- **The third-party notices read as the public document they are.** The file ships with every download and is published beside the licence, and it was carrying a block of internal notes and a handful of internal file paths out with it. Both are gone, and it stays that way.
-
-- **A picture a tool took is shown in the coding window, under the call that took it.** Until now a screenshot taken during a run reached Skales Code as a message from YOU: pictures travel to a model in a user-role slot, because that is the only slot every provider wire has for an image, and the window read that transport detail as authorship. So a live check with two screenshots in it produced a turn on the right-hand side, in your name, carrying the fence written for the model - with an edit button and a send-again button on it. The chat has told the two apart for months; the window now asks the same reader, and goes one better than the chat can: it has a line per step, so each picture hangs under the call that produced it rather than under the round. A tool that merely shows a picture again does not draw it a second time - it says it showed one from further up - while two captures of a screen that did not move stay two looks at the screen, however identical their bytes.
-- **The pictures you attach to a coding turn are in the transcript.** They were counted ("2 pictures") while the session was live and vanished entirely once it was reopened, because saving replaces the parts with plain text and keeps the files beside the session. The same thumbnails the composer showed you before you sent are in the log afterwards, and after a restart, and any of them opens full size.
-- **A block of code in an answer is finally treated like code.** It arrives tinted by the same scanner the file viewer uses, so a fence and the file it was cut from are not two different colour schemes; it carries the language it claims to be; and it has a copy button - the same button the open tool step has, not a second one that looks like it. It also moved off the dark terminal plane onto a card, where those colours were chosen to be read; the terminal plane now belongs to actual command output, which is what it was for.
-- **You can see what the model thought.** The window has offered a reasoning-effort dial for months and then showed a blank column while the effort was being spent. The thinking is folded under the answer after the turn lands, and it is open WHILE the turn is still only thinking - it closes itself once, the moment the answer starts, and after that the fold is yours.
-- **A file a step wrote is a card, not a path in a sentence.** Name, type, size, download, open the folder - the card the chat has had, with the check that the file is still there, drawn in the coding window's own colours.
-- **A turn that looked something up brings its cards with it.** Repositories, articles, places and their map render in a coding session the way they do in a chat.
-- **One card, drawn twice, is one card.** The three shared cards above are the chat's own components mounted inside the coding window through a token bridge: the window defines the names those components read out of its own palette, so they inherit Skales Code's colours in light and dark without a chat value ever reaching them. The alternative was seven hundred lines of card rebuilt a second time, and two copies that disagree with each other within a release.
-
-- **The + that starts a new chat is the one filled button in the chat header.** It sat in the same grey outline as the six tool buttons beside it, so the way OUT of a conversation looked exactly like the controls that act on the one you are in - seven identical outlines, and the one that leaves is not the same kind of thing as the six that act. It is filled in the accent of whatever theme and accent tone you run, and the lettering colour is COMPUTED against that accent rather than assumed dark, which is what keeps it readable under Skales X, Classic and Flat, in light and in dark, and under an accent you picked yourself. It is the only filled control in the row on purpose: a second one would take the distinction away again. Nothing about its size, its spacing, its radius or its icon changed, so the row still collapses the same way on a narrow window, and it is the same button in the compact window.
-- **Wrapped stops being a switch you can forget you turned off.** It never had a menu row - its one door is the button in the Discover header beside Spark - so a switch of its own could do nothing but take that button away and leave the page unreachable from the navigation. The switch, its card on the Add-Ons page and both readings of it are gone, the button is drawn for everybody, and whoever does not want Wrapped does not open Discover.
-- **The Playbooks switch is gone, because it decided nothing.** Recording and replaying live in the chat and inside the Browser surface, and the two tools were always handed out with the browser tools - so the switch on the Add-Ons page changed nothing except what Skales SAID about playbooks, and it could say "switched off" while playbook_run sat in the very same turn. The page, the palette entry and the tools are untouched; what is gone is a control that never controlled anything, and the assistant now reads the Browser Control switch when it answers for them.
-- **Organization is opened from the chat, and from nowhere else.** `/organization` is its door on this machine exactly as it is on the phone: no sidebar row, no palette hit, and a card that names the command instead of a menu entry. The switch keeps everything it gates - the team tools and the team block the assistant reads.
-- **The sidebar stops asking every minute about a badge it cannot draw.** It fetched the weekly summary once a minute, for the whole time a window was open, to light a dot on the Wrapped row - and that row has not been drawn since the entry was taken out of the navigation. The request, its timer and the dot are gone; the Notifications dot beside it is untouched.
-- **The Add-Ons page and the first-run picker stop promising doors that were closed.** Two entries still said Organization and Playbooks had their menu rows back and one said Wrapped's row follows its switch, months after the rows were taken out; the check that would have caught it existed only on the phone. It exists here now: an add-on offered during setup whose whole promise is a page has to name a door - a row, or a written reason why it has none - and the same reading found one more card that could never have shown the mark. What Skales tells you about Templates, Organization, Wrapped and Playbooks matches what the windows actually do again.
-
-- **The Plugins page says what a plugin is for, instead of how it is stored.** It opened on "a plugin has its own menu entry, its own page and its own storage" and an empty text field, and the one route that actually works - describing what you need in a chat - was named only in the empty state, which is to say only until the first plugin existed. The page now opens on what you get, and under it a section that explains the whole thing once: what a plugin is, that anything which would send, publish or spend waits for one click from you even at three in the morning, and the three ways one gets here as three cards rather than one button and two sentences. It closes with an X in the corner and does not come back - the flag sits in the settings file next to every other machine-level preference, so it survives a restart, a cleared cache and a restored backup rather than living in browser storage like the four other one-time notes in this app.
-- **The shelf and the catalogue are one card language now, not two.** The same object was drawn as a full-width row in one place and a two-column grid card in the other, with different padding, different glyph sizes, and permissions rendered once as a sentence with middots and once as capsules. Both are the same card, the same glyph tile and the same pills, from one shared file - and what a plugin may do reads as separate facts instead of a run-on line.
-- **Which tools a plugin may use is a question you can answer.** The picker was a flat list of up to two hundred bare names; the category each tool belongs to was being loaded and never drawn. It is grouped now, with a count of what is on per group, the house switch instead of a raw browser checkbox, and a line naming what is never on offer at all.
-- **The page holds together on a narrow window.** The row's controls used to sit in a column that never wrapped, so as the window got smaller the description was squeezed into a strip one word wide. Everything below the glyph now shares one column that wraps, the tool search takes the full width when there is no room beside it, and the sentence explaining a tool is the first thing dropped rather than the thing that crushes the layout.
-
-- **A plugin page can have icons now.** It never could: a plugin runs in a sandbox the app's own icon components do not reach, so all ten bundled plugins shipped with not one glyph in them, which is most of the reason they read as documents rather than as tools. Sixty-two lucide glyphs are available to a page by name, handed in the same way the house fonts are - only the ones a page actually asks for, so a page without icons costs nothing. A name this build does not know still draws the plugin square, because a `use` pointing at nothing renders nothing and a row that loses its glyph shifts for a reason nobody can see. The geometry is generated from lucide rather than traced, and the build fails if the two ever fall out of step.
-
-- **A plugin now follows the theme in shape as well as in colour, and it keeps following it.** The frame handed a plugin page seven values, and all seven were colours - so a plugin looked exactly the same under Flat, which is square and unshadowed, as under Skales X, which is neither. The radius, the shadow and the easing travel now too, with Flat's own values for them arriving for free. And switching theme or accent while a plugin is open used to leave that page on the colours it opened with until you navigated away and came back: the new values are pushed into the running page instead, without reloading it out from under whatever it was doing.
-
-- **Every plugin page now stands on the same floor.** There were three: the ten bundled plugins carried a real little design system, a plugin started by hand got twenty lines with its own hardcoded colours and a fixed English sentence, and anything the chat built got eight lines - box-sizing, no margin, a colour and a font. The one that was good was reachable from ten files and nowhere else. It now lives in one place, is laid under EVERY plugin page by the frame, and has grown the parts that were missing: a head band with a glyph, a title, a sentence and one primary action; sections and toolbars; rows that lift under the cursor and give way when pressed; a figure with its label for the counts; a skeleton in the shape of what is loading instead of a spinner over a blank rectangle; a failure band that is not the same thing as a remark; a focus ring; and a staggered entrance that is off entirely for anyone whose system asks for less movement. It is a floor and never a ceiling: it sits in the head, so a page that styles something itself still wins, and it contains no `!important` anywhere. A page that never thought about its edge gets one, and takes it back with one line of its own.
-- **A plugin you start by hand opens on the house look, in your language.** The blank page it creates carries the same floor every bundled plugin does and says what it is in the language the app is set to, rather than in English regardless.
-
-- **The Business page now shows what Skales already does instead of claiming what it might.** The page opens on one night's work - receipts read out of the mailbox into a month sheet, an invoice chased with the reminder already written, a post published with its image and its SEO, a summary sent to the phone - with the tool behind every line named. It says out loud that the reminder was drafted and not sent, because a scheduled run has nobody at the screen and sending waits for a click. The old opener promised an agent that "finds clients", and there is no lead, no source and no duplicate check anywhere in the tree to make that true.
-- **Six things Skales already stands for, before anything about a price.** It runs on your machine; your key or no key at all; Skales IQ carries zero data retention on the request itself rather than as an account setting; it asks before it sends, publishes, deletes or spends; nothing locks you in, since a plugin exports as one file and the free agent keeps working with your own key; and erasure is a branch that runs, with a 36-month sweep behind it. Underneath them, the boundary in one line: this is not a certificate, and what a procurement reviewer gets is a document naming the providers a request can reach.
-- **Skales Business is on the page, and marked as not built.** Workspaces, cost control with a usage journal and budgets that reserve before a run starts, infrastructure profiles for a company gateway, department packs, joining by invitation, SSO and SCIM, and an admin portal that runs without us. It is drawn as a rail with hollow marks rather than a row of cards with ticks, because a tick is a claim that something is there.
-- **The free half is named before the paid half.** Ten plugins, every tool, triggers on the clock, a folder and the mailbox, and any provider key stand in their own column beside what the seat adds. A buyer does that subtraction anyway.
-- **The offer is an object rather than a sentence in gold.** What it is stands at the display step with its scope beneath, and the five conditions run down the same object as a hairline column instead of five full-width cards that outweighed the offer itself. The geometry does not care whether the left cell holds a name or a figure, which is what lets the terms arrive over the wire. The partner terms and the v13 sunset are two statements on two lines, not one that trails off.
-- **Enterprise and white label read as an invitation:** "Tell us what you need and we will talk about it. Starting in 2027."
-
-### Fixed
-
-- **A placeholder cannot reach the screen by itself any more.** The sentence that names a licence change was one of two places where a line asked for a piece of text and never handed over the value it leaves a gap for, so twelve languages would have printed the gap instead - and the other was the FTP result in Lio, which put the file count in front of a sentence that already had room for it inside. Both say what they mean now, and the check that holds the twelve language files against each other reads the code as well: a line that asks for a sentence with a gap in it has to fill the gap.
-
-- **The licence notice reaches the page it was written for.** An update that comes under the new agreement announces itself once and then waits four hours for the next check - and the page carrying the notice is reached by FOLLOWING that announcement, which is a navigation. By the time it drew, the announcement was gone, and the only thing left that still knew was the main process, which was not being asked about the licence at all. The notice was therefore missing on the one route anybody actually takes to it, leaving the Download button standing where the notice should have stood in front of it. The main process now answers that question too, the download's own progress no longer drops the answer, and the sentence names the version instead of printing a placeholder at the reader in twelve languages.
-
-- **A link preview stops printing `&` at you.** GitHub escapes its own description twice, so "Android & iOS" arrives as `Android & iOS`, and one pass of decoding left the second escape standing in the middle of the sentence. It is decoded to the end now, in both the chat and the coding window, without ever turning text a page escaped ON PURPOSE into markup - and four entities the reader had never heard of (`'`, `<`, `>`, ` `) are read too.
-
-### Removed
-
-- **The install count, the monthly-active count and the star count.** The install figures came from the public aggregate's `installs_lifetime` and `installs_30d`, which are the count of DISTINCT IP ADDRESSES that asked the update endpoint for a version - the same source reports 639 devices active in thirty days. A figure nobody can explain when asked what it measures does not belong on a page that asks a company to sign something. What is left is countable: tools from the generated manifest, providers and models from the aggregate, countries, and languages from the locale files.
-- **The Revenue Dashboard tile and the "lead scoring" claim.** Neither exists anywhere in the tree. The usage dashboard arrives with Skales Business and is listed there instead.
-- **"A growing library of business plugins - included."** It promised plugins that do not exist, in the row of things you get. The five that do ship - Open Invoices, Receipt Inbox, Notice Radar, Meeting Prep, Page Watch - have a section of their own now, on the free side of the line where they belong.
-- **"...and nothing to bill"** from the Skales IQ line. Beside something a company pays for it read as "free", and it was not true.
-- **"WHAT PEOPLE ARE BUILDING".** Six unsourced claims about customers, under a heading that presented them as real.
-
-### Fixed
+**A release that shows its work, under a licence of our own.**
-- **The weather line names the place that was found, not the words you typed.** The lookup has been forgiving for a while - it tolerates a misspelt country and a town the gazetteer has never heard of, and works its way outwards through the rest of the field until something matches - but the greeting printed the field back at you, so somebody who had typed "Aregentina" read their own typo beside a temperature that belonged to a place the app had correctly identified as being in Argentina. The line now says which place the temperature is for, and says the same thing on the second load as on the first. What it does when the field is EMPTY is untouched: no place is named at all, exactly as before. What it does when the field names a place nobody can find is untouched too, and deliberately - there the words you typed are the only honest thing to show. The assistant's own weather answer and the phone always named the found place; this was the one surface that did not.
-
-- **The first participant in a Group Chat is readable in light mode.** Participants are told apart by colour, and the first one uses your accent - but its name tag on the opening screen, its letter badge in the setup and its whole message card were filled with the accent AND lettered in the accent, which is not low contrast, it is a solid block with the writing invisible inside it. The four participants below it, which draw from fixed blue, orange, purple and pink, were always right. The same pairing sat on the card that carries the closing verdict, and that card belongs to no participant, so it was unreadable in EVERY round rather than only in rounds that had a first participant. All four places now tint the fill and keep the accent for the writing, which is how the blue and orange participants beside them have always worked. Dark mode never had the fault and is untouched, line for line.
-
-- **The mini chat window stops putting its header under the system's own buttons.** Every other window in the app leaves a strip clear at the top for the buttons the operating system draws over the page - the close/minimise/zoom lights on macOS, the caption buttons on Windows - and the compact window was the single one excluded from it, so on macOS the lights sat on the agent avatar and on Windows the caption buttons sat on the microphone, screen, + and expand controls. Two different corners, one missing reservation. It reserves the same strip as the full window now, from the same measurement, so the two cannot drift apart; on Linux, where the window keeps a real title bar and nothing is painted over the page, the strip is nothing and costs nothing. In the same place: the invisible band you drag the little window by was laid OVER the header rather than above it, and a drag band takes the click - so the top half of every button in that header quietly did nothing. The band IS the reserved strip now, and it stops short of the system buttons on either side, because a right-to-left Windows install puts them on the other one.
-
-- **The figures under an answer are readable on a phone again.** Hovering the model line showed a box that ended mid-word, because it was drawn inside the message bubble and the bubble clips what leaves it. The breakdown and the bar's own figures now come from the one tooltip the app has had all along, which hangs off the document rather than inside the bubble and is placed against the line it explains: nothing is cut off at any window width, and it clears itself when the transcript scrolls instead of sticking after a tap. The numbers and the sentence a screen reader gets are unchanged.
-
-- **An answer that writes an HTML tag keeps the words around it.** A tag in an answer was not shown and not spelled out either: it was dropped, and with it every character it covered - a sentence lost the piece the tag sat in, and a tag on a line of its own took every line under it off the screen until the model wrote a blank line or closed it. On a long answer that is most of the answer, arriving one visible line at a time. The tag is now shown as the characters the model typed, so nothing leaves the screen. Code blocks, comparisons like `a < b` and plain links are untouched.
-
-- **A picture is read by the model you picked for the chat, even when the caption asks for work.** With the Advisor Strategy on, the picture turn was sorted by the caption alone - and a caption like "analyze this" or "find the error" reads as a task, so the planner answered instead of your chat model, which is exactly what that strategy was fixed not to do. A turn carrying a picture now goes to your chat model, whatever the caption says.
+DeepDive is named after what the coding window does now. Instead of a summary and a path in a sentence, a run puts what it actually produced on the table: the picture a tool took, under the call that took it; the thinking behind a step; the cards a lookup brought back; code finally treated like code. That habit runs through the whole release, from the transcript to the plugin shelf to the Business page, and it is the reason a hundred and forty five fixes are listed here one by one instead of as "various improvements".
-- **A picture no longer costs a capable model its tools.** Whether a picture turn keeps its tools was decided in the window, against the model shown there, and then held for the whole answer - while the model that actually answers is settled later and, with the Advisor Strategy on, is never the same one. So a model that can see the picture and act on it was disarmed by a verdict about a different model. The question is now asked of the model that answers, at the moment it is known, on every surface that takes pictures.
+Skales also ships under its own licence from here on. Private use, study and teaching stay free of charge, exactly as they were, and commercial use needed a licence under the previous terms too. What changes for somebody already using Skales is nothing at all. What changes on our side is that it is written down: the whole agreement, in the language you chose, in front of you once rather than folded away, and the release you already have keeps the terms it was supplied under.
-- **Joining through the Business page is now recognised on the Skales Plus page.** Both routes write `skalesplus_waitlist_email` into the settings file, but the Skales Plus screen only ever read the browser's local copy - so somebody who joined through Business was shown an empty form and asked for the same address a second time. It reads the same authority the Business page reads, keeps the local copy as the fallback when the settings file cannot be reached, and draws nothing decisive until it knows, rather than offering a join button and taking it away a tick later.
+DeepDive opens a new season, and the record starts here.
### Added
@@ -106,75 +33,38 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
- **Nothing that talks to the model opens before the licence is answered.** The lock screen held the main window, and the desktop companion, the on-screen pointer overlay and the introduction story opened beside it regardless: a machine that never accepted could use Skales through any of the three. All three now wait for the same answer, and they come up the moment it is given, without a restart.
- **The licence screen now says when the text in front of you is the English one.** Two complete texts ship with Skales, English and German, and every other language is served the English one until its own translation is ready. Nothing on the screen said so, so a reader whose app was in Spanish, Turkish or Japanese met a wall of English between a Spanish, Turkish or Japanese heading and a Spanish, Turkish or Japanese Accept button, with nothing to explain it. The line under the text now names it: the binding full text is shown in English, and a translation into your language will follow. It is shown only where it is true - a German or English reader is looking at their own text and is told nothing of the sort.
+
- **Deleting a conversation is not final any more.** A chat you delete was gone the moment you clicked, and the cards, the plan and everything else that belonged to it went with it. A delete is now a move: the conversation goes to a trash under Settings > Security and can be put back for a month, with the list saying what was removed, when, and how long is left to change your mind. Putting one back brings the whole of it - the transcript and the helper cards that hung under the answers - and it is in History, in the search and in the export again, where it was. Clearing a selection in History goes there as ONE row, named by how many conversations are in it, so one press is undone by one press instead of eighty; it takes one slot, so it cannot push out what you deleted yesterday. The trash empties itself, so nothing you deleted quietly lives on for ever. Incognito chats are not kept, because the badge said they would not be. Deleting a conversation stops whatever it was doing first, whether you delete one or a whole selection, so nothing keeps working and keeps costing on something you threw away - and what is kept for you is the finished answer rather than half of one. A conversation you put back comes back as a conversation and not as a job: it is where it was, it is idle, and it waits for you.
- **A task says what it actually did.** A task that ran on its own left one sentence when it worked and one when it did not, and nothing in between: which tools it used, in what order, and which one failed were gone the moment the run ended, so "why did it not send the summary" had no answer. Each task now carries an account of its recent runs under itself - how the run ended, when it started, how many tools it used and how many of those failed, then the steps in order with the failed ones marked. A run the app was closed in the middle of still has its record. Deleting a task takes its account with it.
- **Iris shows the helpers a spoken turn sent out.** Ask the voice surface for something that several agents share out between them and the eye showed a single line, while the chat window, the Code window and the phone all drew a card with one chip per helper and what it was really doing. Iris now shows that same card under the spoken line - live while they work, and still readable when they are done.
-### Changed
-
-- **Your memory has a ceiling now, and the Memory page says what it is.** The three record folders behind the counters grew by one file per turn and nothing ever trimmed them: on this machine 519 recent notes, 556 long-term records and 598 journal entries in eight weeks, of which 465 were written overnight by the dreaming pass rather than by anybody. That is not a disk problem, it is a reading one - the list is read from start to finish every time Skales builds a prompt, and at thirty-six thousand files that is a sixty-five millisecond pause on every single turn. Each folder now keeps its newest thousand records. Nothing you can still reach is lost by that: the prompt reads the newest five, the Memory page shows thirty, and the search reaches four hundred, so long-term had quietly passed the point of being findable months ago. What was filed on purpose is never rotated out, however old it is: a fact you asked Skales to remember, a todo or calendar note you gave the pointer, a pinned record, and the closed records that let a fact show its history. In the long-term store only the overnight pass is ever trimmed. One line under the three counters says all of this on the page itself, and Skales knows it too, so asking whether it keeps everything for ever gets an honest answer.
-
-- **Building a prompt stopped re-reading the memory folder every time.** The list of recent notes was read fresh off the disk on every turn, and twice on a turn where the profile is withheld - once for the turn, once for the rebuild. It now reads through the same short-lived cache its sibling has used all along.
-
-- **The sidebar opens with four entries you cannot lose, and five you can.** Dashboard, Chat, Code and Studio are the floor of the column now. Teams and Iris Orbit used to stand there too and could not be taken away, which is the wrong rank for a messenger a lot of people never open and a voice surface behind its own switch - both are ordinary rows now, and both can be sent into More like anything else. Nothing moves on a machine that is already set up: whatever your column looked like, it looks the same after the update, in the same order. A fresh install opens with Voice, Memory, Discover, Business and Skales Mobile under the four, and takes any of them away with one press. Custom Skills and Plugins are no longer handed out at the start - both are shelves for things you have built, both are empty on a new machine, and a row that opens onto nothing is how you learn that the rows are noise. They are one tap away under More, and whoever already had them keeps them.
-
-- **Templates moved out of the column and into the System button.** The library of starting points and the Guide answer one question between them - what can I start from, and what can this machine do - and they stood a whole column apart. Templates is now a row in the same popup, directly above the Guide, and it still follows its own switch on the Add-Ons page.
-
-- **Pointing Skales Code at a repository no longer throws you into a session.** Picking a folder with nothing typed opened a session on the spot, which took you off the landing before you had chosen a model, attached a file or picked an agent - off the one page whose whole job is making those choices. Picking now binds the folder and leaves you where you are: the chip names it, the branch appears beside it, and the @-menu lists the repository's files, all before a session exists. A press on the x lets the folder go again. The way in is the send button, or the "Open the session" button that appears beside the folder for the times you want the file tree and the terminal without typing anything first. Opening a folder from outside - `skales .`, a folder dropped on the dock, the folder's own context menu - still opens a session straight away, because that is a different request.
-
-- **A session started from the Skales Code landing keeps what you set there.** The model you picked, the effort you dialled, the mode, the agent, the extensions, the skills you attached and the files you dropped were all read from a stale copy the moment the session opened: the model was never pinned, so the composer went on naming it while the run used the settings default; the effort quietly fell back to medium; and anything staged without a message was cleared. All of it is now written onto the session before it opens, and read back from there - so the phone opening the same session finds the same composer. If the model could not be pinned for any reason, it is no longer shown as pinned either.
-
-- **Every question Skales asks before doing something now looks and reads the same.** The permission prompt, the offer to install a language checker, the offer to import another editor's MCP servers, the chat's approval card and Flow's all asked the same four things - allow this once, allow it for this session, allow it always, refuse - in five different shapes, and in three different orders: the refusal was last in the Code window, on the right in the chat and on the LEFT on the phone. They are now one card with one order everywhere: one answer per row, full width, the narrowest yes first and the refusal last and set apart, and each answer carries a plain sentence inside it saying how far it reaches and for how long. The language-checker card, which used to offer four bare words, says it too.
-
-- **The Skales Code window's title bar no longer breaks apart on a narrow window.** A session title runs to sixty characters and the bar is a fixed 44 pixels that does not wrap, so on a phone-width or portrait-tablet window the title spilled out of the chrome. It is now cut with an ellipsis at every width; below 720 pixels the repository and branch step aside, and below 560 the title does too - both are still on screen, in the status bar at the foot and in the history column.
-
-- **Wrapped is opened from Discover, not from the sidebar.** The weekly card draws its week from the feed and its Sparks, and a row of its own in the Main list was a door standing away from the thing it belongs to. The row is gone; a Wrapped button stands in the Discover header beside Spark, follows the add-on switch like the row did, and the palette still finds the page.
-
-- **The chat sidebar's More block holds History and the Cockpit, and both can be pinned.** Group Chat, Browser, Playbooks and Add-Ons stood there a second time, opening as windows over the app, while the same four are rows of the Home column that open as pages; two doors with two behaviours for one place. The block now holds the two surfaces you reach for between conversations rather than during one - History searches what was said, the Cockpit shows the work that came out of it - and either can be pinned into the column, where it stands under Agents. The four others are reached where they live. Group Chat carries speech bubbles instead of a second pair of shoulders, so it is told apart from Teams at a glance.
-
-- **A More row with nothing behind it is not drawn.** Pin every row out of the block, in either column, and More stayed - a chevron that opened an empty space and said nothing about anything. Both sidebars now leave the row out while it hides nothing, and bring it back the moment something is unpinned or a plugin, a widget or the DevKit gives it something to hold again.
-
-- **Playbooks has no menu entry any more.** A playbook is recorded and replayed where the browsing happens - from the chat and inside the Browser surface - so a row in the Tools list was a door to a workbench nobody navigates to. The row is gone. Nothing else is: the page answers at /playbooks, the palette finds it, the add-on switch is where it was, and run_playbook and list_playbooks work from any chat.
+- **Skales can now check whether a tool server it set up actually answers.** A reader wrote in: asked to wire up an MCP server, the agent pointed it at a folder it had guessed, checked neither the folder nor the connection, and reported the task complete and the connection working. The audit found the uncomfortable half of that - of every tool Skales has, not one could test an MCP connection. The handshake existed, behind the Test button in Settings, and nothing the agent could reach ran it. So no model, however good, could have checked; it could only guess or refuse. There are two tools now: one lists the servers this machine has and what each is doing, the other opens a real connection and comes back with the tool names - and says "needs a sign-in" as something different from "broken", because those send you to fix different things.
-- **The sidebar row that opens Iris Orbit is called Voice.** The row read "Iris Orbit" in every language, a product name where every other row says what it does; a person looking for the place to talk to Skales did not find the word. The row is now Voice, in all twelve languages, and the window it opens keeps its name.
+### Fixed
-- **Organization is opened from the chat, not from a menu.** A team of your own agents working one job is something you put to work from the conversation, and its row in the Tools list and in the chat column's More block was two doors to a place the chat already reaches. Both rows are gone; typing /organization opens it, the command stands in the slash menu with the others, and the palette still finds the page. An add-on that is switched off is not opened blind: the command names the switch instead. The Add-Ons card says where the door is now rather than marking the surface as parked.
+- **A placeholder cannot reach the screen by itself any more.** The sentence that names a licence change was one of two places where a line asked for a piece of text and never handed over the value it leaves a gap for, so twelve languages would have printed the gap instead - and the other was the FTP result in Lio, which put the file count in front of a sentence that already had room for it inside. Both say what they mean now, and the check that holds the twelve language files against each other reads the code as well: a line that asks for a sentence with a gap in it has to fill the gap.
-- **The line under the wordmark at start changes.** The splash screen carried one fixed line. It now draws one of twelve at every start, from a list the phone and the computer share byte for byte, and the fixed line stays as the fallback when the list cannot be read.
+- **The licence notice reaches the page it was written for.** An update that comes under the new agreement announces itself once and then waits four hours for the next check - and the page carrying the notice is reached by FOLLOWING that announcement, which is a navigation. By the time it drew, the announcement was gone, and the only thing left that still knew was the main process, which was not being asked about the licence at all. The notice was therefore missing on the one route anybody actually takes to it, leaving the Download button standing where the notice should have stood in front of it. The main process now answers that question too, the download's own progress no longer drops the answer, and the sentence names the version instead of printing a placeholder at the reader in twelve languages.
+- **A link preview stops printing `&amp;` at you.** GitHub escapes its own description twice, so "Android & iOS" arrives as `Android &amp;amp; iOS`, and one pass of decoding left the second escape standing in the middle of the sentence. It is decoded to the end now, in both the chat and the coding window, without ever turning text a page escaped ON PURPOSE into markup - and four entities the reader had never heard of (`&#39;`, `&lt;`, `&gt;`, `&nbsp;`) are read too.
-- **The licence names its licensor at the end, not in its first lines.** The agreement opened with a full postal address, telephone number and e-mail address in its fourth line, before it had said a single word about the software. The address, the telephone number and the e-mail address now stand together in a section of their own at the end, section 12, where a reader who needs to write to us finds them; the first lines say who the licensor is and point there. The English and the German text changed in the same way, and nothing else in the agreement moved.
-- **The licence screen says "du" in every language it speaks.** Skales addresses one reader, informally, and English and German did - but the licence screen changed to the formal voice the moment it was read in Spanish, French, Croatian, Turkish, Russian or Chinese, so the one screen everybody has to read was the one screen that spoke like a form letter. All twelve languages now use the same voice there: pronouns, verbs and imperatives together, not half a sentence at a time. The computer and the phone also say each of those sentences with the same words again - the line explaining that the binding text is in English had drifted into two versions, formal on one machine and informal on the other. The agreement itself is untouched and keeps its legal register, which is a different thing from how a screen talks to you.
+- **The weather line names the place that was found, not the words you typed.** The lookup has been forgiving for a while - it tolerates a misspelt country and a town the gazetteer has never heard of, and works its way outwards through the rest of the field until something matches - but the greeting printed the field back at you, so somebody who had typed "Aregentina" read their own typo beside a temperature that belonged to a place the app had correctly identified as being in Argentina. The line now says which place the temperature is for, and says the same thing on the second load as on the first. What it does when the field is EMPTY is untouched: no place is named at all, exactly as before. What it does when the field names a place nobody can find is untouched too, and deliberately - there the words you typed are the only honest thing to show. The assistant's own weather answer and the phone always named the found place; this was the one surface that did not.
+- **The first participant in a Group Chat is readable in light mode.** Participants are told apart by colour, and the first one uses your accent - but its name tag on the opening screen, its letter badge in the setup and its whole message card were filled with the accent AND lettered in the accent, which is not low contrast, it is a solid block with the writing invisible inside it. The four participants below it, which draw from fixed blue, orange, purple and pink, were always right. The same pairing sat on the card that carries the closing verdict, and that card belongs to no participant, so it was unreadable in EVERY round rather than only in rounds that had a first participant. All four places now tint the fill and keep the accent for the writing, which is how the blue and orange participants beside them have always worked. Dark mode never had the fault and is untouched, line for line.
-- **A chat that is working keeps saying so, all the way to the last step.** The conversation decided whether anything had changed by looking at a handful of facts about the last stored message - how many there were, whose it was, how long it was, when it arrived. Everything a running turn knows that is not in that sample moved without the screen noticing: which step it is on, which tool is open, what its helpers are doing, what the browser is looking at. Measured over one ordinary turn with tools, seven polls out of thirteen carried a change nobody looked at - which is how an answer from an earlier turn could end up standing over a picture you had just re-sent, with no thinking panel under it. The run now carries its own version number, and the chat redraws when that number moves and at no other time; it also notices when something else has replaced what is on screen and puts its own view back. The working indicator hangs on the run as well: it used to leave at the first token and never return, so a turn with eleven more steps to go looked finished for all of them. It now steps aside while an answer is being typed out and comes back for every further step, and a thinking panel you opened stays open across the whole turn. The Skales Code window asks the same question of the same field, so the two windows cannot disagree about the same piece of work.
+- **The mini chat window stops putting its header under the system's own buttons.** Every other window in the app leaves a strip clear at the top for the buttons the operating system draws over the page - the close/minimise/zoom lights on macOS, the caption buttons on Windows - and the compact window was the single one excluded from it, so on macOS the lights sat on the agent avatar and on Windows the caption buttons sat on the microphone, screen, + and expand controls. Two different corners, one missing reservation. It reserves the same strip as the full window now, from the same measurement, so the two cannot drift apart; on Linux, where the window keeps a real title bar and nothing is painted over the page, the strip is nothing and costs nothing. In the same place: the invisible band you drag the little window by was laid OVER the header rather than above it, and a drag band takes the click - so the top half of every button in that header quietly did nothing. The band IS the reserved strip now, and it stops short of the system buttons on either side, because a right-to-left Windows install puts them on the other one.
-- **Skales Desktop now runs on a newer generation of its web engine.** The
- ground the app stands on was a whole generation behind, which meant every
- release carried weaknesses that are no longer repaired at the source. Moving
- up puts the app back on a foundation that is still maintained and still gets
- security fixes, and it changes how pages and requests are handed around
- inside the app. What you see and what you click stays exactly where it was:
- every surface, from Chat and Code to Studio, Flow, Iris and Settings, was
- re-checked against the new foundation before this release.
+- **The figures under an answer are readable on a phone again.** Hovering the model line showed a box that ended mid-word, because it was drawn inside the message bubble and the bubble clips what leaves it. The breakdown and the bar's own figures now come from the one tooltip the app has had all along, which hangs off the document rather than inside the bubble and is placed against the line it explains: nothing is cut off at any window width, and it clears itself when the transcript scrolls instead of sticking after a tap. The numbers and the sentence a screen reader gets are unchanged.
-- **A picture the agent goes back to now says it is the same picture.** A run
- that takes a screenshot and later puts it back on screen drew it twice, with
- nothing to tell the two apart, so a long run read as if the agent had looked
- again when it had only shown you what it already had. The second one keeps
- its picture, because taking it away to save a few lines would hide that the
- agent went back to it, and carries a small "Shown again" chip beside the one
- that says it came from a tool. Two real looks at a screen that did not move
- are still two looks, and are never labelled that way, however identical they
- are.
+- **An answer that writes an HTML tag keeps the words around it.** A tag in an answer was not shown and not spelled out either: it was dropped, and with it every character it covered - a sentence lost the piece the tag sat in, and a tag on a line of its own took every line under it off the screen until the model wrote a blank line or closed it. On a long answer that is most of the answer, arriving one visible line at a time. The tag is now shown as the characters the model typed, so nothing leaves the screen. Code blocks, comparisons like `a < b` and plain links are untouched.
-### Added
+- **A picture is read by the model you picked for the chat, even when the caption asks for work.** With the Advisor Strategy on, the picture turn was sorted by the caption alone - and a caption like "analyze this" or "find the error" reads as a task, so the planner answered instead of your chat model, which is exactly what that strategy was fixed not to do. A turn carrying a picture now goes to your chat model, whatever the caption says.
-- **Skales can now check whether a tool server it set up actually answers.** A reader wrote in: asked to wire up an MCP server, the agent pointed it at a folder it had guessed, checked neither the folder nor the connection, and reported the task complete and the connection working. The audit found the uncomfortable half of that - of every tool Skales has, not one could test an MCP connection. The handshake existed, behind the Test button in Settings, and nothing the agent could reach ran it. So no model, however good, could have checked; it could only guess or refuse. There are two tools now: one lists the servers this machine has and what each is doing, the other opens a real connection and comes back with the tool names - and says "needs a sign-in" as something different from "broken", because those send you to fix different things.
+- **A picture no longer costs a capable model its tools.** Whether a picture turn keeps its tools was decided in the window, against the model shown there, and then held for the whole answer - while the model that actually answers is settled later and, with the Advisor Strategy on, is never the same one. So a model that can see the picture and act on it was disarmed by a verdict about a different model. The question is now asked of the model that answers, at the moment it is known, on every surface that takes pictures.
-### Fixed
+- **Joining through the Business page is now recognised on the Skales Plus page.** Both routes write `skalesplus_waitlist_email` into the settings file, but the Skales Plus screen only ever read the browser's local copy - so somebody who joined through Business was shown an empty form and asked for the same address a second time. It reads the same authority the Business page reads, keeps the local copy as the fallback when the settings file cannot be reached, and draws nothing decisive until it knows, rather than offering a join button and taking it away a tick later.
- **Attaching a picture no longer takes the tools away from an Ollama Cloud model.** A reader reported that the thinking and tool cards vanished from a chat the moment an image was in it. Half of what they saw is how it should be: a model that reads the picture itself answers in one go, so there is no vision step to show and, on a question that needs no thinking, no reasoning to fold open - ask something harder and the fold is back and stays. The other half was real. A model served from ollama.com through your local daemon - anything carrying a `:cloud` tag - is frontier-size and can call tools, but Skales judged it like a small local captioner whenever the daemon had said nothing about it, which is the ordinary case on an older daemon or on a model list fetched before the Ollama key was set, and sent the picture turn with the whole toolset stripped. The answer came back and nothing could be done with it. The tag decides now: a cloud model keeps its tools, a model your own hardware runs keeps the careful rule, and what the daemon itself reports still outranks both. A picture sent to the Telegram bot was a second version of the same mistake - there the toolset came off for every model, whatever it could do - and it now asks the one question the chat window asks.
@@ -195,134 +85,247 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
- **A failing command is not tried five times in a row any more.** Screen and browser actions were exempt from the brake that stops an agent repeating itself, and for a good reason: looking at a screen twice gives two honest answers, because the screen moves. But the exemption covered the failures too, where nothing between two identical attempts can change the outcome - so a wall that could not be climbed was walked into five times, and a question that had already been answered was asked four times, before the run gave up on its own. Repeating something that worked is still free. Repeating something that failed, unchanged, is now stopped at the third try and the agent is told what failed and to try another way - and if it succeeds in between, it starts with a clean slate.
- **A screenshot Skales took is drawn as the tool's picture, and a second look at the same screen is still a second look.** The record that carries a tool's image to the model travels in a user-role message, because that is the only slot a provider wire has for a picture. The half that keeps a reader from mistaking it for something the person typed was working: no blue bubble, no internal fence text on screen, one icon under the picture, and an exported transcript with only the person's own words in it. The other half was not. Where each picture came from was assembled and then dropped on the way to the session file, so a stored image had no provenance at all; and a picture was identified by the path it happened to lie at, which meant the same capture shown again under a copied name counted as a second photograph. Both are fixed, and identity now follows the bytes, with the tool that produced the record deciding whether it claims to have made the picture or only to show it again - two screenshots of a screen that did not move are still two captures.
+
- **A Claude model can use a tool and carry on.** With Anthropic picked as the provider, the first step of a run went through and the next one came back as a refusal from the provider the moment a tool had actually run - so anything that needed a tool ended on an error instead of an answer, and starting several helpers at once looked like only the first one worked, because every helper after it inherited the finished work of the one before. Claude is now told what was called and what came back, everywhere Skales talks to it, so a run continues where it left off, several results of one step arrive together instead of one at a time, and a picture a tool produced is still seen. A call whose arguments arrived damaged is handed over as what arrived rather than as nothing, so the model can correct itself instead of repeating the same empty call, and a step that is refused says in the diagnostics that it was the step returning results - the one line that tells an outage apart from a model that stopped mid-answer.
+
- **The products keep their names in every language.** A Skales Visual was a "Skales görseli" in Turkish, a "Skales vizual" in Croatian, "Visuales de Skales" in Spanish and a transliteration in Japanese and Chinese; Swarm was an "Enjambre" and a "Sürü", Discover was "Explorar" and "Keşfet", Wrapped was an "Özet", and Teams, Workflow, Playground, Playbooks, Studio and Spaces were spelled out in kana, hangul, hanzi and Cyrillic. A name is not a word: a reader who sees something else cannot find the thing again in the sidebar, cannot search for it, and cannot ask about it by name. All nineteen coined names now read the same letters in all twelve languages, on a tab and in the middle of a sentence alike - and the words that really are words, Settings, Tasks, History, Memory, Browser, Chat, Dashboard and Autopilot among them, stay in the reader's own language, where they belong. Grammar is left alone: Croatian still says "u Discoveru", Turkish "Skales Local'i", Korean "Skales Visuals-ro", and German still writes "Skales-IQ-Daten" the way German writes a compound.
+
- **The English text was the source of it.** The /image and /video command lines asked for "a 1:1 Skales visual" in lower case, and eleven translators were told by that, correctly, that it was an ordinary noun - so eleven of them translated it. English now writes the name too, and the gate that holds the list reads the English file as well as the other eleven, so the next name can no longer be lost by being written small.
+
- **Skales Visuals now works the way it reads, and says so when it cannot.** Asking for a graphic in a chat could take five minutes and come back with "The model returned nothing", or with three words that named nothing at all, and Skales could not afterwards tell you whether anything had been made. Three things were true at once and all three are fixed. The model that designs was never the one you chose: it came from the code builder's own saved configuration, a screen whose field says "writes the code" and which nothing in the app connected to your graphics. Designing now follows the studio row under Settings > AI Providers > Per-Mode Model Overrides - a control that was already there and did nothing - and with the row untouched it follows your active provider, or your own machine when you set Skales Local as your main model. The prompt enhancer, the captions, the storyboard, the scene animator, the template prompt writer and a visual made with /image or /video all follow the same choice; the code builder keeps its three models for building software and decides nothing outside it.
+
- **A design that comes back empty now says why, and tries the one thing that helps.** A model that thinks before it writes can spend everything it is allowed to produce inside that thinking and hand back nothing, and running the same request again spends the same minutes the same way. The design is now asked again with more room to write, and if that changes nothing, once more with the thinking switched off - the same rescue the chat has had for months, now on the same wire. When nothing still comes back, the answer says the budget went into the thinking, roughly how much of it, and which kind of model gets past it, instead of saying the model returned nothing.
+
- **A visual that fails leaves a record.** A failed design wrote nothing anywhere, so asking Skales afterwards what went wrong was answered out of an empty log - and the emptiness was read as proof that nothing had happened, while the picture was sitting in the gallery. A failed attempt is now written down like a successful one, where the diagnostics you can read back already look, and it is never counted as a design you made.
+
- **A long design is no longer cut off by the clock it was measured against.** The chat's request for a graphic and the design itself were given exactly the same amount of time, so the request always gave up a moment before the design could explain itself, and what reached you was "fetch failed". The chat now asks for a graphic the same way Studio and Flow do - watching it as it is written rather than waiting in silence - and it is given more time than the design itself has, so the sentence the design had ready is the one you get. A failure that really is a connection fault now names what failed instead of hiding it.
+
- **A provider outage reported mid-answer is no longer read as an empty answer.** Some providers report that the model behind them is down by saying so inside an otherwise healthy response. Skales read past it, ended the turn with nothing, and said the model had returned nothing - so an outage at the other end looked like a fault on your machine. Both the chat and the design path now read it and say what the provider said.
+
- **Studio and Flow render video again.** Before encoding, Skales proves that the composition really paints the frame it is asked for: it seeks to eight frames, out of order, visits three of them twice and compares what it got. The comparison was made byte for byte, and a browser does not draw type byte for byte - measured on a finished eight-second film, 18 of 2,073,600 pixels came back one unit apart on one colour channel, from the way the letters of one headline were smoothed the second time. Invisible on screen, invisible in the file, and enough to refuse the render with "frame 239 changed when sought twice". The two captures are now compared as pictures: a frame that MOVED is still refused, and a frame the rasteriser rounded differently is not. A refusal also says what it saw - how many pixels differed, and by how much - instead of naming only the frame.
+
- **Reporting an answer now works in every language, not only in the ones written with Latin letters.** The report was measured in characters and Skales accepts it by size, so a flagged answer in Japanese, Chinese, Korean, Russian - or simply a German one with enough umlauts - was two to three times too large, was turned away, and was then retried five times with exactly the same content while the confirmation said it was on its way. Every report is now trimmed to a size that actually gets through, and it says in the text when it had to be shortened. An answer whose cut fell in the middle of an emoji was rejected outright for the same reason, on every language; the cut can no longer land there. The ordinary bug report travelled the same road and had grown too large in the same way - a long one now arrives instead of vanishing, and the crash details that used to be sent in a field nothing stored are written into the report itself.
+
- **A report that cannot be delivered says so, instead of promising it is on its way.** Skales could not tell "you are offline" from "Skales refused this report", so a report that would never be accepted was announced as waiting and quietly given up on days later. A refusal is now said plainly and at once, a report that really is waiting still goes out at the next start, and one that has run out of attempts is written into the log the Diagnostics card in Settings shows, so the promise and the outcome can no longer disagree. The report is also put into the waiting list BEFORE it is sent rather than after the send fails, so closing the window mid-send no longer loses it, and every report carries an identity of its own so a slow answer cannot turn one report into two.
+
- **A report about an answer names the machine that wrote it.** The provider was read from whatever was selected at the moment of reporting rather than from the answer, so switching provider and then flagging an older answer accused the wrong one. The provider is now written onto the answer when it is produced and read back from there. The Report entry also disappears from the messages it never belonged on: a goal checkpoint, a budget card and a turn that failed are Skales speaking as itself, not generated content. And the entry is called the same thing here as on the phone - "Report answer" - in all twelve languages, with the field marked optional the way the phone marks it.
+
- **Every published copy of the user guide, the licence summaries and the third-party notices now says what the shipped ones say.** The guide published on the docs site and in the repository that carries the releases still named the licence Skales has left behind - under the current version number, so it read as a statement about this release. The published list of third-party components named one component fewer than the package actually ships. And two documents in this repository still named the previous version of the agreement while the newer one was what installed. All of them are drawn from the texts the app itself carries, and a check now holds the published copies against those texts, so the next correction cannot travel to one place and stay behind in another.
- **A package that lost its third-party notices can no longer be built at all.** The attribution other people's licences require of us was recorded when it was there and simply absent when it was not. It is now required exactly as the licence and the brand notice are, so a build that dropped it stops rather than shipping a copy that honours our terms and nobody else's.
+
- **Approving something from your phone "for this session" now lasts the session on the computer too.** The phone has sent that choice along for a while; the computer read every approval as a single one and asked again at the next step. It now honours what you chose, and a phone that does not say still gets the one-time approval it always did.
- **The chat can hand you an animation as a video file now.** A visual you asked for in the chat - a poster, a title card, a moving 16:9 scene - could be drawn in the answer and saved to the gallery, but asking for it "as an MP4" got you a description instead of a file: only Studio Classic could record one. The chat now runs the same recording as Studio, on the same machine, with no video provider and no key: the finished film appears in the answer with a player and a download, in your Workspace and in the Studio gallery. Two renders in a row are two files. If the machine is missing what the recording needs, it says which of the two pieces is missing instead of failing quietly.
+
- **Making videos no longer waits for an add-on it does not need.** The storyboard pipeline runs on this machine, costs nothing and needs no account - and it was hidden behind the same switch as the paid video services, so a machine that could make a film said it could not. The local pipeline is now offered wherever it can actually run; the switch stays in front of the services that bill you. What the assistant says it can do and what it can actually reach are decided by one and the same rule, so the two cannot disagree.
+
- **A small local model reaches your projects and the design tool without a detour.** With tools loaded on demand, a model on this machine was expected to ask for a second helping before it could read a project or draw a poster - and a small model does not ask, it says the thing is not possible. Those few now ship with every turn on a local engine. For a cloud model nothing changes: the turn stays as light as it was.
+
- **A chat bubble from yesterday says "yesterday" in your own language.** The timestamp under a message was translated everywhere except the one word that matters most in a long conversation, so a German, Japanese or Turkish transcript read "Yesterday 09:47" between otherwise translated lines. All twelve languages now say it in their own words, and today, yesterday and older are still told apart by the calendar day rather than by hours elapsed.
+
- **The "not enough memory" message says which memory it means.** Turning down a local model because the machine is too small named a number that reads as free memory and is in fact the machine's total, which invited closing a few apps and trying again - which cannot help. It now says so in all twelve languages, and where the local engine reported a figure for the graphics card, the message names that too. It is a detail in a sentence, never a new condition: no turn is refused for it.
- **A scheduled job you told not to interrupt you still runs in the open.** Every saved job whose "pause before executing and wait for your confirmation" box is unticked - the default for a new job - had quietly been turned into the kind of run the assistant uses for its own housekeeping: no entry in the list of what is working, no way to stop it while it worked, security gates switched off, and a one-line answer with the rule that delivers a reminder to your notification channel removed. Unticking that box now means only what it says: the run starts without asking first. It is listed, it is stoppable, every gate that guards an action still applies, and a job that says "remind me" reaches you again.
+
- **A scanned PDF is saved to your Workspace again instead of arriving empty.** A PDF without a text layer was being read as a document that simply had nothing in it, so the file was not rescued to your Workspace, the note about OCR disappeared, and the model was handed a stack of page headings with nothing under them. A scan is recognised as a scan again, with the same explanation and the same rescue as before, while a PDF that does have text keeps the page headings that let you jump to page 21.
+
- **A checklist for your next request no longer inherits the last one's leftovers.** Ask for something new in the same conversation and the plan card showed five steps when the assistant had sent two, with the old task's unfinished steps still blocking completion. A list that shares no step with the plan on file is recognised as a new piece of work and replaces it, with what the old plan still had open written into its history. An update to the task actually running is untouched: it still carries the steps it did not re-send.
+
- **A saved browser flow runs to its end again.** A step this replayer has no primitive for - a hover, a select, an extract - stopped the whole flow, so a saved playbook with one of those in the middle never reached the steps after it. Such a step is now named as skipped and the flow carries on, and an ordinary failed step is counted rather than ending the run. A navigation that fails still stops it, because everything written after it was written for a page that never came up.
+
- **A long agent run that ran into a limit teaches the agent something again.** A run that hit its step, token or cost limit, or stalled, is reported as unfinished rather than completed - and the lesson the agent writes into its own memory afterwards was only written for completed runs. Exactly the long runs with the most to learn from were learning nothing. They now leave a lesson too, and it says the run stopped short.
+
- **Iris takes your spoken "yes" again - and your typed one.** When she asks whether she may delete or run something, she waits for an answer, and the rule that the answer must come from a person and not from her own voice in the room was being measured from the moment the microphone opened rather than from the moment somebody spoke. Hands-free opens the microphone the instant her sentence ends, so every yes was judged to be an echo, she said so, opened the ear again, and refused the next one exactly the same way - with the buttons on screen as the only way out. The rule now starts where speech starts: wait a moment and answer, and it lands. A yes you type is never treated as an echo at all, because a keyboard is not a microphone. Her own voice coming back through the speakers is still not consent.
+
- **A dictated sentence stops asking to be sent again.** Stopping a recording on purpose puts the words in the input field to be corrected - but the recording behind them stayed on the "still to be sent" shelf for ever, so a repeat/discard pair sat on screen from then on, and "repeat" sent the same sentence a second time with the same tools. Dictation now ends that recording's business, and the words you send from the field are still the same one-and-only turn.
+
- **Iris no longer calls finished work unproven.** She speaks about a result only against proof, and the strongest proof this product keeps - a sealed project version, meaning the files on disk really changed - was never being read, so a Flow project she had just rebuilt was still announced with "nothing has confirmed this". Sealed revisions and the acceptance criteria a run actually checked now reach that sentence. A turn that changed nothing is still named as such, in those words.
+
- **Flow starts on a local engine whose address was typed without `/v1`.** The readiness check in front of Flow built its own address instead of using the rule every other request uses, so an engine that was answering Chat perfectly was declared silent and Flow refused to open. Both spellings of the address now work, and an engine that really is not answering is still reported as not answering.
+
- **The Flow start page stops reading from disk while nobody is looking at it.** Its "what is still being built" shelf refreshed on a fixed timer even when the window was minimised or behind another app, and each refresh pulled the whole cockpit - goals, scheduled jobs, agent tasks, team runs and every session - to find its own two rows. It now reads only the making shelves, only while the window is in front of you, and waits longer when there is nothing on it.
+
- **The installed app can hold a conversation again, on macOS and on Linux.** The packages the built-in server needs were being placed in a second folder beside the app instead of beside the server itself, and one of the two ways the server loads a package cannot look there. Nothing said so: the app started, every screen opened, and the first message answered with an error. They now ship where the server actually looks, on every platform and in both build recipes, and a check before each release refuses to build if that ever moves again.
+
- **A reply that never came is no longer handed over as an answer.** A model can end a turn with nothing at all - no words, no action, no error - and Skales used to pass that on as an empty but successful reply: no sentence on screen, nothing written into the conversation, and nothing to try again. Now it says in your own language what happened, keeps the turn in the conversation with the reason attached, offers the retry, and reports the turn as not delivered - so a scheduled job, an agent or a connected channel no longer files silence as work done. A long run that one of its own limits ended now closes the same way: if the closing summary comes back empty too, you get the completion line and the named reason in your own language with the retry beside it, instead of a line of untranslated English.
+
- **A conversation that stops working now says so by itself, instead of looking busy until you press Stop.** A turn that fails on its way in - a session bound to an agent that was deleted, a piece of machinery that will not load - used to end quietly: the record closed, nothing was said, the conversation kept the mark that says it is working, and a paired phone kept its banner up until somebody stopped a run that had been over for minutes. The end is now announced wherever the start was, the record names what ended it rather than reporting the run as completed, and a run that dies is surfaced the way an interrupted run already is. A tool that merely fails is untouched: that is a result the model answers, and the turn goes on.
+
- **A coding session now reads the file it just listed.** In a conversation bound to a folder, a file named without a full path was looked for in Skales' own workspace while the folder listing and the shell were in your project - so the model listed a file, was told it does not exist, and listed it again. A plain name now means the folder this conversation works in, for reading, writing, moving, searching and listing alike. A full path still points exactly where it is written, and what you allowed under file access decides as before: a path that leaves it is refused with the reason.
+
- **A project folder your phone may not open now gets a sentence, not the word "policy".** The file tree and the file viewer of a coding session answered a refusal with a single word that means nothing to a person. They now give the same explanation this machine's own Code window gives for the same refusal, in the language this machine is set to, including where to widen the rule.
+
- **A job you hand from the phone to your computer shows up in the list of what is working.** Handing one delegated piece of work across created a run and not a task, and the card that counts sub-agents only counted tasks - so it said nothing was working while a model was working and being paid for. The handover now appears with what it was asked to do and how it ended, and stopping the sub-agents really stops it instead of reporting a stop that changed nothing.
+
- **A tool written into a plugin means the same thing wherever the plugin was built.** What a tool does, what it is for and what arguments it takes are read from the package's declared shape rather than guessed at afterwards, so a plugin built on the phone and run here describes itself identically on both. Plugins already installed load unchanged; a tool that declares nothing about what it does is still treated as the safest of the three.
- **Your phone is now a working seat at a coding session on your computer, not a window onto one.** Every tool step arrives as it happens - what started, what it was about, how it ended and whether its effect is proven - instead of a pause followed by a paragraph. From the phone you can read the change list and one file's parts, keep or undo single parts, stage and unstage, preview a commit and make it, and take back one recorded change; a part or a commit whose file moved on since you looked is refused with the reason rather than applied to something you never saw. You can type into a process that stopped to ask a question, set the session's mode, model and reasoning effort, watch and stop the sub-agents running here, hand one delegated job over to this machine, and save a file you edited - which is refused, with your draft kept, if the file changed while you were reading it. Everything goes through the same doors and the same approval cards as the panel here, so nothing became possible from the phone that was not possible at the machine.
+
- **A background command that asks a question still ends by itself when nobody is there to answer it.** Making processes answerable from the phone gave every background shell an open input line, including the ones no person can reach - so a `git commit` without a message or an installer waiting for a password would have waited forever instead of finishing. The input line is now only held open when somebody is actually sitting at that command's own coding session on a paired phone; a phone that merely happens to be online elsewhere changes nothing, and every other command gets the same immediate end-of-input it always got. A line sent to a shell that was started without that channel is refused by name rather than disappearing.
+
- **Opening a desktop session on the phone no longer re-sends the whole conversation.** A session with screenshots was costing megabytes on every check, which is why the phone had to check slowly. It now asks for what is new and fetches a picture only when it draws one. A phone from an older release keeps receiving exactly what it did before.
+
- **A conversation bound to a project stays on that piece of work when it moves to the computer.** The binding was travelling from the phone and being discarded here, so the same chat did the same job unbound on the other machine. Two mode changes that cross on the way over now settle by which one you made last, instead of by which one happened to arrive last.
+
- **Reading a picture no longer needs the same key typed a second time.** The picture reader had its own key field, so an install with a Google or OpenRouter key already saved still answered every image with "not configured". It now uses the key you already gave that provider, and when no reader is chosen at all the provider you are working with stands in, but only when its model can genuinely see. A provider you chose by hand keeps its precedence, a model that cannot read pictures is never pressed into the job, and with no key anywhere the message names the settings page that fixes it.
+
- **A picture your computer's tools took now arrives on the phone as that tool's picture.** Opening a desktop session from the phone used to hand over those rounds as ordinary lines you had typed, so a screenshot showed up as your own message, complete with the internal warning text, and was exported under "You". The computer now sends the same origin the desktop transcript keeps. The "your computer is working" banner also carries the instant its run began, so a run that has already ended can no longer come back as working on the next reconnect.
+
- **The thirteenth version of a Flow project no longer costs you the first one.** Older versions used to be dropped from the history and deleted from disk without a word, so a state you might want back was simply gone. They now move into an archive: the recent shelf still shows twelve, the older ones are marked as archived, and every one of them can still be opened, previewed, branched and restored. A seal that moves something off the shelf says which version it moved.
+
- **An action still waiting for your approval is never reported as done.** A write, an edit or a commit stopped at the approval card used to be summarised as finished, so the answer named a file that was not on disk, and at the channel door nothing came back that could have shown you the card at all. The turn now hands the stopped action to whoever asked, says in your own language that it is waiting and was not carried out, and a goal cannot close while one of its own actions is still a question. A model that asks for the same action again gets one card, not a third copy of the same question.
+
- **A tool result now reaches the model exactly as the tool wrote it.** The live clock was appended to the last message of the turn, which during a tool run is the tool result itself, so Skales' own sentence landed inside the fence that marks tool output as untrusted data and left that fence unbalanced - the file body a model was meant to quote ended in a removal note and a time announcement. The clock now arrives on a turn of its own, behind the result, so a read file, a fetched page or a searched folder is handed on byte for byte.
+
- **A tool you write into a plugin now really runs, from the chat, from its own page and from a schedule.** A plugin's own tools used to be checked and stored and never registered, so calling one by name answered that no such tool exists. Each carries a name of its own, an argument schema, a version and what it does before it runs, and it reaches exactly what its plugin was granted: a plugin installed without network stays without network however its tool file is written. The plugin test calls a read-only tool for real before it reports that it works, and loads the others far enough to see that there is something to run, instead of reading the file's text.
+
- **Custom skills now run in a bounded host under rights they declare.** Skill code used to load into the app itself and was handed every setting and every decrypted secret, so an endless loop could take the app with it and a small helper could read any file on the machine. A skill now runs beside the app with a stop that works, a time limit that ends a runaway, no memory of its last run, and a contract that names which files, which addresses and which secrets it may use. Skills written before the contract keep working: their rights are read from their own code once, shown to you as such, and anything that ought to be your decision stays switched off until you grant it. Refusals are named on the screen instead of failing somewhere deeper.
+
- **A skill is now one package that says what arrived with it.** Every file is listed with its size and its own fingerprint, together with the version, where it came from and when. An import that could not fetch everything says so instead of installing quietly, and the agent can read a skill's reference document or script on demand rather than never seeing the files its instructions point at.
+
- **Templates now say which surface they drive, what they need from you and what you should end up with.** Using one starts its own run, so a template used a second time never arrives carrying the answers from the first. A run reports success only when the thing it promised is actually there; one that merely opened a surface is reported as unfinished rather than as done.
+
- **Custom widgets and plugins stop being two answers to the same question.** Every widget is carried over as a plugin page with its title, its id and every one of its saved entries intact, and nothing is granted by the move. The original stays on disk as the way back, a second migration changes nothing, and the existing entry points keep working where they are.
+
- **The Cockpit says how old what it shows is, and why it stopped updating.** A panel that loses its connection used to keep its last reading and present it exactly like a current one. It still keeps the reading, and now names the source, how long ago it was confirmed and what went wrong, in all twelve product languages.
+
- **Spoken work is only called finished once something has confirmed it, and interrupting now stops the work as well as the words.** A completion read aloud used to be the model's own sentence, so the better the voice sounded the more convincing a wrong answer became. A turn that changed nothing, a send whose outcome is unknown or a half-met checklist is now said as exactly that. How the voice performs - which engine, which language, how long until the first sound - is measured and shown separately from whether the task succeeded, and speaking over Skales stops the run she was speaking about instead of only silencing her.
+
- **Work you started on one making surface is now findable from the others.** A render still encoding, a cut still exporting and a project that stopped halfway through a turn used to live only inside the panel that started them, so closing it was the same as losing them. They now appear together with everything else that is running, on the phone as well as here, each saying how far it is and continuing behind the right door. A quick-start tile for a surface you switched off is no longer offered.
+
- **A local model that is downloading, loading or silent now says which of the three it is.** How far the weights have arrived, whether the engine is holding them, and whether it actually answers are three separate lines with three separate errors, and a background download of a second model no longer makes the one you are using look unready. Studio and Flow start a project only once the engine has really answered, and when it has not they name the reason instead of failing in the middle of the first turn.
+
- **Provider continuation remembers effects before an external call starts.** A durable, session-owned journal prevents replay after interruption, concurrent attempts, or a provider change, and recovered runs receive the recorded outcomes. Unconfirmed effects remain explicit, while a new user run stays independent. Effect identity covers the complete arguments, including changes in the middle of long requests.
+
- **A page Skales read for you stays the page it read.** When a long page has to be worked through in parts, every part comes from the one copy that was fetched - the address it really ended up at, what kind of page it was and exactly how far the reading got - rather than from a fresh fetch that might now answer differently. Asking for the next part costs no second visit to the site, and on a small model the reading advances only over text that was genuinely shown.
+
- **Research tools preserve complete source bodies until the shared spill boundary.** HTTP, raw page and extracted-text results now carry requested and final URLs, status, content type, full length, partial state and the complete body. The short human preview remains bounded, while the model can page the retained spill without losing a source's middle or ending.
+
- **A screenshot a tool took is shown as that tool's picture, not as something you wrote.** The pictures a tool hands the model travel on the same message slot a person's attachment does, so a screen capture appeared as your own blue bubble, complete with the internal warning text, could be edited as if you had typed it, and was exported under "You". It now carries the run, tool call and picture identity from the tool to the transcript, the export and the version history. The same picture shown again is recognised as the one picture, a genuinely new capture stays its own, and the screen label carries one monitor symbol instead of two.
+
- **The thinking and tool panel stays on screen for the whole turn.** It used to disappear once anything had been written to the conversation - so from the second step of a tool-using turn onwards, and for the entire run after you edited a message with a picture in it, Skales looked finished while it was still working. The panel now follows the run itself: it appears with the first activity, survives every tool step, stays open across a reload, and steps aside the moment the real answer arrives. A model that delivers no reasoning shows its working status instead of an empty thinking box.
+
- **Attaching a picture no longer takes away the model's tools.** Image turns were sent without any tools at all, which is right for a small local model that can only caption and wrong for every model that can both see and act - including resending the picture you just edited. Tools now follow what the model can actually do, read from the local runtime's own capability report or the provider's contract.
+
- **An edited picture message says so when the picture is gone.** If the stored copy and the thumbnail have both disappeared, the edit stops with the reason instead of quietly sending the text alone and answering a question about a picture nobody could see. Your message stays where it is.
+
- **Read-only modes now weigh a browser or desktop click like a file write.** Plan, Ask and the composer's read-only dial refuse every acting browser and computer tool, including the click that was classified as harmless, while reading a page or a screen stays free. A click aimed at a control that deletes, buys, cancels or pays asks first in every safety mode, naming the button it is about, and that answer covers only that one click.
+
- **A sub-agent asked for the same model now runs on the model this conversation is using.** The per-chat model pick, a custom agent's pin and the effort dial travel down to synchronous children and to fan-out jobs, instead of falling back to the global default with a different price and a different privacy decision. Each child's card names the model it actually ran on, and "fast" only means a runtime on this machine, judged by its address rather than by a missing API key.
+
- **Team runs started from the chat now belong to the run that asked for them.** They carry the session, the run id, Stop and the bill: stopping the turn stops the team, and every model call the leader and the members make lands on this conversation's spend, with an unpriced call reported as unknown rather than as zero. A run's saved state records which conversation it belongs to, so a reload finds the team run that is still going.
+
- **Switching the Organization off no longer looks like losing your teams.** Delegation inside a turn and jobs that already use a team keep working, and a chat with teams on disk says the add-on is switched off and where the switch is, instead of answering that there are none.
+
- **The card that starts helper agents, a fan-out or a team now says how the work will run.** Before you approve it: who is working, where they work (a scratch folder, the folder this session is bound to, or their own separate copy of it), what they can see and use, and where the result will land. The terminal and the preview each have one name everywhere, in all twelve product languages.
+
- **A project can now be the folder a conversation actually works in.** Give a project a folder or repository and the chat bound to it works there: relative paths resolve in it, the shell runs in it, and a trigger or schedule created from that chat runs there too. The binding belongs to the conversation, so it comes back after a reload while a second conversation keeps its own project, and an armed Code folder still wins. Saying "work in project X" from chat is one visible switch, and "none" goes back to the plain workspace.
+
- **Standing jobs can be created from chat as one complete object, including folder and mailbox triggers.** "When a new invoice lands in this folder, book it" now subscribes to the real event instead of saving a prompt or pointing at a page, and the same object carries the instruction, the folder the run works in, which agent runs it, where the result goes and whether it may act unattended. The watcher is armed the moment the job is created rather than at the next start, a job can be fired once as a test, its runs and their results are readable from the conversation, and removing the job stops its watcher.
+
- **Documents are now read with their structure, and never cut without saying so.** Word files bring their headings, their tables as tables, and the footnotes, endnotes, comments, headers and footers that used to be skipped entirely; embedded pictures and diagrams are named and can be handed to a vision-capable model. Spreadsheets keep cell ranges and formulas instead of collapsing to tab-separated values, and PDFs can be asked for any page range rather than only the first twenty. Every read reports how much of the document it returned and where to continue, so a report, contract or table is never summarised from an unmarked excerpt.
+
- **A working checklist may now change during a run, and no step can leave it in silence.** The agent can insert a step, split one in two, re-word it or skip it, each time saying why, and the checklist card shows the revision and the reason. A step the agent simply stops mentioning stays on the list as an open decision, long plans keep all of their steps instead of being cut at thirty, and a run cannot report itself finished while the plan it showed you still has open or unexplained steps. Asking for something else starts a fresh plan and records what the previous one still had open.
+
- **The Browser Control approval switches now decide every browser and computer action.** Auto-approve navigation and the approvals for logins, form submissions, purchases and downloads are read at the moment an action is decided, not only stored in Settings. A click on an already-filled order form is weighed as a purchase, not as a keystroke, and the four classes about identity, money, someone else's system and your disk keep asking even when prompts are switched off wholesale. Consent is remembered per action class and per site, so an unchanged authorized action stops re-asking and another site never inherits the answer.
+
- **A saved browser flow stops at the first step that fails or cannot be replayed.** Every step now carries what must be true before it runs, how it proves it worked, and what happens when it does not. A flow made only of checks can no longer report a clean success without having checked anything, a click after a failed navigation does not run, and the steps that were never reached are listed as not run instead of quietly disappearing.
+
- **A team run only says "completed" when its work is actually done.** A model error, an exhausted step budget, a stop, a blocked branch and an answer that only appeared after every tool failed are each their own end state, on the page, in the chat reply, in the exported document and in the history. The leader's summary is told what is missing and may no longer present an unfinished run as a finished deliverable, and a failed run no longer counts as a finished one anywhere else in the app.
+
- **Team tasks run as the dependency graph they declare.** A task never starts before the task it needs, circular and dangling dependencies are named and cut instead of stalling or being ignored, independent branches still run side by side, and a failed prerequisite blocks everything downstream with the failing task named. Plans larger than a handful of tasks keep every task.
+
- **A team run picks up where a restart left it.** The run remembers the steps it proved and the conversation it belongs to, so a reload resumes at the first unfinished task rather than paying for finished work twice or losing the run entirely.
+
- **AIPointer answers picture questions on the model you chose, and every call it makes is on the meter.** A multimodal model now receives the screenshot itself and keeps its Skales actions on that turn, instead of being replaced by the separate Vision Provider and stripped of them; the Vision Provider still answers when the chosen model cannot see. The vision call is counted against the pointer's daily ceiling like any other, so asking with a picture no longer walks past the limit. Desktop element search also works with a local, keyless vision server, the same way every other surface already accepted one.
+
- **A recorded desktop procedure now replays the keys it recorded, follows its windows, and can be asked for values.** Modifier combinations such as Cmd+S are actually held, punctuation and the number row are pressed rather than dropped, and a key this machine cannot press stops the run before the first click instead of being skipped silently. Clicks are stored relative to the window they happened in, so a moved window no longer turns the procedure into clicks on the desktop. Any recorded text can be marked "ask each time" and becomes a question at run time, F10 stops a running replay, and the result says which step it reached and why it stopped.
+
- **A desktop click now belongs to the screenshot it was read off.** Every screenshot returns the display it shows, its position, its scaling and a revision; a click or scroll carries that revision back and is mapped to the right point even on a Retina panel or a second monitor. If a monitor was attached, removed or moved in between, or the look is too old, the action is refused with the reason instead of landing somewhere else.
+
- **Computer Use can read the screen instead of only looking at it.** Controls, windows and their frames now come from the accessibility layer with roles and names, so a button can be clicked by name, an app can be brought to the front and confirmed, and an action can state what should be true afterwards. An effect that could not be checked is reported as unverified, which is neither success nor failure, and a missing macOS Accessibility permission is named on screen instead of appearing as an empty desktop.
+
- **A spoken conversation now sees the pictures it is about.** Iris and Call Mode hand the runner the whole conversation, including the tool results and the pictures attached to them, so "what was in that screenshot?" reaches a model that can actually see it. A camera frame, a dropped file or something pasted into the window goes into the same turn as any other attachment, stored with the session so it is still there after a reload.
+
- **Choosing the live channel now opens a real one.** Where the provider has a bidirectional audio API, Iris opens it: the provider's own turn detection, interruption that stops the voice mid-sentence, the same tools the chat has with the same approval prompts, and what it costs written into the session's own spend. Providers without such an API are not offered it, and when the channel cannot be opened or drops for good, the window says so and carries on listening the usual way instead of quietly changing what the mode means.
+
- **A long spoken instruction survives a failed transcription.** The recording is kept under the conversation before anything is sent, so a provider error, a missing key, a reload or a restart leaves it recoverable with one button instead of asking for the whole paragraph again. Transient failures retry themselves with a growing delay; a missing key waits for the key. Retrying never sends the same words twice and never runs the same tool twice.
+
- **Iris no longer takes her own voice for your approval.** A spoken "yes" only counts when it came from a microphone that opened after she stopped talking, so a permission question coming back through the speakers cannot approve a tool. Saying no is still always heard, the buttons always work, and a refused answer says why instead of being swallowed.
+
- **The orbit shows the whole result, not the first part of it.** A long report, document or answer is paged inside the frame with a place to read on and a line saying how much is left, instead of being cut at a fixed length with nothing to indicate it. Rows and columns are drawn as a table, and every row of a list arrives. Closing the window and coming back rebuilds what was on screen from the conversation itself.
+
- **A video you cut keeps everything you set, all the way to the file.** Where each clip sits, its own sound, the crossfades between them, how loud the voice and the music are, the music ducking under the voice, the subtitles, the shape of the frame and whether it leaves as MP4 or WebM are all part of the saved project now, instead of being rebuilt from guesses on the way out. Export a second time and you get a second file, of exactly the cut you were looking at.
+
- **Code follows the rules your project already has, at the place they apply.** A repository often carries rule files at several levels - one at the top, another in a subfolder, another for one kind of file - and they were being applied loosely. Before anything is written, Skales now works out which of them apply to that exact path and in which order, whether the change comes from the main run, a helper agent or your phone. And a preview can be checked for more than how it looks: what the page logged, what it asked the network for, what failed to load, and a real picture of it. An app that needs its own address, a sign-in or its own stored data can be moved into the session's shared browser rather than an isolated one.
+
- **A request that is really a job is recognised in all twelve languages.** Asking for several steps of work in one sentence turns into a Goal at the same point whether you wrote it in English, German, Russian or Japanese, short or long. A plan you are only quoting or explicitly ruling out stays an ordinary message, something you want repeated stays a schedule rather than becoming a Goal, and whenever a message is promoted you are told in the conversation and the Goal keeps saying so afterwards.
+
- **You can edit a file yourself in Code, and the tree keeps up.** The review panel saves whole text files, with the keyboard where you expect it, and refuses to overwrite a file that changed on disk while you were typing - it tells you instead of silently winning. The file tree comes back with the project's own selection and folders still open, reloads only the folder that actually changed, shows a folder it could not read with a Retry rather than an empty space, and can be walked entirely from the keyboard.
+
- **A finished goal is checked against every requirement you actually stated.** A long brief no longer loses its must-haves after the seventh one. Where the proof is a file, the file's content is fingerprinted and checked again later - so a proof that was true an hour ago and has since been edited is caught - while something already done out in the world stays done and is never done a second time to re-prove it.
+
- **Lio tells an interrupted build apart from a finished one.** A build that was stopped, failed or only got halfway is recorded as that, worked out from the steps that really finished and the files that really have something in them. Picking it up again starts at the first unfinished step, and deploying or asking for changes stays locked until the build has actually proved itself.
+
- **A small change to a Lio project no longer costs a full rebuild.** Asking to change a word, a colour, a label or a bit of spacing reads just the part of the file it concerns and makes one precise edit, while a real structural change still goes through the full plan-and-review path. Each project keeps a record of what it was before and after, how many rounds it took, which files were read and changed, how long it took, and anything that changed that you did not ask for.
+
- **Falling back to another provider no longer does your work twice.** When a provider fails halfway through a turn and another takes over, Skales now knows which actions had actually gone through - a mail sent, a file written, something created out in the world - and does not repeat them. That memory survives a reload, so a resumed run does not send the message again either. A separate run you start yourself is untouched by it.
+
- **A picture sent from the phone is a real file on the computer again, and the answer says where.** An image attached on the phone lands in the same folder a picture dropped into the desktop composer lands in, so asking to edit, convert or upload "this image" finds something to open. The sentence naming the file comes from the same place the desktop's does, so a model that cannot see the picture is told that plainly instead of guessing at it.
+
- **Working from the phone now runs the same way as working at the computer.** A request from a paired phone enters exactly the machinery a desktop request enters, gets a confirmation that it was really accepted, and is then followed as the one true run - through queueing, an approval card, Stop, losing connection and restarting. The transcript comes back whole rather than cut at a fixed length, and asking for several agents no longer quietly puts the work on a second, different track.
+
- **The browser you can see and the browser Skales drives are the same browser.** The panel and the automatic tools now agree at every moment on which page is open and which page an instruction refers to. Switching between them does not leave an instruction pointing at a page that is gone, reloading rejoins the page that is actually there, the panel follows along when Skales navigates, and two runs that both want the browser are told which one has it instead of silently pulling the page out from under each other.
+
- **Helper agents get the same senses and the same off switch as the run that asked for them.** A helper started inside a turn inherits that conversation, pictures included, and Stop reaches it and its tools rather than only the parent. The planning and reviewing roles in Code can read files and run checks through the same guarded set the writing role uses. And browser actions keep the run that opened the page as its owner: a competing request from a chat, a coding session, a playbook or a plugin is visibly refused instead of quietly taking the page over.
+
- **Memory recall only brings back something that actually matches.** Its own bookkeeping - scan state, lists of deletions, a malformed record - could crowd out the whole recall window; it no longer counts as a memory. Recency now only decides the order of things that are relevant, instead of pushing an unrelated fresh fact into your next answer. And short names and words stay findable in all twelve product languages, both when they are stored and when they are looked up.
+
- **Forgetting a conversation forgets what was learned from it.** Deleting a session also removes the memory built out of it, including any day summary it fed, and leaves a marker so a later background pass cannot quietly reconstruct the record you deleted.
+
- **Pushing a day plan to the calendar cannot duplicate it.** If the push is retried - a hiccup, a lost answer, a resumed run - the events that already exist are reused rather than created again, an event that was made but never confirmed is found, a block you edited is updated, and a half-finished push carries on from where it stopped.
+
- **Code keeps what you had half-done with the session it belongs to.** The text you were typing, the files you attached, the skills you picked and how hard the model was set to think all follow the session you opened, across switching sessions and across a reload. A first send that fails keeps the turn instead of eating it, attaching more than fits is named rather than silently dropped, and a team run receives the same files, pictures and skills a single agent would have received.
+
- **A command you started in Code keeps running while you are not looking at it.** Hiding the panel, reloading, or going off to another project no longer kills it; reopening the panel shows you everything it wrote in the meantime. Closing that tab still ends it, and quitting Skales still ends what is left.
+
- **A Flow video is proved before it is encoded.** The renderer waits until the motion engine is really ready, steps forwards and backwards through the difficult frames, and refuses a frame that is missing, broken or different every time it is asked for - naming the frame. A file that was encoded and a file that passed its checks on length, size, frame rate and colour are now reported as two different things, so a render that failed its checks can no longer be handed to you as finished.
+
- **A Flow render or review survives closing the project.** The work is recorded against the project, the exact version of the source it came from and the stage it had reached, so reopening picks it back up - rechecking if the source changed since, continuing a correction that was already in progress, and only accepting that a correction worked after seeing that it changed the very thing it was meant to change.
+
- **A proposed day plan is checked before it replaces the one you have.** Blocks have to be well formed, sit at real times in your own time zone, last longer than nothing on the same day, fall inside your working hours and not collide with your calendar, your fixed blocks or each other. A malformed answer or a provider failure stays visible as what it is and leaves the plan you already had alone.
+
- **Stop stops the helpers too, at the same moment.** Pressing Stop reaches every helper agent's model request and every tool it is running, so nothing new starts afterwards and no helper quietly begins another paid round. Helpers running side by side also spend from the one conversation budget: calls already in flight may finish, and then they all stop before the next model call once their combined spend reaches the ceiling you set.
+
- **What you review in Code is what gets committed.** A commit preview shows the exact state you staged and refuses to go ahead if that state changed while you were reading it; an action on part of a change is refused if the change has moved on; and undoing a checkpoint keeps later edits and shows you the conflict instead of overwriting them. Undoing a whole turn is worked out in full first, so repeated or multi-file edits either all come back or nothing is touched.
+
- **Work done by a helper agent in its own copy of the project is never lost in the handover.** What it changed is measured from where its copy started, so work it already committed counts too. The parent confirms the handover arrived before cleaning anything up - and if the handover is refused, the helper's result fails loudly and its work is kept for you to recover rather than swept away.
+
- **Chat shows the local context window without a trip to Settings.** The meter and its detail read from the local runtime that is actually serving, keep what they say tied to the endpoint and model you selected, and stop an older reading from reinstating a stale window after you switch models.
+
- **Stopping the voice keeps it stopped while a fallback is still loading.** A reply that arrives late can no longer start the old voice over the next thing you play, and a failure to speak keeps its real reason instead of a generic one.
+
- **A second local turn no longer refuses to use the model already loaded.** The check before a run now compares the memory the model needs with the memory the machine has, so the memory the loaded model itself occupies no longer looks like a reason to say no.
+
- **A closed connection no longer takes the local server down with it.** Someone closing a page while a local answer is still being written is recognised as exactly that and survived. Every other write failure still goes down the normal error path, and the diagnostics keep the real reason.
+
- **Local hardware diagnostics keep the whole device name.** The server's output is read across however many pieces it arrives in, so a detected graphics card and how much of the model it is carrying survive startup and repeated loads instead of arriving cut in half.
+
- **A local context window follows the model it belongs to.** The window a model is really being served with is carried for models that are idle as well as loaded, refreshed when the Local panel looks, and kept separate per endpoint - so the same model served from two places no longer shares one stale number.
+
- **A signed-in ChatGPT subscription is recognized consistently in Flow.**
Flow's landing check, project model choice and shared surface routing now use
the same provider readiness rule. Older valid sign-ins without a stored
provider switch are accepted, while a cloud endpoint address alone is no
longer mistaken for a usable credential.
+
- **Pictures attached in Chat and Code survive the session handoff.** The first
Code request, queued turns, follow-ups, app restarts, resumed runs, and paired
mobile sessions reopen the same session-private image bytes before building
the provider request. ChatGPT subscription models receive native image input,
including the user's per-model vision override.
+
- **An agent run is only called finished when its work is done.** A run knew
just one way to end badly - every tool failed - so everything else was filed
as completed: an action still waiting for your approval, a run that used up
@@ -334,6 +337,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
reads a run says the same thing - the agent's card in a team run shows it
stopped rather than done and is left out of the ranking, and a room is told
the action is waiting instead of being handed an answer that sounds complete.
+
- **One tool offered twice no longer breaks every tool-using turn.** The set of
tools a conversation carries is put together from the app's own, your
imported skills, your plugins and your connected servers, and some model
@@ -344,8404 +348,158 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
left out, and both sides are named so the duplicate can be renamed. One such
duplicate was already there and is gone - with imported Agent Skills switched
on, the tool that opens a skill's own files was being offered twice.
-- **The picture reader's key field no longer fills itself with a key you never typed there.** Letting the reader borrow the key you already gave a provider also put that key into the Browser Control card, so the field showed a secret nobody had entered there, and pressing Save laid a second copy of it down on disk and froze it. Rotating the key in AI Providers then left the reader quietly using the old one until every picture came back as a failed request. The card shows what is stored again, an empty field stays empty, and the borrowing happens at the moment a picture is read - so a new key is in effect on the next image, with nothing to re-save.
-- **Clicking a control by name now lands on that control, on a Retina screen too.** A button addressed by its name is found through the accessibility tree, which answers with a point of the desktop - but it was then read a second time against the last screenshot, so on a scaled display the click landed at half the distance from the top left corner, and if another app had come to the front in between it was refused with a sentence about a screenshot that was never involved. A point that came from the tree is now used as it is; a coordinate read off an image is still checked against that image as before.
-- **Looking at the screen stays allowed in Plan and Ask.** Four ways of reading the desktop - listing the displays, listing the open windows, reading what is on screen and checking whether something is true there - were refused in the read-only modes, while taking a screenshot went through. They only observe and are now treated that way, so a plan can be worked out by looking before anything is changed.
-- **A skill you installed before there were permissions keeps the doors it always used.** Giving skills a bounded space of their own closed four of them at the same time: a skill that fetches from the internet could not reach it even when its own permissions named the site; the data folder and the workspace were handed over and then refused on every path inside them; the settings it used to read came back empty; and a skills folder reached through a shortcut was not recognised as belonging to you. All four are open again and none of them wider than before - the internet only to the sites the skill's permissions name, the two folders only because they were handed over, and the settings with anything that looks like a secret removed, exactly as a command run from a shell gets them. A request carrying something other than text is refused by name instead of being sent empty, and what a site answers with is readable again.
-- **The same skill folder now gets the same package fingerprint here and on the phone.** The fingerprint exists so two copies can be compared without trusting where either came from, but the two platforms listed the files in different orders - and the order is part of the fingerprint, so any folder holding a SKILL.md or a README.md never matched itself across devices. Both sides list in one order now.
-- **Building a release no longer leaves the terminal and the global hotkey broken behind it.** Both are pieces that have to be rebuilt for the machine they will run on, and that rebuild was being skipped at the end of a release - so afterwards each failed with an unreadable error that looked like a fault in Skales itself. It happens again now, on every kind of release build, and there is only one place it can be forgotten from.
-
-### Security
-
-- **A team of agents started from an incognito chat now stays off the record too.** Ask several agents a question inside an incognito conversation and the run itself behaved: it stayed in memory and refused to file a summary in your documents. Every agent underneath it did not. Each one filed the first hundred characters of your question under recent conversations, pushed up the counter of how long you and Skales have worked together, wrote a line into the activity log that feeds Wrapped, posted to the Discover feed that a conversation had happened, and fed the knowledge graph. A named agent also wrote your question, word for word, into its own run record on disk, and taught itself a lesson from the conversation afterwards. None of that happens off the record now. There is one place in Skales that answers the question "is this off the record", it answers it from the conversation itself and not only from the switch, and everything that runs a turn asks it.
-
-- **The Discover feed and saved images ask the conversation, not the screen.** Both asked only whether the incognito button looked pressed. That button is screen state and a reopened window clears it, while the conversation underneath is still the off-record one - the same gap that made the badge report possible. A picture sent in an incognito chat could still land, at full resolution, in your workspace uploads folder, where the file tools can read it long after the chat is gone.
-
-- **An incognito chat can read its own notes again.** Closing the memory leak took two blocks with it that were never memory: the notes Skales keeps for itself during a conversation, and the summary of a goal it had just finished. Both are read from the conversation, and an incognito conversation lives only in memory and dies with the app, so neither could outlive anything. Worse than the loss: the tool that writes those notes was still switched on, so Skales kept writing notes to itself that it could never read back. Both come back now, for incognito conversations only.
-
-- **Incognito now covers what Skales already knew about you, not just what it would write down.** The badge promised no memory and delivered half of it: the conversation got no file and the memory tools were switched off, so nothing could be looked up - but everything Skales already knew was in the prompt before you typed a word. Your profile, your saved facts, the topic index, the answers you gave on the personalisation card and a line naming what your last conversations were about all travelled into an off-record chat, which is why one could still open by knowing who you were. An incognito conversation now gets the same treatment a remote provider gets under Privacy Mode: language and clock, nothing else. It is told plainly that it is off the record, so it says that instead of inventing a reason for not knowing you, and the receipt under an answer no longer lists memories the reply never saw. A chat you leave and come back to is covered as well - it used to wear the badge on screen and run as an ordinary chat underneath, because coming back forgot which kind it was while the conversation itself had not.
-
-- **An incognito conversation no longer leaves anything behind in your memory.** Every finished turn was recorded: the counter of how long you and Skales have worked together went up, the first hundred characters of your message were filed under recent conversations, and a fact was extracted from what you said - none of which knew that incognito existed. So an off-record chat quietly wrote the very lines that were read back to you in the next one, which is how it could remember a session that was supposed to have left no trace. Nothing from an incognito chat is written now, and the refusal sits at the one place all four channels pass through - the chat, the desktop buddy, WhatsApp and Telegram - rather than at each of them, where it had already been forgotten once.
-
-- **Skales stops inventing a memory about your computer.** Any message containing the word "server", "environment" or "where are you" filed a permanent memory saying it had detected it was running on localhost, on Windows. It had detected nothing: no address, no provider and no operating system were ever read, the sentence was fixed text, and the match was loose enough that "observer", "serverless" and "environment variables" all triggered it - each repeat counting as another confirmation. It was then read back as a fact, and on a Mac or a Linux machine it was simply wrong. The entry is gone, and what it pretended to answer Skales can check properly and on request. The neighbouring rule had the same fault: saying "my name is not important" stored your name as "not", and a fact you asked to have remembered was stored with every occurrence of "that", "this" and "save" cut out of the middle of it. Both now read the sentence you actually wrote.
-
-- **The record of what you recently discussed is being kept again.** Every short note about a finished conversation was silently thrown away before it reached the store, on every machine where the setting for keeping identifiers is off - which is how Skales ships. The check that refuses to store anything shaped like a password was reading the whole record including Skales' own internal conversation number, and that number contains a long run of digits, so the check took our own plumbing for your personal data and discarded the note with it. Only these short notes carried such a number, which is why the memory page kept filling up while the recent-conversation half of it quietly stopped. The check now reads what was said and ignores the envelope it arrived in.
-
-- **The licence now names the cases it used to leave to good manners, and the download carries nothing that could hand its own source back.** It says that Skales is to be run as it was published, that it is not to be built into another program, wrapped, or pointed at from somebody else's tool, that access to it is not to be sold on or bundled, and that nobody may present themselves as a partner of Skales or claim that Skales stands behind their product. Because the text itself changed, every installation is asked once more, with the new points in the summary in all twelve languages and the binding English text one click away - and a further way in which a package can give up the readable original of its finished code is now refused before a release can be built at all.
-- **The Apple Silicon download is now built on an Apple Silicon machine.** The release was cut on one computer, and the half of it meant for newer Macs came off a machine of the other kind - which meant that half travelled without the protection its Intel twin had, and nothing in the build said so. The recipe now hands that half to a machine of the right kind, checks the finished package there, brings it back and checks it again here before anything is published. A release that skips that step says out loud what the Apple Silicon package is missing instead of leaving it to be found later.
-- **A copy that removes the licence no longer looks untouched.** The record every package keeps of its own files covered the program and its texts, but not the two pages that say who the copy belongs to and what may be done with it - so deleting them or rewriting them left a build that still reported itself as the one that shipped. They are part of that record now: a package that leaves the factory without them is not built at all, and an installation from which they were removed or edited afterwards says so, in the diagnose panel and wherever the build talks to us. It still starts and it still works; it just stops claiming to be an original.
-- **Skales now ships under its own licence, and the download says so in one place.** The old terms were borrowed, carried a date on which they would have turned into something else, and left the name and the logo in nobody's terms at all - so somebody who took the published build, put their own name on it and offered it as their own product had nothing in writing to trip over. Skales is proprietary software, licensed and not sold, and it stays that way. Private use, study and teaching stay free of charge, exactly as before; an organisation gets thirty days to try it out; company use, hosting it for other people, white-labelling and passing copies on need a yes in writing first. The licence and a short brand page now stand at the place people actually download from, they travel inside the package next to the third-party notices, and the same two lines go under every release. The open-source licences of the components Skales is built on are unchanged and still carried with it.
-- **A Skales build now identifies itself to Skales' own services.** Until now a
- request only said which release was calling, which made an original build and
- a copy of it indistinguishable from the outside. Every build carries its own
- identity from here on and presents it wherever it talks to us. Nothing about
- this is about you: it says which build, never who is using it, and an older
- Skales keeps working exactly as it does today.
-- **The downloaded app no longer carries a readable copy of the program it is built from.** Every package Skales has shipped so far came with the app's own uncompiled source alongside the finished program - not needed to run anything, and there for anyone who opened the download. It is gone, while everything the app genuinely reads from disk while running - all twelve languages, the built-in skills, the templates and the design style packs - is untouched. More of the desktop half now ships compiled rather than as plain text, which leaves less of the machinery behind the update and the local engine lying open in a downloaded copy.
-- **A Skales install can now tell you whether it is still the build that was published.** Every release writes down what its own packaged files are, and the app checks them once after it starts. When something has been swapped, that is said plainly - in the diagnostics you copy into a bug report, and in the sentence the assistant is given when you ask it why something failed - so a report from a modified copy is no longer mistaken for a report from the real one. A mismatch never blocks the app or hides a crash: it is a statement, not a lock, because a crash report you can read is worth more than a lock. On your own machine, and in a developer checkout, there is nothing new to see.
-- **A release build now refuses to ship a package that is not what it should be.** Building a release reads the finished app back and stops if the source crept in again, if the compiled pieces are missing or came with readable copies beside them, if the hardening switches in the binary are not what the build set, or if the record of the packaged files does not match them. Before this, every check of that kind read the source tree - which is not what anyone downloads, and is exactly why the source in the package went unnoticed for months.
-- **A copy of Skales that was not built here can now be told apart from one
- that was.** Every release already carried a stamp of its own; the package it
- ships in now also identifies itself as a Skales build in a way that survives
- being taken apart and put back together, and a check before each release
- refuses to build if that identification is no longer intact. None of it is
- visible while you use the app, nothing about you or your machine is part of
- it, and nothing is sent anywhere.
-
-## v12.9.27 - Hotfix
-
-### Added
-
-- **Image tools give the working model the actual picture.** Chat, Code and
- the shared agent loop attach screenshot and `show_image` results as image
- inputs, including the ChatGPT subscription adapter. Session-private copies
- let a resumed run reopen its pictures without resending every old screenshot.
- In Code, `show_image` now resolves relative files inside the bound project and
- refuses images outside it, matching the other file and image tools.
-- **The ChatGPT subscription follows the existing effort control.** Chat and
- Code send the selected supported reasoning level through the shared request
- builder, including session overrides and the model's available levels.
-
-- **Diagnose now says what actually left this machine.** Mails, messages,
- uploads to your server, a post to Discover and a finished video are written
- down at the moment they happen, with what the far side answered - successes
- and failures both. Until now a mail Skales believed it had sent and a mail
- that really went out looked exactly the same afterwards, and a bug report
- about "it said it sent it" had nothing in it to settle the question. A
- half-finished upload says so, and a render whose progress bar reached the end
- without producing a file counts as failed, because that is what it is. The
- record is local, the recipients are shortened, and it goes into the report
- you copy out.
-- **A working sub-agent says what it is doing, not just that it is working.**
- The chips under a run now carry the step the child is on, the tool it just
- reached for and how long it has been going, ticking on its own. A child on
- its ninth careful step, one looping on the same tool and one whose provider
- stopped answering used to wear the same spinner for as long as they lasted.
-
-- **Diagnose has a model check: every model you set up must place a real tool
- call.** Some models write their calls in a spelling this build cannot read.
- When that happens the machine text is hidden, as it should be, but the call
- is dropped - so the answer bubble looks finished and nothing actually
- happened. The check asks the models themselves rather than a list of known
- spellings, which is the only way to catch the next one. It runs one model at
- a time, shows each row the moment it is known, and the result copies out into
- a bug report.
-
-- **Settings > Appearance now has a switch for the animated background, and
- it starts off on Intel Macs.** The Skales X theme draws three soft gradients
- that drift behind the shell; on a Mac with an integrated Intel GPU that
- layer costs real frames, and a user watching their fan reported it. The
- switch turns the layer off entirely - it is not rendered at all, so it costs
- nothing - and takes effect at once, in every open window, without a reload.
- On an Intel Mac it starts off and the row says why; everywhere else it
- starts on, exactly as before. An x64 build translated onto Apple Silicon is
- not an Intel Mac and keeps the aurora. Whatever you set beats the default,
- survives a restart, and survives a cleared browser store. Switching it off
- also parks the ambient bloom behind the conversation, which is the other
- layer that never stops moving - but only when you switch it off yourself:
- the machine-picked default speaks for the drifting gradients alone and
- leaves every other theme exactly as it was.
-- **A PDF now opens inside Skales instead of sending you to another
- application.** A PDF the agent wrote, one you find in a file card, and a link
- to a PDF in an answer all open in the app's own panel, with the scrolling,
- zoom, search and print you know from any reader - drawn by the Skales window
- itself, so nothing is downloaded and nothing new is installed. There is one
- panel, so opening a second document replaces the first rather than piling up
- windows, and Escape, the X or a click beside it closes it. Download and Open
- Folder stay in the panel's title row, so a file can still be taken into your
- own reader. A PDF lying in a project folder is shown the same way in the Code
- window's preview pane, where it used to be handed over as a download.
-- **Skales Code can now run a task through three models instead of one:
- an architect, a coder and a reviewer, each on its own provider and model.**
- Press Roles in the Code composer, or type `/roles` with what you want done.
- The architect plans the change and names the files without writing code, the
- coder makes the change as an ordinary coding turn - the same tools, the same
- approvals, the same diff cards - and the reviewer reads back the difference
- that actually landed and says what is wrong with it. `/architect` and
- `/reviewer` ask one of them on its own. The three roles can sit on three
- different companies, which is the point: a blind spot is not reviewed by the
- model that has it. Set them up in Settings > Chat & Code > Skales Code roles;
- a role left empty runs on the model that session would have used anyway, and
- the row says which one that is. Each answer lands in the session log as its
- own folded block naming the role, the model and what that call cost, so it is
- still there after a restart, and all three prices count against the session's
- spending ceiling. Stop stops all of them. Nothing runs by itself: a session
- where you never press the key behaves exactly as it did before.
-- **A custom endpoint is no longer one model.** An OpenAI-compatible server
- usually serves several, and Fetch models on the endpoint's card has been
- asking it for a while - but only the card read the answer. Every model the
- endpoint returned now appears under that endpoint's own name in the chat
- model picker, in model search, in Flow and in the model field of an agent,
- with the model typed on the card leading the list. Several endpoints can be
- set up side by side as before, each with its own name, address, key and
- models. Typing a model id by hand still works, for a server that answers with
- one it does not list. Pointing an endpoint at a different address drops the
- previous server's models instead of offering ids the new one refuses - and a
- different address now includes a different port, so two local runtimes on one
- machine no longer share one model list. In the chat model search, a model of
- an extra endpoint is now captioned with the name you gave that endpoint
- instead of the internal slug.
-- **An agent on your own endpoint picks its model from a list.** The provider
- field has held every extra endpoint by name for a while; the model field
- beside it was a free-text box, so the one model that had to be typed from
- memory was the one the app had already fetched and cached. It is now the same
- catalogue the chat picker shows, with the free-text box still there for an
- endpoint that has never been asked.
-- **The Analyze report now says what a run cost, and what each single answer
- in it cost.** The price used to live only in the meter under the composer,
- which answers "what has today cost" but never "which answer cost it": a
- screenshot step that charged a quarter of a dollar disappeared into a
- session total. The report's header now carries the same sum the meter
- carries - the same adder, so the two can never print different totals - and
- every turn in the log names its own price beside its tokens. A provider that
- reported no price says so rather than showing $0, because a zero reads as an
- answer that was free. The copy-as-text report carries both figures too.
-- **The bar under an answer now names that answer's price when you hover it.**
- The bar that opens Analyze already broke a turn's tokens into where they
- went; it now adds the money on the same card, including what sub-agents and
- the behind-the-scenes calls charged on that step's behalf. Shown only when
- the provider named a price.
-- **The AIPointer hotkey picker can now learn the key by listening for it.**
- Press "Choose by pressing" beside the picker and hold the key you want:
- AIPointer takes the name from the key itself. Which modifier sits in a given
- position is a property of the keyboard, not of the operating system - on
- many non-Apple keyboards the key where right Cmd sits reports as right Alt -
- so picking a name from a list could leave you holding a key the app was
- never listening for, with nothing on screen to say so. Only modifiers are
- accepted, anything else is refused with a sentence, and the list stays for
- keyboards and for the Web UI, where nothing can listen.
-- **The row under the hotkey picker now names the key it last saw, not only
- whether the selected one arrived.** A status dot and one plain sentence say
- what is going on - "Last key seen: Right Alt (2s ago)", or, when that is not
- the key you selected, "You are pressing Right Alt, but Right Cmd / Win is
- selected". The way to fix it is the button beside the sentence, and while a
- freshly pressed key is taking effect the row says so instead of naming the
- old one. A hotkey that never fires and a hotkey you are simply not pressing
- used to look identical from the outside: nothing at all.
-- **You can now say how full the context window may get before Skales
- summarises a long run.** The threshold sits in Settings under Memory Mode,
- next to the other memory dials, and defaults to the same 75% it has always
- used - so nothing changes unless you move it. Set it to 0 and Skales never
- summarises mid-run, which is what a very large window wants and what a small
- local model must not have. The hint beside it names where the size of that
- window itself is decided.
-
-### Changed
-
-- **The ChatGPT subscription is its own provider card, above OpenRouter.** It
- used to be an accordion folded INSIDE the OpenAI card - one vendor's sign-in
- living inside another vendor's key card, although the two are different
- relationships with different credentials and different model catalogues. It
- is now an ordinary card called "OpenAI Codex (OAuth)", with a tile in the
- provider grid, a switch, a Test button, the timeout and context-window
- settings every other provider has, and the model dropdown with the Fetch
- button beside it. The card that asks for a key is now called "OpenAI API", so
- the two are told apart at a glance. Signing in is what its card shows where
- every other card shows a key field, and the sign-in is also offered during
- first-run setup, which only ever listed providers you can paste a key into.
-- **Skales Local now reports the context window it is actually serving to the
- rest of the app.** The Skales Local tab printed the real number while the
- chat meter, the budget and the compaction threshold read a deliberately
- small default - so one model on one machine showed 8K on one screen and
- 65,536 on another, and nobody could tell which one the conversation
- followed. The running server's own answer is now written where every one of
- those readers looks, the moment a model is loaded, and it is removed again
- when the server stops rather than left behind as a stale promise.
-- **A document the agent writes now says where it went.** The tool described a
- "Document panel" and named no way to it, so an answer could send somebody
- looking for a tab that does not exist. It now names the panel behind the
- document icon in the chat header, says it is not a page or a tab, and the
- same place is in the app's own map, so asking "where is my document" gets
- the real answer.
-- **Asked what it can do, Skales now answers about the model as well as about
- itself.** The reply used to describe the app and leave out the half people
- actually ask about: which model is answering, how big its window is and
- whether it reads pictures. Those three facts now ride in the same answer,
- resolved from the same sources the turn itself is run on.
-- **The per-model context override is findable.** Setting a context window for
- a model has been possible on every provider card for a while, folded away
- under a word that named none of it. The fold now says what is behind it, the
- fields say what they are for, and Settings search finds them under "context
- window", "context size" and "max tokens".
-- **The time zone follows you.** It was read from the browser once during
- setup and never again, so moving to another zone left the current time,
- reminders, quiet hours and new schedules on the old one while only the
- weather followed. Skales now re-reads it at every start and updates the
- stored zone when it changed. Settings > General has a new Time zone row:
- Automatic follows the browser, Manual pins a zone you pick from the full
- IANA list. Schedules you already created keep the zone they were created in.
-- **The location field says what it is for.** It is now called Weather
- location, and its hint says plainly that it does not set the time zone -
- which is what the Time zone row above it is for.
-
-### Fixed
-
-
-- **Skales Local explains why a model cannot load before a turn starts.** Chat
- now checks the selected local model's installation and free-memory
- requirements before sending, and reports the actionable reason instead of
- forwarding an opaque llama-server failure. Code sessions remain visible in
- Chat history, approval cards keep a neutral session-scoped third choice for
- every pending tool batch, and queued replies are consumed at the next agent
- step.
-
-- **Chat can remember approvals for this session, and Settings has one save.**
- The approval card offers the same session choice as Code for each named
- action. The fixed Settings save also saves Goal activation and call limits,
- and switching tabs keeps unsaved values. Changes to the action
- classification still pass through the existing permission checks.
-
-- **Markdown files in Skales Code open as documents, and their file links stay
- inside the workspace.** The file tree sent Markdown into a browser preview
- that served plain text; it now opens the existing rendered document panel.
- Links to another project document, PDF, image or HTML file use the same file
- opening path as the transcript instead of navigating the Code window away.
-- **Sub-agents have one progress card in Chat and Code.** Direct children now
- use the same trace as dispatched tasks, instead of a second hand-built card
- and a generic working-tool line. Hover or focus a child for its step budget,
- current tool, elapsed time, model and reported cost. The result keeps those
- details after the run and after reopening the conversation; a finished child
- does not appear twice while its parent continues working.
-- **The ChatGPT subscription can fetch the models your account actually has.**
- The catalogue requires a client version; leaving it out made a valid sign-in
- fail with a misleading address or firewall error. The request now identifies
- its client, so newly available models such as GPT-6 Astra arrive from your
- account's catalogue instead of waiting for a hardcoded list to change. Their
- reported context windows and supported effort levels stay with the model.
- A limit or outage during sign-in renewal keeps its real cause instead of
- becoming a request to sign in again.
-- **Ollama uses the same context window as the conversation it receives.** A
- long chat could be kept against the model's larger window while its image
- request quietly asked Ollama for only 32K. The request and the conversation
- now follow the same explicit setting, live model information or catalogue
- value. The output budget also counts only the current pictures, after older
- image turns have become text placeholders.
-- **Code Auto now covers the script runner and remote sessions keep their permission controls.**
- PowerShell and other multiline scripts use the same in-project Auto decision as
- shell commands, while dangerous or out-of-project scripts still ask. Turns sent
- from Mobile now persist their explicit Code and access modes before any work is
- queued or started, so Desktop enforces the same choice.
-- **A long Code run no longer turns its compressed context into a Goal of its
- own.** The forty-step handoff used the machine-written compression block as
- if you had typed it, and it ignored the Auto-trigger switch entirely. It now
- keeps your real instruction before compression, rejects internal context and
- image messages, and only hands a plain chat to Goal when Auto-trigger is on
- and that instruction clears the sensitivity you chose. Codework stays in its
- Code run instead of spawning empty-criteria goals that die forty steps later.
-- **Code and its shell can see the tools a Windows terminal can see.** Shell
- children now combine the app's inherited search path with the current
- machine and user PATH from Windows itself, without creating competing `Path`
- variables or changing which existing command wins. PowerShell 7 detection,
- Node, npm, npx, MCP servers and Codework all use that same path, so installed
- commands no longer fail with a bare `ENOENT` inside Skales.
-- **A picture returned by a tool stays available on the next turn.** Its private
- session reference now survives the Code handoff, a queued follow-up, Goal
- resume, ordinary chat-history reconstruction and the image-turn path. A
- screenshot the model could inspect when it arrived no longer turns into a
- text-only placeholder after a reload or when you ask about it again.
-
-- **A temporary ChatGPT refresh failure keeps its real cause.** Network errors,
- server outages and usage limits no longer inherit the original request's
- “sign in again” message. Refreshing credentials also leaves global model and
- effort preferences alone. Both plain chat and tool runs use this behavior.
-- **The isolated read-tool benchmark finishes even when a tool rejects.** Its
- timer is released on errors, and the source-tree measurement explicitly binds
- its own folder instead of depending on personal file-access settings.
-
-- **The ChatGPT subscription can finally be asked which models your plan has.**
- Every provider card has a "Fetch Available Models" button; on the
- subscription it was permanently greyed out, because the button asks whether
- an API key is filled in and a subscription has no API key at all - the
- credential is the sign-in. So the one provider whose catalogue nobody can
- know from memory, because a Plus account and a Pro account do not see the
- same models, was the one that could not be asked. The button works now, on
- the computer and on the phone, and it asks the Codex backend with the
- account header the way a real turn does.
-- **The sidebar no longer says "Not Connected" under a subscription that is
- answering.** Two separate reasons, both fixed: the name under the status was
- built by stripping "_oauth" off the id, which printed the name of a
- DIFFERENT provider, and the status itself was only re-read when you changed
- page - so signing in on the settings page left the red box standing until
- the window was reloaded. It now names OpenAI Codex and updates the moment the
- sign-in, the active provider or a provider switch changes.
-- **Speaking to a paired Skales Pocket no longer sits on "Understanding" while
- a slow speech provider takes its time.** Every leg of the speech-to-text
- cascade now has its own clock, and the recording moves on to the next one
- after 25 seconds instead of waiting forever - only your own endpoint had a
- clock before, and the two callers that have no browser behind them (the
- paired device and the video tool) had none at all. The whole cascade is also
- capped for a paired device, so the puck hears what went wrong rather than
- running out its own patience in silence.
-- **A dedicated speech provider now transcribes before a general router.**
- Running the chat on OpenRouter used to hand it your dictation too, with Groq
- - which answers Whisper in seconds - waiting behind it. Which model writes
- your answers and which one hears your voice are two different choices, and
- only one of them was ever made on purpose. A speech model you pinned on
- OpenRouter yourself still leads.
-
-- **The weather understands your place names in your own language.** The
- gazetteer only knows a town under the names of the language it is asked in,
- and it was always asked in English: "Wien" came back as Wien, Missouri, and
- "Graz, Steiermark, Österreich" as nothing at all. It is now asked in your
- language first, then in the language of your region, then in English, and
- the answers are merged - the same rule the phone has used all along. The
- widget and the greeting line take the language of the window, the weather
- tool the language you set in Settings.
-- **Reminders, schedules, calendar events, the planner and the greeting run in
- the zone you pinned.** The planner tool stamped a new reminder with the
- zone of the machine the server happens to run on, a recurring task carried
- no zone at all, a calendar event took the server's zone, and the greeting
- read the browser's clock - so a pinned zone or a server without one moved an
- absolute reminder by the difference. All of them now read the same zone the
- clock line in the prompt reads. A second walk through the setup, or a
- profile save, no longer overwrites a zone you pinned by hand.
-- **The setup pre-selects Fahrenheit for a US locale.** A default, not a
- question: the three regions that read Fahrenheit get it, everyone else keeps
- Celsius, and Settings still lets you change it.
-- **A run stopped after a few words of thinking leaves no empty bubble.** The
- transcript kept the message because reasoning existed, while the trace is
- only drawn from about twenty tokens on - the result was an avatar, a token
- bar and a time with nothing between them. One rule now answers both.
-- **A PDF the agent produced sits above the fold.** The View button of a file
- written by the PDF tools lay behind the collapsed Reasoning fold, unlike a
- file written by the edit tools; a produced file is the result of the turn
- and shows like one.
-- **After a stop, the late re-read of the session actually runs, and Enter
- cannot start a second run on the stopped one.** The re-read hung on a flag
- that was cleared a line later, so it always turned back; the composer was
- freed before the run record was, so an Enter in that window doubled the run.
-- **A custom endpoint that serves hundreds of models gets the same short
- picker group as every provider.** The search still lists all of them, and
- the model you are on always leads. The agent editor re-reads a stored model
- once the catalogue has arrived instead of showing an empty dropdown.
-- **A weather location written the way you say it now finds your town
- instead of nothing.** A field like "Los Nogales, Tafi Viejo,
- Argentina" was sent to the gazetteer in one piece, which answers nothing for
- it; the dashboard then fell back to browser location and, when the window
- had no permission for that, to Vienna - so a user in an Argentinian winter
- was shown 31 degrees. The field is now searched one segment at a time, and
- the remaining segments decide which of the answers is meant, forgiving
- accents and up to two typos in the province and the country. The chat's
- weather tool takes the same route.
-- **A location that cannot be found is named on the widget instead of being
- swapped for another city.** As long as a location is set, neither browser
- location nor the Vienna default can take its place: the widget says which
- field it could not resolve and offers the way to Settings, in all twelve
- languages, and the weather tool answers the same way in chat. The greeting
- line above the dashboard follows the same rule - it was the one that showed
- 31 degrees to a user in mid-winter. The old fallback chain stays for the one
- case it was written for - no location set at all.
-- **The browser no longer keeps running after the job that opened it is
- done.** A page opened for one turn used to stay up for as long as Skales did
- - a headless Chromium with a graphics process on several cores, nowhere on
- screen and nothing to click to end it. A browser now closes with the run
- that opened it, however that run ended, and any browser left sitting
- untouched closes itself after the session length set under Browser Control.
- Nothing is lost: cookies and logins live in the browser profile, so the next
- page opens still signed in. Headless browsing also no longer asks for
- graphics acceleration it cannot show, and the visible window you sign in by
- hand stays open while you type. Quitting Skales sweeps up any browser an
- earlier crash left behind. Browser Control's "Max session time" now says
- what it does, and the tools themselves know the browser does not outlive the
- run.
-- **A place search no longer bills your Google key at the most expensive
- rate.** Every `search_places` call asked Google for the rating, the review
- count, the opening hours, the price level, the phone number and the website,
- and in Places API (New) a single one of those fields prices the whole search
- in the Enterprise tier, whose free monthly allowance is a fifth of the one
- below it - so an agent that only needed a name and an address quietly spent
- the expensive quota on every lookup. A search now asks for name, address,
- coordinates, type and photo, tells the agent in its own answer which fields
- it left out, and takes `rich` when the rating really is wanted for a whole
- list. The new `get_place_details` tool fetches all of it for ONE place,
- which is billed per place instead of per result list, and the Settings hint
- next to the Places key says which tier a search stays in. Place cards are
- unchanged: they draw those fields, so they still ask for them.
-- **Editing a message right after Stop now actually sends it.** Pressing Stop
- and then correcting the question used to look like Save did nothing: the
- second and third attempt were refused in silence and the fourth arrived
- together with the ones before it, picture and prompt three times over, and
- the answer reasoned about the question you had already replaced. Stop now
- releases the chat the moment it is pressed, and a corrected message waits
- for the run it replaced to be gone before it goes out - once, in order,
- about the new text. If the old run will not end, you are told so instead of
- watching Save do nothing.
-- **Stopping a turn keeps what you already read.** The half-written answer and
- the reasoning behind it used to be wiped in the same instant the run ended,
- so a paragraph you were in the middle of reading disappeared and nothing of
- it came back on reload. What had arrived when you pressed Stop is now part
- of the conversation, trace included, and it is still there after a restart.
-- **The thinking and tools panel no longer disappears in the middle of an
- answer.** On a model that thinks out loud, the panel showed up with the
- first thought, vanished the moment the model reached for its first tool, and
- only returned once the whole answer was finished - which on a long run was
- most of the wait. It now stands from the first thought through every tool
- step to the last word, and it keeps whatever you had opened open.
-- **Switching an endpoint to a different address no longer keeps offering the
- old one's models.** The provider card added everything a previous fetch had
- cached to whatever the current endpoint serves, and because that list only
- ever grew, pressing Fetch models could not clear it - so models that answer
- 404 stayed on the menu. The card now reads the catalogue the way the chat
- picker always has: a list belongs to the address it came from, and a fresh
- fetch replaces it instead of adding to it.
-- **Forcing "this model can see images" on Skales Local no longer sends a
- picture into a model that cannot receive one.** The switch outranked what
- the running server reported, so the picture went out, the engine dropped it,
- and the model answered about something it never saw - or started hunting for
- the file with read_file. For our own local server the loaded model's actual
- modalities now decide, and when the switch is on for a model without vision
- the picture is set aside with that said plainly, to you and to the model.
-- **An MCP server no longer answers "HTTP 400" when several tool calls arrive
- at once.** Sub-agents work in parallel, and two of them reaching the same
- not-yet-connected server used to open two connections: the server issued two
- sessions, Skales kept the second, and calls carrying the first were refused
- with a bare 400 and no explanation. There is now one connection per server
- no matter how many callers ask for it at the same moment.
-- **A tool call no longer overtakes the connection handshake.** Skales counted
- a streamable-HTTP server as connected before the final step of the handshake
- had been delivered, so a strict server rejected the first call of a session.
- The handshake is completed first, and a call that is refused because the
- server no longer knows the session re-opens it once and is repeated, instead
- of failing in front of the user.
-- **The MCP server log now says what was asked, not just what came back.** A
- failing tool call can be reproduced by hand: the log carries the tool name
- and its arguments, shortened, with API keys and other credential fields
- masked.
-- **Ollama no longer resends every picture of the conversation on every
- turn.** A chat with several image turns used to put all of their images into
- each new request, so a model that answered five pictures well was handed
- fifteen and replied with nonsense. Only the pictures of the turn being
- answered travel; earlier turns say that an image was there.
-- **Every shell command works again on Windows.** `execute_command` and
- `run_script` failed with ENOENT for everything - `powershell.exe`, `cmd`,
- `node`, `git` - because child processes were handed a search path containing
- nothing but Skales' own fallback folder. Windows spells the variable `Path`,
- the filter copied it under that name and then wrote a second one called
- `PATH`, so the child ended up with two search paths and ran with the wrong
- one. The search path is now written under the one key the platform already
- uses, and it is the inherited system path with Skales' fallback appended -
- never a replacement for it. The same collision is closed in the MCP server
- spawn, the language-server spawn and install, the dependency check and
- browser control.
-- **A trigger key you picked for AIPointer now survives switching AIPointer
- off and on again.** Turning the master switch wrote only the on/off state,
- and that write replaced the whole AIPointer block: the trigger key, the hold
- duration, the appearance, the accent, the screenshot crop and the
- only-when-Skales-is-hidden gate were all gone the next time it was touched,
- so the hotkey silently fell back to right Cmd and the rectangle that hangs
- off holding it went with it. The switch now only changes the switch.
-- **The hold duration under AIPointer now actually governs how long you hold
- the trigger key.** The slider saved its value and nothing read it - every
- hold was two tenths of a second no matter where it stood.
-- **AIPointer no longer asks the Keychain whether you pressed its trigger
- key.** Deciding that took the full settings read: four provider keys
- decrypted through the system Keychain and the whole settings file re-parsed
- - once per keystroke, twice counting the release, and again for every mouse
- move at screen refresh rate. All of it ran on the same thread the global
- input hook waits on while it holds the OS event tap open, which is why the
- keyboard half could go quiet on a busy machine while mouse wiggle kept
- working. The hotkey and the wiggle switch are now read on their own, and the
- settings file is re-read only after it actually changes.
-- **AIPointer says when a trigger key never reaches it.** macOS can let an app
- watch the mouse while withholding keystrokes, which looks exactly like a
- broken hotkey: mouse wiggle summons the box, the key does nothing, and the
- app said nothing at all. The row under the hotkey picker now names the state
- - key arriving, waiting for a first press, keystrokes withheld while the
- mouse is seen, Accessibility off with a way straight to it, or the input
- hook not running - instead of leaving a dead control on screen.
-- **The drag rectangle can no longer get stuck armed.** If the box was stopped
- from opening at the last moment because Skales itself was in front, region
- mode was armed anyway for a box that was never shown, and nothing ever
- disarmed it - the next rectangle started in a state the user had not asked
- for.
-- **A room agent now gets at most five pictures per turn, the newest ones.** A
- Space agent was handed every image attached to every message in its window,
- so three messages of three pictures arrived as nine in a single request -
- more than any other surface in Skales can send, and more than a local Ollama
- model accepts at all. The ceiling is now the same one the chat composer has:
- the most recent five travel, the pictures left behind are marked as set
- aside in the line they belonged to, and the agent is told it did not see
- them instead of guessing. The task card says how many stayed behind, so
- nobody has to work it out from a bad answer.
-## v12.9.26 - Grip
-
-### Added
-
-- **A paired device can now have this computer turn a recording into text.** A
- phone transcribes on the phone, but a small paired device with a microphone
- and no room for a speech model previously had to reach this machine at its
- address on the local network - so it only worked at home. It now asks over the
- same connection everything else uses, gets the transcript back, and the
- recording never leaves that path. It runs through the same speech setup the
- microphone in Skales uses, so whichever provider you configured is the one
- that answers, and a recording that is too large to travel is refused with a
- reason instead of dropping the connection.
-- **The Studio render button now asks before it exports, and the export can go
- through the server.** Pressing Render opens a card in the panel itself: how
- many scenes, which aspect and size, how long, and what should happen to the
- subtitles - nothing is written before that card is answered. Scenes you have
- previewed or edited are still rendered from exactly what you saw; a
- storyboard nobody has previewed goes to the render pipeline in one request,
- which keeps running when you leave the page and is the same path the chat
- tools take. Whatever produced the file, it ends up in the same place: the
- download button, the gallery entry, the subtitle file, and the toast.
-- **Campaign, as a Flow template.** One brief, several finished pieces: a hook,
- a main clip and a call to action by default, each its own file named by its
- role, with one palette, one type pairing and the same closing frame across
- all of them, plus a small index page to review the set in one place. It rides
- the motion mode and the existing template picker - no new mode to learn.
-- **Skales asks before it exports a video - in every mode.** A render writes a
- finished file and takes minutes of the machine, so the card comes up every
- time and names what is about to come out: how many scenes, which aspect and
- size, how long, and what happens to the subtitles. Auto and Unrestricted do
- not lift it, and the card deliberately offers no "always allow this" button -
- the next film is a new question. A preview frame still runs without asking:
- it writes nothing anyone keeps.
-- **A rendered Studio video can carry its subtitles - as a file, burned in, or
- both.** The captions come from the storyboard's own scene text and timing, so
- nothing is transcribed and nothing is guessed: a scene without text still
- holds its place on the clock, so no later line arrives early. The `.srt` is
- written next to the MP4 and stays editable; burned in, the words survive an
- upload to a network that throws sidecar files away. Off is still the default,
- because burned-in pixels cannot be taken back out.
-- **Skales Code can see what is in its context, and shorten it.** The figure in
- the status bar now opens a breakdown of the turn: what the system prompt
- takes, what the tool definitions take and how many there are, what the
- conversation itself takes, how much of that is the newest tool results, the
- total against the model's own window, and what the session has cost so far.
- Every number is the one measured on the payload as it went out, so the meter,
- the breakdown and the hint cannot disagree about the same session. The meter
- itself now counts the whole turn rather than only the transcript, which is
- why it reads higher than before and why it now reads true.
-- **A shortened stretch of a coding session reads as a lid, not as a wall.**
- `/compact` folds the older half of a conversation away and leaves one card in
- its place saying how many turns went under it; opening the card shows the
- summary. It survives a reload, because the session file carries it.
-- **The run line offers to shorten when the window fills up.** Past four fifths
- of the model's context the line adds how full it is and that `/compact`
- shortens it. A suggestion only: nothing is ever folded away without being
- asked, because the conversation belongs to the person having it.
-- **An MCP server behind its own certificate authority is reachable.** A server
- on a company network or a self-signed one takes a `CA certificate` in its
- settings - the path to a .pem or .crt file, or the certificate pasted in - and
- that certificate applies to that server's requests and to nothing else on the
- machine. Verification is never switched off, and nothing else Skales talks to
- starts trusting the new root. A certificate that cannot be read says which of
- the three things is wrong - the path, the format or the date - when it is
- saved rather than on the next failed connection, and a connection that fails
- anyway now names the real reason instead of "fetch failed".
-
-- **Skales Code says what it is doing, step by step.** A coding run used to
- show one word - usually "deciding" - for as long as it took, which looks the
- same as a run that died. There is now one live line under the turn: which
- step is on, which tool on which file, and for how long. Parked on a question
- it says so; gone quiet longer than it should, it names what is hanging
- instead of counting up under a word that is no longer true. When the run
- stops, the same line becomes its receipt - steps, files, how long, and what
- the turn cost - read from the session itself, so reopening the window later
- shows the same one.
-- **A diff is no longer painted like a failure.** Removed lines and the minus
- counters beside them come from the diff palette rather than from the red this
- window keeps for a step that actually failed, everywhere they appear: on a
- tool line, on a card head, in the status bar and in the session list. Failed
- steps and script errors stay red, because those are errors. A long diff still
- arrives folded, but never blind: its first three changed lines stay on screen
- and the lid says how many are still behind it.
-- **Every step says how it ended, as a shape.** Tool steps carry the same tick,
- cross or warning sign the live line uses, in the same slot and at the same
- size, and the rows keep one height as their details fill in.
-- **A permission question says how far a yes reaches.** Each button on the card
- now carries one sentence of its own: this call only; every action of that
- kind until this session ends; or every one of them, in every session, until
- you take it back in Settings. An unattended session that stops to ask anyway
- says why, in the words of the check that stopped it.
-
-- **Devices: your other computers, under a name you can read.** The surface
- that used to be called Agent Swarm is back in the menu as Devices, under
- Tools. Devices are YOUR machines and Teams are other people, and that is the
- whole difference between the two entries. It lists the Skales instances on
- your network, takes a remote one by address, shows what was handed back and
- forth, and delegates a task from a form. Nothing about how it works changed:
- the /swarm command in the chat, the peer scan under Settings, Advanced and
- the switch on the Add-Ons page are the same ones, and the old /swarm address
- still answers - it forwards to the new one, so a bookmark keeps working.
-- **A long run says what it is doing, step by step.** The Skales Code window
- used to show the word "deciding" for minutes at a time, which is the same
- thing it shows when a run has died. Every step now records what it is - the
- tool, the file or command, how long it has been at it - and the status bar
- reads the newest one: "step 7 - write_file src/InputManager.ts - 12 s". A run
- waiting on your approval says that instead of claiming to work. After 45
- seconds of nothing at all (it used to be two minutes) the line names what went
- quiet, so you can tell a hung command from a slow model without stopping the
- run to find out. It survives a reload, because it is kept with the run and not
- in the window. Skales can also now tell you which script interpreters your
- machine actually has, rather than guessing and learning from the error.
-
-- **Fold the sidebar with Cmd/Ctrl+B, and it stays folded.** The column
- remembers whether you folded it, across a restart and in both readings of the
- sidebar. Before this, a window you had given the extra width back to came
- back full-width every single time.
-
-
-- **Skales opens the page instead of telling you where it is.** When the chat
- makes something that has a home - a poster or a video in Studio, a team run in
- the Organization, a task in the Cockpit, a document or an older conversation -
- it can now take you there. Ask to be shown something and the app goes; nothing
- moves that is not really there, because the addresses come from the navigation
- itself, and a surface whose add-on is switched off says which switch rather
- than opening a page that would only turn you away. In a Telegram or WhatsApp
- conversation, and while a job runs with nobody at the screen, nothing is moved
- at all - the place is named instead.
-
-- **Your teams can be started from the chat.** The Organization no longer needs
- its own screen to be used: ask which teams there are, give one a job, or ask
- what a team is doing right now. A run started this way is the same run the
- Organization surface shows - it appears there while it happens - and it asks
- before it begins, because a team is minutes of work and one model call per
- member.
-
-- **Each agent in a team run uses its own model again.** An agent card has had a
- provider and a model of its own for a long time, and a team run ignored both:
- every member answered from one shared model, however different their roles.
- Now a card's own model is used when it names a provider that has a key on this
- machine, and when it does not, the run log says so in a line instead of
- quietly falling back.
-
-- **Four more places the conversation can reach.** The Group Chat room can be
- looked into and asked a question from the chat; the Desktop Buddy can be asked
- how it is and given a line for its bubble; your own Custom Widgets can be
- listed and read, so an answer about one is about the real thing; and the
- Discover feed can be read, with an honest "offline" instead of an invention
- when the machine has no connection.
-
-- **Every conversation shows what it costs, and can stop itself at a ceiling.**
- A running price sits beside the context meter in the composer footer, each
- answer carries its own price in its token line, and the hover card says how
- much of the input came out of the provider's cache instead of being paid for
- a second time. Under Settings, Goals you can give a conversation a ceiling:
- at half of it and again at the ceiling itself Skales stops and asks, with
- three answers - carry on, switch to a cheaper model, or stop - and the
- question is answered here rather than by the model, so being asked costs
- nothing. A fresh installation starts with five dollars; a machine that
- already has settings is left exactly as it was.
-
-- **Mail does what a mail program does.** Attachments are named on every
- listing and can be fetched into the workspace, a message can be read whole
- instead of as a preview, the mailbox can be searched on the server rather
- than by pulling everything down first, and a message can be put back to
- unread. Searching, reading whole and unread need a mailbox reached over IMAP;
- a Google account connected through Google says which of the four it cannot
- do, instead of answering as though it had.
-
-- **Retrieval over your indexed documents has an off switch.** It sits at the
- top of the block that used to only ask which vector database to use, above
- that choice, because whether anything is looked up comes before where from.
- Switched off, nothing is read from disk and no embedding is paid for, the
- chat command for it says it is switched off instead of "no results", and
- Skales stops offering a search it would not run. A profile that never touched
- the switch keeps retrieval on.
-
-- **The permission window in Skales Code says what a yes never covers.** Every
- window that asks you to allow a step now carries a line you can unfold: the
- handful of things that are always asked, whatever you allow, with one
- sentence each. A session-wide yes is a yes to the ordinary work and not to
- everything, and now it reads that way.
-
-- **Searching your Code sessions finds what they say.** The box in the session
- column searches what was written in a session, not only its title, and the
- results replace the list in the same column: title, folder, date, and the
- passage the hit sits in with the found words marked. The box itself used to
- be two rows tall with the magnifier stranded above the field; it is one row.
-
-- **The scratch folder a coding session works in has a door.** Skales gives
- every coding session a private folder for working files, and nothing in the
- window ever showed it. A line under the All tab of the file column names it
- with its file count and size, opens it in the same file tree the project
- uses, and a file in it opens in the same frame a project file does. It
- appears only when something is really in there, and clearing the session
- empties it.
-
-- **The meter counts the model calls you never see a bubble for.** Two of them
- were free as far as the footer was concerned: the summariser that shortens a
- long conversation to keep it inside the model's window, and the vision model
- that describes a screenshot for a browser or computer-use step. Both are real
- charges on your key, and on a screenshot-heavy afternoon they are the larger
- half of the bill. They now ride on the step that made them and appear on their
- own line in the hover card over the token badge - never folded into the
- sub-agent line, which would name work that never happened. A price that cannot
- be known (a local vision model) still says nothing rather than showing a zero.
-- **A sub-agent may not spend past the ceiling its conversation was given.**
- run_subagent and dispatch_subtasks had a step budget and no money budget at
- all - forty steps per child, as many children at once as the model asked for,
- and the session ceiling only saw the bill when they were all back. Each child
- now runs against what the conversation has left, and a child that reaches it
- says so in words instead of coming back with a bare "did not finish".
-
-- **A recording that could not be transcribed is no longer lost with it.** A
- dictation that failed - a refused key, an empty balance, a dropped
- connection - printed "Transcription failed." and threw the clip away, so the
- only way forward was to say the whole sentence again at the worst possible
- moment. The recording is now held and the error offers to send it once more,
- through the same path that would have sent it the first time.
-
-### Changed
-
-- **Every Skales window now closes the way your operating system closes
- windows, and nothing else.** Iris Orbit had no title bar at all and drew its
- own small close disc - and only once it was listening, so during the intro
- and the name question there was no visible way out. Flow drew one too, in the
- corner Windows draws its own Close button into. Both are gone: Iris, Flow,
- Code and the main window now carry the same title bar - the traffic lights on
- macOS, the window buttons on Windows, the native frame on Linux - and the
- page paints nothing into that corner. Escape, Cmd/Ctrl+W and the right-click
- menu still close what they always closed.
-- **On Windows, the window buttons no longer sit on top of anything.** Every
- header, rail and floating panel that reaches the top edge of a window now
- reserves the width the caption buttons actually take, measured from the
- system rather than guessed - on the left as well, for a right-to-left
- install. A window opened after you switch themes also opens in that theme
- instead of coming up with a black title bar until the page loads.
-- **Studio Classic is down to what it is for: Media, Audio, Type and Gallery.**
- The Design and Scenes tabs are gone. Flow does that work now - it writes the
- same pieces from a conversation, and the multi-scene video lives there as
- Motion, with the timeline, the audio and the export card. The 3D room is
- untouched and still opens from the same door. An old link to one of the two
- retired tabs lands on Flow instead of an empty stage, and a scene render that
- was still running when you updated finishes into the gallery.
-
-- **"Lean prompt" is now its own switch, so a hosted endpoint can send less
- without pretending to be local.** How much goes out per request - the short
- system prompt, the capped tool set, the rest a `load_tools` call away - used
- to hang entirely off "this endpoint runs on my own machine". That single
- question was doing two jobs: the tick decides privacy and what runs where,
- and it was also deciding your bill. There is a second switch next to it now,
- on by default for an endpoint that runs on your machine and off for a hosted
- one, settable either way per endpoint, per provider, or per model in an LLM
- profile. The line the model reads when its tool list was shortened names that
- switch instead of blaming "local models".
-
-- **Reading a file no longer drags ninety tools in behind it.** The always-on
- tools ship on every turn anyway, but each is filed under a category, and the
- bookkeeping read "read_file was called" as "the whole core group is in use" -
- so the first file the model opened put ninety-odd extra schemas into every
- step that followed, paid for on all of them. A tool from the always-on set now
- brings itself and nothing else; the group behind it is still one load_tools
- call away. Measured over twelve steps of ordinary file work: 8.618 to 2.862
- tool tokens a step.
-- **A missing tool costs one group, not the whole catalog.** When the model
- reached for a tool that was not in front of it, Skales reopened every schema it
- has for that turn. The refusal names the tool, though, and a named tool has a
- group - so that group is loaded and the turn runs again, which is the road the
- model would have taken itself. Only a name that belongs to no tool at all still
- opens the full list, because for that there is nothing to load.
-- **The skill tools are one family again.** create_skill, test_skill, read_skill,
- list_skills, delete_skill and the two switches sat in a different group from
- the one the app said they were in, so asking for that group handed the model
- half a family. They are all in `system` now.
-- **Running one at a time is now a property of the model, not a switch for the
- whole app.** The ask was to be able to run requests to a model one after
- another; a global switch would have taken parallel conversations away from
- everybody who runs a cloud model, which is most people. So the run order
- follows the model that will actually answer: a local runtime, or a model small
- enough that the machine is the ceiling, takes turns by default, and everything
- else stays parallel. A choice made in Settings still overrules that, and an
- LLM Profile overrules both - it says `one at a time` on the profile's own line,
- beside the other properties of that model. A team run whose seats do not agree
- takes turns, because the card that holds one model decides for the run.
-- **Asked how it is built, Skales answers instead of refusing.** The prompt used
- to forbid the model outright to look at Skales' own source, which made it
- stonewall even a user who had bound that very folder as a project. It now asks
- for honesty: answer from the docs, the changelog and check_capabilities as far
- as they reach, say "not documented" when they do not, invent no internals - and
- a bound Skales source folder is a folder like any other.
-- **The Cockpit reads as one screen again.** The board no longer grows without
- a ceiling: at fifteen cards it took most of the page and pushed everything
- under it out of sight. The four columns now stop at a share of the window and
- each one scrolls itself - and a board with three cards on it is still three
- cards tall, not a fixed grey rectangle. Under it, Goals, Schedule and Tasks
- are the same accordion with the same shape: the name is said once, in the lid,
- and the page headings that stood one line below it are gone, along with
- Schedule's screen-high floor and the second "Goals" inside the Goals lid.
- Each lid now carries the same one-line summary - how much is moving, how much
- is done, when the next run fires and what it has cost - with the words in the
- tooltip. On their own routes the three pages are unchanged.
-
-- **"Beta" is said in one place, so it means something again.** The word stood
- on the Plugins page, on the Teams surface, on navigation rows and under the
- Skales+ waitlist - five marks for surfaces that have shipped. It now sits
- beside one heading only: Skales Local, under Settings, which brings its own
- inference server and its own models. The bundled manual carries the same one
- mark, and no menu row carries any.
-
-- **Choosing Auto is the whole answer.** Auto mode pre-approves file and shell
- work inside the folder a coding session is bound to - that is what it is for -
- but it only did so when a second, separate opt-in had been recorded somewhere
- else. The Code window never asked for it, so a session started there on Auto
- stopped on the first write into its own bound folder, and the card explained
- the stop with *Auto waves through changes inside the folder, and this is not
- one of those* about a file that was inside it. Picking the mode is now the
- consent, on every surface that offers it. Nothing else moved: work that
- reaches outside the folder, a push, a deploy, a destructive command and every
- network tool still ask, and the folder rules and the blocked-command list are
- untouched.
-- **Every mode says what it may do.** Ask, Code, Plan and Auto are four words on
- four small pills that decide how much a session can do without you. Each one
- now carries its own sentence, on the pill, so the reach of a mode is readable
- where the mode is chosen.
-
-- **Organization, Browser and Playbooks have their doors back.** Two people
- asked for Organization in one week after it lost its menu entry, which is the
- answer to whether taking it out was right. All three stand under Tools and in
- the chat column's More block, and each opens the same surface its own page
- renders.
-- **An add-on you switch off takes its menu entry with it.** Turning Wrapped or
- Skales Code off left the entry standing in the sidebar and the page answering
- as though nothing had been decided. Now every surface with a card on the
- Add-Ons page obeys that card everywhere: out of both sidebars, out of Cmd+K,
- out of the pins. The page itself stays reachable and says where the switch is
- instead of showing you a blank screen.
-- **The Cockpit's Control Room is gone, and its eight controls are where they
- belong.** The deep-dive interview, the daily stand-up and the button that
- lifts a spending pause are at the head of the Cockpit, beside Pause and
- Resume. When the autopilot may be awake, what it may spend in an hour and
- when it stops by itself are settings rows under Settings, Goals. The line you
- type straight into the queue is at the head of Tasks, over the queue it
- writes into. The reset for a wedged scheduler is under Settings, Advanced,
- Diagnostics, where a person looks when something is stuck.
-- **One popup frame instead of two.** The settings jump out of the Guide had a
- hand-copied frame of its own, which is why it kept clinging to the window
- edges on a wide screen while every other panel had learned to keep a margin.
- There is one frame now and both open it.
-- **The Guide has one door.** It lives in the System button in the sidebar
- foot, with Diagnose, Update and Report Bug - the four things about the
- machine itself. The row that said the same thing a second time is gone.
-
-- **The prompt behind every turn got smaller.** Skales handed every model the
- full description of every tool it might want, on every single turn, plus a
- few lists that only matter on the rare turn they are about. The permanent set
- is a third smaller, the rest is fetched by group the moment a turn needs it,
- and every promise the prompt makes about a tool that is not in the turn yet
- names the group it comes from, so a model asks for it instead of guessing at
- a name. On a provider that charges for the prompt, that is money back on
- every step of a long run.
-
-- **A sub-agent runs on the model you chose for it, and says what it cost.**
- Handing work to a sub-agent allowed exactly two answers, the parent's model
- or one fixed fast one; any model you have configured can be named now,
- including one written the way your provider writes it. Each child appears in
- the chat as its own chip with its state, its model and its price, so a step
- that quietly ran on the expensive model is visible while it happens instead
- of arriving as a line on the bill.
-
-- **Telegram and WhatsApp count money per conversation and per day.** The
- spending leash started again at zero for every incoming message, which is no
- leash at all on a channel where messages arrive one after another. It runs
- over a conversation and over a day now, and when it stops a run it says in
- half a sentence which of the two you met and how it differs from the ceiling
- that governs a conversation at the computer.
-
-- **A new installation starts with memory switched on.** The knowledge graph
- and the nightly distilling of finished conversations shipped switched off, so
- the two features that make Skales remember you were only ever found by people
- who went looking. They are on for a fresh install, each row says what it
- costs before you decide to leave it on, and a machine that already has
- settings keeps every choice it made.
-
-- **A long diff arrives folded, and Keep and Revert appear only where they mean
- something.** A change of a hundred and fifty lines unrolled into the
- transcript in full; it comes folded now, with a line saying how much is
- behind it, and a short one stays open. The two decision buttons used to show
- under anything that vaguely resembled a file, including text that was not
- one; they appear on a real change to a real file, and each says what it does
- when you rest on it.
-
-- **A tool group that nobody uses stops riding along.** Loading tools on demand
- only ever added: a group the model pulled on the third step was still being
- paid for on the fiftieth, because the whole conversation was re-read as
- "loaded" on every turn. Now a group nobody has asked for or called in four
- steps is dropped, and pulling it back is the same one-step request it always
- was. A group in genuine use refreshes itself with every call and never
- expires. Measured over a twelve-step run with three loaded groups, the tool
- payload falls by up to 22% per step; over a fifty-step one, by 20% on average.
-- **A model that keeps asking for tools gets the ones it asked for.** When a
- weaker model loops on the load-tools request without ever using what it
- loaded, Skales steps in so it cannot stall. It used to do that by handing over
- the entire catalogue for the rest of the turn - the most expensive corner of a
- turn, on the model least able to use it. It now hands over exactly the groups
- the model named. The full catalogue stays for the one case the narrow answer
- cannot serve: a model that named nothing that exists.
-
-- **A voice you picked is only ever sent to the service it belongs to.** The
- voice Iris and Call Mode speak with was stored as a bare name with no record
- of which service it came from, so an ElevenLabs voice went out to OpenAI as
- if it were one of theirs, came back refused, and was reported as "the chosen
- voice produced no audio" about a voice that works. A voice now carries its
- engine, and a choice made for one service is never handed to another.
-- **The female Spanish voice is on the phone's list again.** It shipped in the
- app and was missing from the list the relay serves, and the served list wins -
- so the voice disappeared from the download screen the moment a device came
- online. The list now carries it, states which speaker of that archive is the
- woman, and says of every voice whether a listener hears a woman or a man, so
- a surface that asks for a female voice can be answered instead of guessed at.
-- **One voice, one name.** The same archive was called three things across the
- desktop catalogue, the relay's list and the phone's - somebody who had heard
- of a voice under one name could not find it under another. Every voice is now
- named once, in the spelling that was already published, and a gate holds the
- lists to it.
-
-### Fixed
-
-- **A Studio visual that animates in now plays instead of freezing on its first
- frame.** A headline arriving word by word is invisible at its own start; the
- guard that rescues a page whose entrance never fired read that as a failure
- and switched the animation off. A page whose text is animating right now is
- left alone, and only the last-resort check seven seconds in still overrides it.
-- **The feed no longer names a surface Skales does not have.** A finished
- Codework run posted "completed a Codework session" - and posted the project
- folder's name with it, into a public feed. It now posts the same content-free
- coding event the Code window posts, and the folder name stays in the private
- local log. The Buddy is the Buddy, not the "Desktop Buddy", and the example
- the post-writer is shown says Skales Code like every other line already did.
-- **A Playground Space you share actually arrives.** The share sent an event
- name the feed had never heard of, so the post was refused and disappeared
- without a word.
-- **`http_request` no longer gets turned away by sites that refuse nameless
- robots.** The tool sent no User-Agent at all, so Wikipedia and other Wikimedia
- sites answered a perfectly correct call with 403. It now identifies itself as
- Skales, the same way the page-reading tools already did. A User-Agent you set
- yourself still wins, and no browser is impersonated.
-- **A conversation carried over from Iris keeps its bill.** The transcript Iris hands to the runner now travels with what each turn cost and which rows were Skales' own cards, the way the Code window already does - so the session ceiling counts an Iris conversation too, and its stop card is never read back to the model as ordinary text.
-
-- **The closing report of a coding run counts your work, not your
- dependencies.** A run that installed packages in a folder without a
- .gitignore signed off with "2109 files changed (+956818)" - a number about
- npm. Dependencies and build output are now left out of the count whatever the
- folder's ignore rules say, and a lock file counts as the one file it is
- rather than as forty thousand lines.
-- **The spending ceiling's card has its buttons in the Code window.** The
- session ceiling stopped a coding turn with the sentence "nothing more will be
- sent until you decide" and nothing to decide with: the three answers were
- written into the session and this window drew only the prose. Continue,
- Cheaper model and Stop are now on the window's own question card, answered
- the way the chat and the phone answer them - Continue re-runs the turn that
- was refused rather than sending the word "Continue", Cheaper model opens the
- picker in the composer, Stop stays stopped - and none of the three costs a
- model call. The card also loses its "something else" box wherever Skales is
- the one asking, on both surfaces: free text there was read as "carry on".
-
-- **A budget notice appears while you are looking at it.** A turn the ceiling
- refuses is over before it starts, so the Code window - which re-read the
- session only when a run visibly moved - showed an unanswered question and no
- reply until the next reload, which looks exactly like a hang. The window now
- follows the turn it just started, and hears it end even when it never saw it
- running.
-
-- **Raising a spending ceiling wakes the conversation it stopped.** After the
- ceiling was lifted the next turn started and the assistant answered "I can't
- run that, this session has passed its cost ceiling" - to the person who had
- just raised it. The stop card is a message from Skales to you, with three
- answers this computer reads itself; it was also being handed to the model,
- which read the transcript refusing to spend and refused again. It now stays
- in your transcript, where you need to see it, and out of what goes to the
- model. The same on the phone.
-
-- **A project folder inside a bigger checkout no longer reports the whole
- checkout as its own work.** Point a coding session at a folder that has no
- `.git` of its own and git answers for the repository above it - so the
- changed-file list, the +N -M counter and the "commit N files" button were all
- describing a checkout this session had never touched, and pressing commit
- would have staged it. The list, the counter, the suggested message and the
- commit are now scoped to the folder you bound, the files are named as you
- would open them, and a line under the branch says the repository root is
- further up.
-
-- **A sub-agent can use the tools it was given, whichever name it types for
- them.** Every child gets a file listing, and children were being told the
- listing was not available to them - so they ran `ls` through the shell
- instead and stopped to ask permission for it. The permission gate compared
- the name the model wrote against a list that only holds one spelling per
- tool, while the tool itself has always answered to both. It now resolves the
- name first, and only through the fixed alias table: a nickname can open a
- door your run already holds open, never one it holds shut.
-
-- **Sub-agents started together each get asked, instead of the first one taking
- the only card.** A fan-out whose children all wanted to write a file used to
- finish one of three: the first child's question went on the card, and the
- other two were told there was nobody to ask and wrote nothing. Children now
- queue for that card and you answer them one after the other, so all three
- finish - or the ones you said no to say so by name.
-
-- **The Code window comes back to the session you had open.** Reloading was
- meant to reopen the transcript, and after visiting any other view it stopped
- doing so: arriving at the Code start screen threw away the pointer to what
- should come back. Opening Code from the sidebar still starts fresh; a reload
- now finds the conversation again, mid-run card included.
-
-- **The spending ceiling applies in the Code window too.** A coding session
- handed its transcript to the runner without the prices on it, so the ceiling
- counted a conversation that had never paid for anything and never showed its
- card - on the surface that spends the most, because a coding turn pays for
- its sub-agents as well as itself. The meter, the warning and the stop card
- now see the same dollars the context panel shows.
-
-- **A Flow project says what it is doing while its tab is closed.** The run
- itself always kept going on the server and reopening the project picked it
- back up - but from the project grid a working project looked exactly like a
- stopped one, and a project parked on its scoping questions looked like one
- that had given up. Cards now say "Still working" or "Waiting for you".
-
-- **A new conversation is named once, not twice, and not with the whole agent
- prompt.** Two title passes ran for every session; one of them sent the entire
- agent identity along to produce six words. There is one now, it carries a
- two-line prompt of its own, and it runs at most once per session - the chat
- page, the Code window and the run itself all ask, and only the first ask
- reaches a model. A naming pass that came back with nothing usable is not
- repeated over the same unchanged first exchange either.
-
-- **The part of the prompt that describes you stops changing under the
- conversation.** The identity paragraph carried three counters that are
- incremented the moment a turn ends - how many conversations there have been,
- how many interactions, a trust percentage - and a list of recent-conversation
- summaries that gains an entry per turn. So turn two of a conversation sent a
- different paragraph than turn one, and it sits in the cached part of the
- prompt. Those values are now held still for the length of a session. Anything
- you changed on purpose - a saved fact, a learning, the character settings -
- still arrives on the very next turn, and the clock still says the real time.
-
-- **The planner no longer generates a plan just because you opened it.**
- Finishing the setup - "Skip to Calendar" included - started a paid day-plan
- request. A plan is made when you press Generate Plan.
-
-- **The clock in the prompt tells the actual time.** It said "this is the real
- current time NOW" while being rounded to the top of the hour, so an assistant
- asked at 05:55 wrote 05:00 into a file, and a whole hour of requests read the
- same time. It is accurate to the minute again, and it sits where being
- accurate costs nothing.
-
-- **A long run stops paying for its own prompt again on every step.** The live
- working-state note - step number, last result, the file ledger - was written
- into the system prompt, which made the prompt a different string on every
- step, so no provider's prefix cache could ever match it: one measured build
- went from 9k to 58k input tokens per call with not a single cache hit, about a
- million input tokens for one job. The note now sits at the end of the
- conversation where it belongs. The model reads exactly the same thing; the
- prompt stays identical from step to step, and providers that cache it charge
- for it once.
-
-- **`src/systems` and `src/scenes` are your folders again.** A project folder
- whose name merely resembled a system location - every game, ECS, Unity or
- Phaser project, and any folder called `dev/` - was refused with a security
- message, and the model quietly built the project under different names. Real
- system locations and credential folders still refuse wherever they sit; your
- own project folders are yours.
-
-- **Setting up your own endpoint now asks which model it serves.** An endpoint
- saved without one made the chat send a hosted catalogue's model id to a
- personal server - which either answered as something it is not, or refused
- with "model not found" - and sent Studio to OpenRouter with a Claude model,
- telling people whose only provider is their own server to add a key for a
- service they never chose. The setup card asks for the model, with a Fetch that
- reads the list off the endpoint and fills the field, and it asks whether the
- endpoint really is a small model on this machine. A turn with no model set
- says so instead of guessing, and Studio stays on your provider.
-
-- **Your Brand Kit actually reaches the design that gets built.** Flow said a
- kit was active, then asked what colours you wanted and built a deck and a logo
- animation in colours the kit does not contain: the kit was never in the
- prompt, and the tool that reads it was the first thing every tool budget threw
- away. Flow now opens with your saved kit selected, the tool that reads it is
- protected the way the file and shell tools are, and the kit's colours, fonts,
- tone and logo go into the run.
-
-- **A custom endpoint is local only if you say it is.** An address on your own
- machine was taken as proof of a small model, so a proxy, tunnel or relay in
- front of a large one was served a compact prompt and a quarter of its tools -
- and then blamed "Max tools for local models" for a tool it had never lost. The
- switch on the endpoint decides now, and an endpoint you have not marked as
- local is treated as a full one.
-
-- **A sub-agent's approval reaches you, and a sub-agent that never got one does
- not claim it finished.** Two helpers were sent off to write a file each; both
- writes needed your yes, no card was ever shown, and the report came back
- "2 of 2 finished" over two files that did not exist. A helper's question now
- lands on the same approval card as the run that started it, saying which
- helper is asking, which tool and which file. Answer once and the helper
- carries on; say no, or send the work off to run unattended where nobody can be
- asked, and it stops and tells you that is what happened - counted as
- unfinished, on the report and on the Tasks page.
-
-- **A long build is no longer mistaken for a stalled one.** A goal that had
- written fifteen modules and got its build green was stopped at step forty
- with "nothing produced", because only bookkeeping counted and a build run
- does its work long before a step is checkpointed. Every file written, every
- edit applied, every command that ran and came back clean now counts as
- progress, and it travels with the goal, so a reloaded run keeps what it
- built. A run that only reads and searches still stops and asks, which is the
- case the halt was built for.
-
-- **A skills repository is imported in one download.** The batch import asked
- GitHub for the branch listing and then pulled every file of every skill one
- request at a time, so a repository of nineteen skills was several hundred
- calls: slow on a good line, and on a bad one a skill landed without the
- scripts its own instructions point at. The whole branch now arrives as a
- single archive, folders and all, and the file-by-file route stays as the
- fallback for what the archive cannot serve. When GitHub does refuse, the
- answer reads its headers and says how long the wait really is instead of
- guessing an hour.
-- **A typed "Stop" stops the run, at once.** The word was let past the queue
- and then walked the ordinary send path, which asks the server a question
- before it does anything else - so the word sat in the composer, nothing was
- drawn, and the turn carried on while the Stop button beside it worked. It now
- presses the same switch that button presses, before anything else happens,
- and clears what you typed. With nothing running the word is an ordinary
- message again.
-- **A typed "Stop" is answered where you can see it.** The confirmation was
- written into the conversation right after the run it stopped, which is the
- same shape a further step of that run has, so the transcript folded it into
- the bubble above and drew nothing. The stop had worked and read as ignored -
- the exact complaint the stop word exists to end.
-
-- **A finished coding run no longer claims Skales was closed under it.** A
- Skales Code run that ended normally, with the window still open, could put
- "Skales was closed while this session was working" over its own summary line
- and leave it standing until the session was reopened. Nothing had crashed: a
- turn ends in two writes, and the window happened to look between them. The
- writes now happen in the order that never spells a death, a run whose last
- moments are still being written is read as finishing rather than interrupted,
- and the window asks twice more before it puts the notice up. A real
- interruption - the app killed mid-run - still says so.
-
-- **A run that ends without you stopping it says so.** A model call that sent
- nothing back for five minutes - a local model finding its first word, a long
- hidden deliberation - was taken for dead and swept away, even though the wait
- Skales allows for exactly that is ten minutes. The run then ended in complete
- silence: no answer, no reason, your question still on screen. Two things
- changed. A step waiting on the model now keeps the run marked alive, the way
- the tool calls around it always did. And when a run is ended by something
- nobody pressed - the sweep, a crashed runner, a restart mid-work - it writes a
- line into the conversation naming which of the three it was, and says that
- nothing above it was lost. A stop you pressed yourself still says nothing
- extra: you already know.
-
-- **Skales no longer asks permission for something it will refuse anyway.** A
- coding session bound to a folder outside your file access rules worked for
- minutes, stopped and asked to write a file, waited for the yes - and then the
- file guard refused the write, because that had been decided before the first
- word was written. The rule is asked first now: when it forbids the path there
- is no card, just the refusal, naming the folders that are allowed and where
- the list is widened.
-- **A folder Skales may not write in says so at the top, not at the end.**
- Opening a project that lies outside the file access rules now puts one line
- above the composer saying exactly that, with the folder and where to add it.
- It was already the empty state of the file column, which is not open while you
- type.
-- **A permission card names the real reason it appeared.** Under Auto the card
- had one explanation for every stop, including stops that had nothing to do
- with folders. It now says *this reaches outside the folder* only when the gate
- actually decided that, gives the gate's own reason when there is one, and says
- nothing extra when the question already speaks for itself.
-- **A run that stops to ask reaches the window you left open.** The Code window
- stopped listening while it was not the window in front of you - which is
- exactly how a coding run is watched: send the job, do something else, come
- back. A run that parked on a permission question went on showing *the model is
- answering* with a frozen clock until the session was closed and reopened. The
- window now keeps following a live run wherever it sits, and a run waiting for
- you counts as live, so the card arrives on its own and the time it has been
- waiting keeps counting.
-
-- **Skales stops telling you it has no Node.** On a machine without a separate
- Node.js install, asking for a script to be checked answered *Interpreter
- "node" not found on this system* - from an application that is itself made of
- Node. It now falls back to the runtime it already runs on: a Node you
- installed yourself is still used first, and only when there is none does
- Skales offer its own. That reaches the shell too, so a command that spells
- `node` finds one as well, and MCP servers and language servers that need
- `npx` stop failing on a fresh Windows machine. npm is the honest exception -
- Skales ships a Node runtime and not npm, so when npm is genuinely missing it
- says so and tells you to install Node.js, instead of failing with a stack
- trace. If a script fails now, the error names which interpreter actually ran.
-
-- **Auto mode stops asking permission for Skales' own scratch folder.** A coding
- session has a working folder of its own that the assistant is told to use, that
- is deleted with the session, and that has nothing to do with your files - and
- Auto raised an approval card every time anything was written there. It was the
- one card that could only ever be answered yes. Writing outside the project,
- pushing, deploying and every destructive command still ask, exactly as before.
-
-- **The work a provider interrupted actually comes back.** After a turn died on
- a 402 or a rate limit, Skales was supposed to put the unfinished items from
- its own checklist in front of the model on the next turn, so "continue" was
- not a word it had to interpret. It never did: the check looked at the end of
- the conversation, and the end of the conversation was the message you had just
- typed, so the answer was always "nothing to resume". The message being sent is
- no longer mistaken for the conversation moving on - and tapping "Not now" on
- the offer still means not now.
-
-- **The session budget also holds for a chat sent from your phone.** The ceiling
- under Settings, Goals is a ceiling on a conversation, but it was only checked
- for turns started at the computer. A turn typed on the phone runs through a
- different path and walked straight past it, on the same conversation and the
- same money. Both paths now ask the same ceiling and produce the same card, and
- the card reaches the phone. The ceiling and the per-day budget on Telegram and
- WhatsApp each say in one half-sentence which of the two you just met.
-
-- **A local model with tools switched off can write a document again.** Skales
- told every model to put documents in the Document panel by calling a tool -
- including models that have no tools at all, because the tool budget or the
- endpoint switch turned them off. Those models followed the other half of the
- instruction, the half that says not to put the document in the chat, and the
- answer arrived nowhere. A turn without tools is now told about the other way
- into the panel, and told to answer in the chat if even that is not possible.
-
-- **Gemini speech models can speak.** Picking a Google text-to-speech model
- under OpenRouter produced "OpenRouter returned no audio" every time. Those
- models accept only raw samples, Skales asked every model for MP3, and the
- 400 that came back said so in a sentence nobody ever saw. The request format
- now follows the model - raw samples for the Google family, MP3 for the rest -
- the samples get a WAV header from the rate the answer itself states, and the
- browser, the Studio, the AIPointer overlay and the Telegram and WhatsApp
- voice notes all play what actually arrived instead of what was assumed.
-- **A voice that will not speak says why.** The reason OpenRouter gives -
- a refused key, an empty balance, an account setting that rules an endpoint
- out, a format it does not take - now reaches the preview line and the note
- under the answer, in OpenRouter's own words. "Returned no audio" is kept for
- the one case that means it: an answer that really was empty.
-- **Settings says who is speaking.** A line at the top of the speech block
- names the voice that will actually be heard, resolved the way the cascade
- resolves it. An on-device voice set as the main one speaks before any
- provider you picked, deliberately, and the line now says that instead of
- leaving a screen where the highlighted tile, the hint underneath it and the
- voice coming out of the speaker were three different answers.
-- **The speech voice is chosen where it is heard.** The voice for OpenRouter
- moved out of the provider card into the speech block, and it offers the names
- the chosen model accepts - the six OpenAI names, or the Gemini names for a
- Google model. The block used to show a browser voice list that did nothing
- while the working field sat on another screen.
-- **Iris keeps her voice on the on-device path.** Iris asks for a female voice.
- The request stopped at the door of the on-device engine, so she spoke with
- whichever single voice was installed. The catalogue now states the gender of
- each voice, the preference picks among the voices for the language being
- spoken, and when there is no matching one the answer says which voice spoke
- and why instead of quietly swapping her.
-- **Listening can be switched on.** Setting the on-device engine as the ear
- needs a model, and the model picker only appeared once the ear was already
- on - a circle with no way in, which sent the setting back to off without a
- word. The picker is always there, a click carries the only installed model
- with it, and a role the store refuses is now said out loud.
-
-- **Stop stops.** Pressing Stop wrote "stopped" into the record and left the
- work running: a script that had been started kept going, everything it had
- started kept going, the next tool in the batch still ran, and the chat could
- sit there looking idle while money was still going out. Whoever ends a run
- now ends the runner with it, on every route into it - the button, the word
- "stop" typed into the chat in any of the twelve languages Skales ships, and
- the sweep that retires a run nobody is watching. The signal is checked before
- every single tool of a batch, a running command is taken down together with
- the processes it spawned, and a browser step closes its page. A long tool
- call no longer looks abandoned while it works, so it is not retired out from
- under you, and while any loop is alive the chat says it is running instead of
- reading a record that says otherwise. Quitting Skales takes the whole tree
- with it, and a server whose app is gone ends itself.
-
-- **Skales no longer hands out its own stored credentials.** The file holding
- your mail login was refused to the file reader and to nothing else: a script
- or a shell command could open it, a search one folder up printed it line by
- line, and Skales' own instructions actually named the file as something worth
- inspecting. In one reported run a model read the password out of it, logged
- into the mailbox by itself and marked seven hundred and sixty-three messages
- as read. Every stored credential file is refused now to the file tools, to
- the shell, to scripts and to search, in every mode including the unrestricted
- one; a search skips them rather than printing what reading them would have
- refused, and the instructions say to work through the mail and integration
- tools instead of signing in by hand.
-
-- **Your mail password is no longer written down in plain text.** Provider keys
- have been kept encrypted for a long time; the mailbox password sat beside
- them readable, in the file and in every copy of it - a backup, a sync folder,
- a screenshot, a transcript. It is encrypted at rest now, in both places it
- was kept, and an existing account is converted the next time Skales loads it,
- with nothing for you to do. Honest about the reach: this stops the password
- from being readable, it is not protection against somebody who already has
- your machine.
-
-- **Nothing changes twenty things in your mailbox without asking.** One
- instruction can carry a loop, and a loop is where a small mistake becomes a
- large one. Past twenty changed messages in a session Skales stops and asks,
- and it asks in the unrestricted mode too, because that mode is about trusting
- your judgement rather than about not being told. The count is kept on disk,
- so it survives a restart in the middle of a run, and a yes covers what was
- asked and not everything after it.
-
-- **Marking a message read works, and so do moving and deleting.** On a mailbox
- reached over IMAP the step that changes a message never opened the mailbox it
- was about to change, so the change went nowhere and the answer said it had
- worked. It hit the three most common ways of naming folders, which is to say
- most of the mailboxes people actually have, and the same one cause broke
- moving to another folder and moving to trash. All three do what they say now,
- and the connection is closed even when a step fails.
-
-- **An answer written while the phone bridge was down is delivered when it
- comes back.** If Skales was closed in the middle of a WhatsApp job the job
- was picked up again at the next start, and its finished answer was thrown
- away with a line in a log, because the restart that revived the job had not
- brought the bridge up yet. The answer is kept now and sent once, with the
- time it was written and a note saying it is arriving late.
-
-- **A Telegram message waiting for your permission stays visible.** A turn that
- stopped to ask you something closed its own record on the way out, so the
- list of what is running showed nothing and the run looked finished. It shows
- as waiting for permission with the time it started waiting, it closes when
- you answer, and a run that was waiting when Skales was killed comes back
- waiting rather than as something to continue.
-
-- **A screenshot is handed over once, not on every step afterwards.** Every
- further step of a turn re-sent the whole picture, so a run that took a
- screenshot early paid for it again and again. The picture goes into the turn
- once; after that the step carries the description and where the file is.
-
-- **A retired model says that it is retired.** A model its operator has
- switched off answers with a status that fell through every rule and arrived
- as a bare error, so it read like a fault in Skales or in your key. It is
- named for what it is now, in twelve languages, with the advice to pick
- another model. An account without credit is deliberately not called a dead
- model: that is a different problem with a different fix.
-
-- **A tool that twice returns nothing useful is not asked a third time.** A run
- could spend its whole budget rephrasing the same fruitless lookup. After two
- empty answers from the same tool that channel is closed with a sentence
- saying what could not be verified and why, and the run carries on with what
- it has. A real find resets the count, and nothing is taken away from a run
- that is getting somewhere.
-
-- **Right-click, Download in Skales Code saves the file.** It worked most of
- the time and failed the rest, for a reason that had nothing to do with you:
- the download was cleaned up in the same breath it was started, so whether it
- arrived depended on how busy the machine happened to be. It arrives now, and
- a download that genuinely fails says so on the screen instead of nowhere.
-
-- **"Always allow, this session" survives a reload.** The answer was held only
- in memory, so reloading the window, or Skales rebuilding a piece of itself in
- the background, threw it away and the same question came back a few steps
- later. It is written to the session now and comes back with it. It still
- covers only that session, it still lifts no guard, and clearing the session
- throws it away in both places.
-
-- **Switching Skales Code off closes Skales Code.** The switch on the Add-Ons
- page left the window and its address working exactly as before. The surface
- says which switch turns it back on, and it waits until the add-on list has
- really been read, so the notice does not flash up on every start.
-
-- **A reload reopens the session you were in.** Reloading the Code window
- dropped you on the start screen, with the session you had been working in one
- click away and no reason given. A reload comes back to it; opening the window
- fresh still starts where it always did.
-
-- **A command you run gets your shell, not the app's build environment.**
- Skales handed its own build settings down to everything it started, so an
- install inside your own project quietly skipped the development dependencies
- and left you with a project that would not build. The build settings stay
- with the app now, and anything you run sees what it would see in your own
- terminal.
-
-- **A command that did its work and returned a non-zero code no longer reads as
- a total failure.** Plenty of useful commands end with a non-zero code and a
- perfectly good answer on their output, and that answer was replaced with
- "command failed", which sent the model looking for a problem that was not
- there. A run like that says which code it ended with and shows what it
- printed. "Failed" is kept for the cases that earn it: no output at all, an
- error stream, or a timeout.
-
-- **A file you attach to a coding session can be found by it.** A session bound
- to a folder was told the file's name and nothing else, so the first thing it
- did was hunt for it, sometimes in the wrong place. The note carries the full
- location now, for video, documents, spreadsheets, archives and anything
- binary alike.
-
-- **The GPU badge in Skales Local now says what happened, not what was hoped.**
- It used to read "GPU" whenever the bundled engine had been built with a
- graphics backend, whatever the machine then did with it - so two Linux users
- with an RTX 3070 watched it say GPU while llama-server was reporting that it
- had asked for 999 layers on the card and put nothing there, and the whole
- model was running on the processor. The badge is now read from what the engine
- actually loaded. Until a model has been loaded it says "GPU build" instead of
- claiming a run that has not happened, and when nothing goes across it says CPU
- with the reason underneath: no Vulkan driver found on this machine, no
- graphics device found, or a card that was found and given no layers - each
- with the next step, which is a different one in every case.
-
-- **A local model on the processor is given the time it needs, and the message
- says so.** A model computing on the processor takes minutes a turn, and the
- fixed request budget cut it and reported "the connection to the provider
- closed before the model finished answering". Nothing had closed. Skales now
- measures how long this machine is actually taking with this model and lets the
- budget follow it, and a turn that is still cut short says how many seconds a
- turn has been taking and where to look, instead of blaming a network that was
- never involved.
-
-- **Qwen 3.5 no longer prints its own tool calls into the chat.** Served through
- llama.cpp, Qwen 3.5 writes a tool call in a dialect the endpoint does not
- translate, so the raw markup arrived in the answer and the tool never ran -
- reported by a user whose model tried to read an image and printed the attempt
- instead. Those calls are now recognised and executed like any other, and
- anything left over is cleaned out of the answer. Skales Local also reads what
- the running engine says about a model's image support, so a local model with
- its projector in place is no longer judged by its name alone.
-
-- **Updating no longer leaves processes behind.** The restart that applies an
- update tore the app down with a shorter routine than an ordinary quit uses -
- one that stopped the web server and the bots and never touched the local model
- engine. Two Linux users were left with Skales and llama-server processes still
- running after the update finished. The update now uses the same teardown as a
- normal quit, including the sweep for orphaned engines.
-
-- **On-device voices no longer fail with "Please check your config!"** Every
- Piper voice on a machine could refuse to speak because the shared speech data
- had once been half-created: Skales checked that the folder existed rather than
- that it was complete, so the full copy inside each newly downloaded voice was
- deleted on arrival and the broken folder was kept. It is now checked for the
- files the speech engine actually validates and replaced when any are missing,
- and if a voice still refuses, the message names the folder and the missing
- file instead of four words that point nowhere.
-
-- **A browser click that moves nothing is no longer called a success.** Asked
- to post on X, Skales wrote the text into the composer, clicked Post at
- coordinates a vision model had guessed, and the page did not move - and the
- click was still reported as having worked. Reading the timeline afterwards,
- the model found a post from a year earlier and announced that it had
- published. Nothing had gone out. A click is now measured on both sides on
- every route it can take, and a page that did not move is a failure that says
- so and says what to do instead. A click aimed at Post, Send or Submit comes
- back with either the address that changed or the words "publication
- unconfirmed", and the answer states plainly that a post visible in a
- screenshot is not proof that yours went out - it may be an older one, or
- somebody else's.
-
-### Security
-
-- **Every connected account's credential is encrypted on disk.** The mail
- password already was; the Telegram bot token, the Google account's refresh
- token, the Discord, Slack, Signal, Notion, Todoist, Spotify, YouTube, Home
- Assistant, Twitter/X, Outlook and VirusTotal keys and the webhook secret were
- written in clear text beside it. All of them now go through one gate on the
- way to disk and back, with the per-installation key the rest of Skales
- already uses. A store that still holds a clear-text key is sealed the first
- time it is read and never again after that, so nothing has to be re-entered.
- A backup still travels as portable plaintext and a restored one is sealed
- again with this machine's key, so moving an install keeps every connection.
-- **Publishing asks once, in every safety mode.** Unrestricted Mode says "stop
- asking me whether this is risky"; it never said "publish in my name without
- telling me". The first call of a conversation that puts content where other
- people can see it - a tweet, a WordPress post or page, a Slack, Discord,
- Signal, WhatsApp or Telegram message, an FTP upload, or a browser click on
- Post, Send or Submit - comes back as the same approval card the mailbox
- volume check uses. Answer it once and the rest of the run goes through; the
- answer survives a reload, so a resumed run cannot publish without ever having
- asked. There is no new setting, and Auto cannot turn it into a standing yes.
-- **The blanket "approve everything in this session" answer survives a
- reload.** Its marker was an invisible character, which the session file's own
- cleanup quietly rewrote into something the reader no longer recognised - so
- the card came back after a reload even though it had been answered. The
- marker is now written out in full, and an answer given by an older build
- still counts.
-
-## v12.9.25 - Backbone
-
-### Added
-
-- **Memory refuses to hold your secrets.** A saved memory is not stored once
- and forgotten: it is read back into the prompt of every later conversation,
- on whichever model you point Skales at next. So a value shaped like a card
- number, an IBAN, an API key, a token, a password or an id number is no longer
- written down. Skales says plainly that it did not save it and where the thing
- belongs instead - a provider key goes in Settings, where it never enters a
- prompt. Ordinary facts are untouched: a budget, a phone extension, a version
- number and a time of day all still get remembered. If you genuinely want such
- a line kept, there is a switch under Settings, Memory; the background pass
- that distils facts out of finished conversations ignores it and never keeps
- one, because nobody asked for what a background pass happened to notice.
- Automatically distilled facts are also capped at one line, so a paragraph can
- no longer arrive as a "fact" and sit in every prompt afterwards.
-
-- **A tool cannot act on something it never saw.** Before Skales sends a mail,
- changes a WordPress post, moves a calendar entry, deletes a file or updates a
- task, it checks that the thing it is about to touch actually came up in this
- conversation - from something it read, or from something you typed. An
- identifier the model produced out of thin air no longer reaches your site,
- your mailbox or your disk: the step stops, says which identifier it does not
- recognise and which lookup would settle it, and the run carries on from
- there. The check happens before you are asked to confirm anything, so a
- confirmation card is never a question about an invented id, and what a
- conversation has looked up stays with that conversation - a second chat
- starts from nothing, and a run that survives a restart keeps what it knew.
-
-### Changed
-
-- **One engine drives every run.** Chats, scheduled tasks, cron jobs, plugins,
- webhooks, Discord, Telegram, WhatsApp and the desktop Buddy now run the same
- agent loop, with the same guards and the same budget. A request no longer
- finishes when you type it and gives up early when it runs unattended.
-- **The budget you set now holds everywhere.** A Buddy bubble, a scheduled
- task and a message sent from Telegram or WhatsApp are counted against the
- same token and spending limits as a chat, sub-agents included. A run that
- stops because it reached one says so, instead of blaming the work - and the
- four of them now show up in the run list while they work, so a long one can
- be watched and stopped. If Skales closes mid-task, a Telegram or WhatsApp
- turn comes back as an unfinished run like any other, and the answer reaches
- the chat it was asked from.
-- **A run survives a restart.** If Skales is closed, updated or crashes while
- a chat, a Code session or a goal is working, the run is no longer thrown
- away. It comes back in the conversation as an unfinished run, says how far
- it got, and one Continue picks it up at that step. Nothing already done is
- repeated. Under Autonomous an interrupted run picks itself up on the next
- start and writes a line saying it did. An approval you never answered
- because the app closed comes back with the run.
-- **Sub-agents that come back.** Skales can hand part of an answer to
- sub-agents and wait for them: ask it to check four files or four sources
- and it runs four workers at once, each with its own context and budget, and
- returns with the conclusions. A sub-agent can work in its own checkout of
- your repository, so two of them never touch the same tree, and what they
- changed arrives in your folder as normal pending changes, hunk by hunk.
- Sub-agents cannot start sub-agents.
-- **Budgets and a context meter.** A run can be held to a token budget and a
- spending limit, not just a number of steps, and says which one it reached.
- A running goal shows how full the model's context is, how many steps and
- tokens it used and what it cost so far; where a cost cannot be worked out,
- the total says it is incomplete. Every chat now carries the same reading
- under the composer, measured on what the last answer actually sent, and it
- names what filled the window: the system prompt, the tool definitions, your
- skills, your memory and the conversation itself. Type /compact in the chat or the Code
- window to shorten a long conversation on the spot. Long runs on Claude
- models reuse the conversation from the previous step instead of paying for
- the whole transcript again. A run that cannot shorten its own conversation
- stops and says so, instead of ending on a provider error that explains
- nothing.
-- **One job store, one clock.** Planner tasks, goal schedules and cron jobs
- were three schedulers with three stores and three failure rules. They are
- one now. Workflows, playbooks, team runs and goals can be scheduled on the
- same form, the Playbook "Scheduled" badge is true, and a job can be fired
- by a watched folder or a watched mailbox instead of the clock. Watched
- mailboxes read each message once, even across restarts. Plugins that
- declare a folder or mailbox trigger are armed and stopped with the plugin.
-- **Discord answers, webhooks become tasks.** The Discord bot starts with the
- app, reaches the same engine your chats do, and an action that needs your
- yes appears as Approve and Deny buttons in the channel. A webhook arrives
- on the Tasks page with its own log under the same confirmation rules. The
- dashboard no longer says a bot is running when its messages go nowhere.
-- **One approval, every channel.** A request to approve an action is a single
- record that Telegram, WhatsApp, Discord and the app all decide; answer it
- once and it is gone everywhere. WhatsApp can approve now, as a numbered
- question. A run that needs approval on a channel that cannot ask says so
- straight away, and an approval that runs out of time says so in the chat.
-- **A System button in the sidebar foot.** Between Settings and Stop, it
- opens Update, Report Bug, Logs and diagnostics, the Guide and the DevKit as
- links; Logs, Update and Report Bug left the menu for it. Crashes and failed
- calls moved from Settings to the Logs page, beside the log they describe;
- Settings keeps a signpost. Cmd+K finds every page, including the ones
- without a menu row. One autonomy switch with one name: Autonomous, with its
- time window; the Goals section that was called Autonomy is Working style.
- The sidebar itself is as it was in 12.9.21: Home and Chat, Dashboard on top.
-
-### Fixed
-
-- **The Zotero preset carries its key.** The llm-for-zotero server protects
- its endpoint with a bearer token, and the preset from 12.9.21 claimed no key
- was needed, so the connection failed at the first call. The quick setup now
- reads the token off your Zotero profile and fills the Authorization header;
- where it cannot, the field is on screen with the path to the token.
-- **Text a tool fetched can no longer imitate you.** Everything Skales reads
- from somewhere else - a web page, a mail, a file, an entry from a connected
- service - is handed to the model inside a marked-off block, so it is treated
- as material to report on rather than as an instruction. That block is now
- cleaned before it is handed over. Characters that render as nothing but still
- read as text to a model are taken out, so a page can no longer hide a
- sentence in the gaps between the words you can see. Text laid out to look
- like the conversation resuming - a blank line and then "assistant:", or a
- fragment of transcript markup - is defused, so a page cannot put words in
- Skales' mouth or invent a turn that never happened. And a page carrying a
- copy of the marker itself can no longer close its own block early and have
- the rest of it read as instructions. Ordinary writing is untouched: a
- sentence about a system, a heading with a colon in it, and a config file full
- of angle brackets all arrive exactly as they were written.
-
-- **Every part of Skales gets that protection, not just the chat.** The marked
- block used to be put on where each conversation type built its own messages,
- which meant a scheduled task, a background agent, a remote session or a run
- resumed after a confirmation could reach the model without one. It is now
- applied at the single point every request passes through, so chat, Buddy,
- Telegram, WhatsApp, agents and scheduled runs all get the same treatment. A
- block that had lost its ending - which a long, compacted conversation could
- cause - is repaired rather than passed on, so the rest of a conversation can
- never be swallowed into it.
-
-- **The clock stopped costing you money every minute.** Skales tells the model
- what time it is, and that line used to carry the minute. Every time the minute
- rolled over, the provider's cache went cold and the whole conversation was
- charged at full price again - on every step of a long run. The line now says
- the date and the hour, which is all it was ever read for, so a turn that has
- not changed anything reads the same bytes as the one before it. Reminders and
- scheduling are unaffected: they are computed from the real clock, and the
- times you see in the app are exact as they always were.
-
-- **A small model that circles is asked for the answer instead of left to
- spin.** Asked for its name and what it can do, a small local model would go
- round and round: what can I do, which skills are ready, what is in the
- workspace, what can I do again. Nothing caught it, because it never repeated
- itself - a different read-only tool every round. Skales now watches whether a
- round brought anything NEW, and after four that did not, it takes the tools
- away for one round and asks for the answer from what is already on the table.
- A run that is genuinely working is untouched, however long it runs. Asked who
- it is or what it can do, Skales answers from what it already knows: its name
- and its capability list are in front of it and were never something to look
- up. Reported on a 0.8B model on Linux, where the same question took twenty-two
- minutes and eighteen tool calls.
-- **A conversation is never named after a tool call.** A model whose server does
- not parse function calls writes them out as text, and a chat in History ended
- up called `search_web text: "sk..."`. The naming pass now skips the machinery
- and looks for the first real answer, and a name that is itself a tool call is
- refused outright, in History and in the chat window alike.
-- **A local model's own context window is what the request is measured
- against.** The safety trim that keeps a request inside the window was
- measuring against a guessed default rather than the window the local server
- reported, and against the chat model rather than the one actually answering
- when a different model had been chosen for code.
-
-
-- **The chat stays still while an image turn is regenerated.** Regenerating a
- turn that carried a picture put the "n/n" version switcher under the new
- answer in and out of the bubble on every poll, and a transcript pinned to the
- bottom stepped up and down with it for as long as the model worked. The
- archive of earlier answers now goes onto the rebuilt list before it is
- shown, so the switcher stays where it is from the first token.
-- **Ollama takes a picture again with keep_alive set to "keep loaded".** The
- "-1" and "0" presets went over the wire as text, which Ollama's native
- endpoint reads as a duration and refuses with "time: missing unit in
- duration". They are sent as numbers now; the chat endpoint had quietly
- ignored the field, which is why only image turns failed.
-- **A batch skill import lists the repository once.** It used to ask GitHub
- for every folder of every skill, and sixty unauthenticated calls an hour ran
- out halfway through a skills repository, so the second half arrived without
- its scripts. One call now fetches the whole tree, the files come from the raw
- host that has no such limit, and the import page lists every error by skill
- and reason instead of saying "1 error".
-
-### Removed
-
-- The Codework and Swarm pages, the parked Business, Cast, Datasets and
- Network pages, and a Studio publish route nothing called. Codework has been
- a mode inside Skales Code for a while and Swarm is the /swarm chat command.
- A build check now makes sure every page is either in the menu or declared
- doorless with the reason written down.
-- The AIPointer overlay can start apps again instead of answering that the
- feature is not built yet.
-
-## v12.9.21 - Cockpit 2.0
-
-### Added
-
-- **Cockpit 2.0.** One head, one action. The Cockpit opens on an execution
- board: everything that stands, in four columns - pending, in progress,
- completed, blocked - with goals, your own tasks, Autopilot tasks and
- schedules as cards you can filter, drag and act on. Goals, Schedule and
- Tasks sit below as sections that fold open; Live and History are one switch
- away; Identity & Memory and the Control Room keep their place as sections.
- The two stacked tab bars, the second title and the icon chips are gone, and
- finished goals finally show up as finished.
-- **Every model gets a "Can this model see images?" switch.** Beside the model
- field of each provider card: Auto, Yes, or No. Skales used to guess from the
- model name and from what a running daemon reported, and there was no way to
- correct a wrong guess except for custom endpoints. Your answer now wins over
- both, in either direction, for that model at that provider. GLM-5 and Qwen 3.5
- are recognised as able to read images by rule rather than by an exact name.
-- **Answers cite the memories they used.** A line under the reply lists what
- Skales recalled, each with its date, and tapping one jumps to that entry on
- the Memory page. The Buddy shows the same line from the same recall.
-- **Agent lessons have a home.** What an agent learned from its runs shows on
- the Memory page as its own kind, correctable and deletable per line.
-- **Skales Code adopts what a project already has.** Rules written for other
- editors are named, MCP servers from a project made elsewhere can be imported
- (switched off, with their command line shown), and files a project tracks on
- purpose under an ignored folder stay visible in the tree.
-
-### Changed
-
-- **The card that asks how you like to work asks each question once.** Answer
- it or wave it off and it is gone for good, in the feed and on the Memory page.
- Every question is answered by tapping - no open text box, no model call - and
- there are six more of them to draw from. Answering shows what was saved and
- then clears the card away.
-- **Plugin pages are the page.** The job box only shows for plugins that have no
- page of their own; a page plugin gets a small "Give a job" in its header
- instead of a permanent box above the tool. Something a run prepared and parked
- now waits where the thing is - on the page, next to the letter or invoice -
- and the frame sits flush in the window with one scrollbar instead of two.
-- **Bundled plugins open on what they are for and what to do first**, instead of
- on a row of counters reading zero. Data Broker Exit offers the step you can
- actually take - fetch the register, record your details, pick the companies,
- then check for answers. Open Invoices no longer tells a fresh install that
- every invoice is paid. Notice Radar, Page Watch and Receipt Inbox name the
- first step, and a button that cannot do anything yet says why. Installed
- copies show "Update to 1.1.0" on the plugin page.
-- **Bundled plugins update themselves when nothing about their rights changes.**
- A newer bundled version that asks for no new tool, network or file access is
- applied on the next start or page visit; one that widens its rights still
- waits for your approval card.
-- **A room with nobody else in it says "sent", not "sending" forever.** The
- message was delivered to the relay; there was simply nobody to confirm it.
-- **The DevKit command line is `skales-dev`.** The app's own `skales` command
- keeps its name; the docs say which is which.
-- **Removing someone from a room now tells the relay.** The removed device stops
- receiving that room's pushes at once instead of when it next says so itself.
-- **Advisor strategy answers plain chat on the executor.** The advisor only
- steps in for a turn that really needs a plan; a normal message no longer
- routes to the planning model.
-- **Every turn carries less prompt.** The release notes and the page catalogue
- ship as short lists and the full detail loads on demand; facts told twice
- are told once; the stable part of the prompt stays ahead of the cache line.
- A model that was handed the whole tool catalogue because it once fumbled a
- call gets the short one back after its first real tool call.
-- **Plugin cards say how a plugin reaches the web.** "Sandbox offline, reaches
- the web through Skales tools" instead of a bare "No network" on a plugin that
- sends mail or fetches pages through the host.
-- The plan header says "Step 2 of 5" instead of "2/5".
-### Fixed
-
-- **A note sent to your phone has a place to land.** `send_to_phone` delivered
- the note and the push arrived, but tapping it opened nothing, because the
- phone had no destination for it. The note now lands in the open conversation
- on the phone under "From your computer", the tool description, the docs and
- the app's own self-knowledge say so, and the guide has a section on it.
-- **Image generation on OpenAI and OpenRouter works again.** The OpenAI path
- sent a `response_format` parameter that the gpt-image models reject, and a
- size they do not offer; both now follow the model family. Image-only
- OpenRouter models were called on the chat endpoint and answered 404; they now
- go to the images endpoint, hybrid models stay on chat, and a model that fits
- neither gets a sentence instead of a status code. The image-model picker in
- Settings lists both catalogues.
-- **A custom endpoint is no longer treated as a local model just for being
- custom.** The address decides, and every custom endpoint card carries a
- "Runs locally" switch for the cases the address cannot tell. A hosted
- frontier model behind your own URL gets the full tool set, not the 70-tool
- cap meant for small local models.
-- **Importing skills from GitHub brings the whole skill folder.** The batch
- import used to write only the SKILL.md and drop the scripts beside it; single
- and batch import now go through one installer that mirrors scripts,
- references and assets.
-- **Zotero is a one-tap MCP preset.** The quick setup lists the llm-for-zotero
- server, so a library can be connected without typing the command by hand.
-- **Small local models that print their reasoning as the answer no longer do
- so.** Thinking that arrives without any tag folds into the reasoning panel,
- while it streams and once it lands; when Skales is not sure, it shows
- everything rather than hiding a real answer.
-- **The chat no longer steps up and down while the assistant works.** The
- status line stayed put once per run instead of remounting on every poll, and
- the auto-follow only scrolls when there is more to see.
-- **An attached picture tells the truth about who sees it.** The note under an
- image no longer claims the model can already see it when the model is blind
- or when a separate vision model is doing the looking. "This model cannot read
- images" names both ways out - the switch and a vision provider - where it
- happens.
-- **Quitting Skales sweeps up llama-server processes** left behind by an
- earlier session or a crash, not only in the emergency backstop. Restarting
- from inside the app now shuts down everything a normal quit shuts down.
-- **A scheduled agent that finishes with nothing to say is marked failed** and
- says why, instead of being filed as a successful run.
-- **Installing the TypeScript checker for Skales Code works again**: the install
- pinned no compiler version and the newest one has no tsserver, so every
- request came back "unresponsive".
-- **The build gates no longer go blind** on a file whose line comment mentions
- a path with a star in it.
-- **A Windows update that cannot install now says so.** The installer used to
- run hidden, so anything that stopped it - Windows warning about an unsigned
- setup, or one last Skales window still holding a file - ended the update in
- silence, and the next start was the old version again. It now installs in its
- own window: the warning and the "close Skales and retry" prompt are on screen
- where they can be answered - and because the setup now knows it is an update
- of an app that is closing itself, it waits for that instead of asking at all.
- Skales still restarts by itself afterwards.
-
-## v12.9.20 - Slim
-
-### Added
-
-- **Your agent carries its working mood in the open now.** The Conscious bar
- and the companion behind it - interests, mood, the day's movements - are on
- from the start instead of hidden behind a switch most people never found.
- It only ever draws: no notification, no sound, nothing sent. One tap in
- Settings > Memory turns it off, and anyone who had turned it off stays off.
-- **A pixel gecko lives on the Code landing.** Between the greeting and the
- composer, 120 by 80 pixels of mascot, with nine little loops - it naps,
- hunts a bug out of a terminal, sheds its skin, catches a fly. Every visit
- picks a different one, never the same twice in a row, and its dark details
- follow your theme.
-
-- **Call Mode is Iris now.** A call used to run on a second, smaller brain in a
- conversation of its own: no skills, no goals, no access limit, and its answers
- landed somewhere you were not looking. It is the same mind as the voice window
- now, in the same conversation, with everything that mind can do - and if it
- needs permission for something it asks out loud and takes "yes" or "no" for an
- answer. Hang up and the chat behind it is where you left it.
-
-- **A result in the ring can be a circle now.** Short answers, a picture, a
- short list and the countdown get a round frame with the text following the
- curve; anything longer, and anything with headings, lists, tables or code,
- keeps the panel it has always had. Long text scrolls by itself, slowly enough
- to read along with - and the moment you touch it, with wheel, trackpad,
- scrollbar, keyboard or by selecting a word, it stops for good.
-
-- **A web page can live inside the ring.** Ask to see a page and it appears in
- the frame as the page, with its pictures, rather than as a wall of text - and
- the browser only opens when you actually ask for the browser. More of what
- Iris finds now has somewhere to be shown: what you searched, what you listed,
- what you read.
-
-- **You can watch a design being made.** Studio, the chat and Flow now show a
- running line while a visual is being written - which stage it is at, how much
- has arrived, how long it has been going. Before this you looked at nothing at
- all for as long as it took, with no way to tell a slow model from a stuck one.
-
-- **A design that got cut off says so, and Skales finishes it.** When a model
- runs out of room mid-page, the half that arrived used to look exactly like a
- finished one - and the next thing you did was share it. Now it is labelled,
- Skales writes the rest once by itself without asking, and if that is still not
- enough there is a button to carry on. It stops on its own when a model starts
- going in circles instead of burning your budget on it.
-
-
-- **Skales remembers what yesterday was about.** Once a conversation has been
- quiet for a while, it writes four short lines about it: what it was about,
- what got settled, what was still open, and which of your things came up. The
- conversations of one day are folded into a single account overnight, so
- "last week" has an answer too without anything having to read eight hundred
- messages again. Opening a new chat brings the open ends of the last one with
- it, quietly, instead of asking you to explain it all a second time.
-
-- **"What did we talk about yesterday?" is a question that works now.** The
- search through past conversations understands periods of time, and it looks
- for meaning rather than for letters - so it still finds the afternoon you
- mean when the words you used back then were different ones. Alongside it
- there is the cheap way: the short account of a conversation, instead of the
- whole conversation.
-
-- **You can see where a saved line came from.** Every line on the Memory page
- now says whether you told Skales yourself, whether it was noted during a
- chat, whether the nightly pass found it, or whether it came over from your
- phone. A line from before this existed says nothing rather than guessing.
-
-- **A saved line can be corrected, not only deleted.** A memory that is right
- about the subject and wrong about the detail used to cost the whole record.
-
-- **Skales brings along what hangs off the thing you are talking about.** Until
- now it could only look something up when the word was there. Mention a
- project, and the people and tools that belong to it come with it, without
- having to be named.
-
-- **What you keep coming back to now shapes an ordinary conversation.** It was
- learned already, and it reached the messages Skales sends on its own, but not
- the chat in front of you. It says what you care about, which is a different
- thing from where a conversation stopped, and both are used quietly rather
- than read back to you.
-
-- **Ten plugins ship with Skales, and they are the point rather than the
- demonstration.** The shelf used to carry two, and neither answered the
- question of what a plugin is for. It now carries three that answer a question
- about your own exposure, three a small business runs on, and four about the
- working day. **Exposure Check** tells you whether a password has been in a
- breach without the password leaving the machine: it is fingerprinted here and
- five characters of that fingerprint go out, and the same page names which of
- the services you use have been breached and exactly what came out of each.
- **Attachment Check** answers for a file somebody mailed you in two steps -
- first a lookup by fingerprint, which publishes nothing, and only then, if you
- press it, an upload that does. It follows a link without opening one, and
- shows you where it actually ends up. **Data Broker Exit** works down the
- Californian broker register, writes each company the erasure request the law
- where you live actually gives you, tracks the deadline it has to answer by,
- and chases the ones that go quiet. **Open Invoices** keeps who still owes you
- money and has the reminder written by the time it is overdue. **Notice Radar**
- keeps every contract with its notice period and warns you long before the day
- you can still cancel. **Receipt Inbox** collects the receipts that arrive by
- mail into a month you can correct and hand on. **Page Watch** tells you what
- changed on a competitor's page since last week. **Meeting Prep** reads
- tomorrow's calendar and leaves a briefing on whoever you are meeting.
-
-- **Nothing any of them can do happens quietly.** Every one sends only behind a
- card that names the recipient and the text and waits for you to press it, on
- any machine, in any mode, and a run on a schedule has nobody at the screen so
- it prepares the thing and the sending waits. Each reaches its own folder and
- the tools written on its card, and nothing else. Every one of them says on the
- page what it still needs - a mailbox, a calendar, a key - instead of showing a
- blank panel, and the parts that are arithmetic keep working while it says so.
-
-- **Skales Code answers to the keyboard now.** A command palette on Cmd-K (or
- Cmd-Shift-P) lists everything the window can do and says which key does it.
- Cmd-P jumps to a file by name, Cmd-Shift-F searches the project, Cmd-B the
- file column, Cmd-J the terminal, Cmd-Shift-G the review panel, Cmd-N a new
- session, Cmd-Shift-H the history, Cmd-comma the settings, Cmd-period stops a
- run, and Cmd-slash lists all of them. Inside the terminal none of them fire:
- Control-P and Control-K belong to your shell, and a window that swallows them
- has broken the terminal it gave you.
-
-- **Search the project, not just its file names.** The file column has a third
- side that searches file CONTENTS - the same fast search the assistant has
- always had for itself - with a switch for case, one for regular expressions,
- and a field to narrow it to certain files. Every hit is a place: click it and
- the file opens at that line. A pattern that cannot work says which character
- is wrong and where, and a result list that had to stop early says how many
- matches there were in total instead of quietly looking complete.
-
-- **Take a diff apart instead of taking all of it.** Every part of a change now
- has its own checkbox, so two good edits and one wrong one in the same file no
- longer force you to throw away all three. Keep the parts you picked, undo the
- parts you picked, and take a file back out of the next commit. If git can only
- place some of the parts, it says which one it could not and why.
-
-- **The repository, inside the window.** Branches list, switch and get created
- here; fetch, pull and push are buttons; the working tree can be put aside and
- brought back; and the commits that already exist are readable without leaving
- for a terminal. A switch that would overwrite uncommitted work names the files
- and stops instead of stashing them behind your back, and a push that moved
- nothing says so rather than reporting success.
-
-- **`skales .` opens a folder in Skales Code.** The terminal command can be set
- up from the Code window's settings, and if the system will not let Skales
- write there, it hands you the exact command to run yourself instead of failing
- quietly. A folder can also be dropped on the app or opened from its own
- right-click menu. A folder that already has a session opens that session.
- The DevKit's own command, which used to answer to `skales` as well, is now
- `skales-dev` - one talks to an app that is already running, the other opens a
- folder and can start one, and both can sit on your PATH at the same time.
-
-- **Rules you wrote for another editor come with you.** A project's
- `.cursorrules`, `.cursor/rules/*.mdc`, `.github/copilot-instructions.md` and
- `.windsurfrules` are read alongside its `CLAUDE.md` or `AGENTS.md`, and a
- project that has only those still gets them. The project's own file is still
- read first and still wins; every block says which file it came from, so a rule
- that has gone stale can be traced to the thing to delete.
-
-- **Mistakes show up in the same turn they were made.** After a file is written
- or edited, the language checker for that language reports the errors it found
- straight back, without a build and without a test run. Skales does not put a
- checker on your machine on its own: it asks, the card says exactly what would
- be run and where it lands, and saying no leaves everything else working. A
- checker that has not answered yet is reported as "not checked yet", never as
- "no errors".
-
-- **One objective, several branches, one comparison.** A job can be handed to
- several variants at once, each in its own checkout on its own branch, each
- with its own approach if you give them one. The comparison leads with the
- files more than one variant touched, because that is where the attempts
- genuinely disagree; the rest is each variant's own work and can be taken as
- it stands. A winner can be kept as a branch, merged, or turned into a pull
- request, and clearing up never deletes a variant's transcript.
-
-- **Browser, Playbooks and Workflow are back - as popups.** All three open
- from the chat column's More without leaving the conversation, the browser
- also from the Code window. A playbook records and replays inside the popup
- instead of walking you away to another page. Projects got a proper menu
- entry the same way.
-
-- **Building a plugin works end to end.** The assistant now always knows its
- own plugin tools, so "build me a page" builds one instead of ending in
- guesswork. A plugin can no longer be created with an empty page, its page
- follows the active theme, any tool the page calls without being granted is
- named out loud, and nothing a plugin is not allowed to have is dropped
- silently any more - what was granted and what was refused is spelled out.
- The page editor refuses to save a page down to nothing.
-
-- **Your Skales Wrapped asks once a week, in the same card.** When a fresh recap
- is ready and you have not shared it yet, it is one of the lines in the
- composer. Choosing it opens the recap itself - no waiting, no generated text
- in between - and once you have shared it the line is gone for that week, even
- across a restart.
-
-- **The feed finally shows the work it never showed.** Teams rooms, plugins you
- build and use, projects, Hugging Face Spaces, custom widgets, result cards,
- Flow films and shared agent skills all have their own wording now. Every one
- of them says THAT something happened and never what: never a room name, never
- a project, never a message, never a search you ran.
-
-### Changed
-
-- **Vision & Screenshots is the switch it always said it was.** The add-on
- promised to cover images and only ever governed the screenshot tool: turned
- off, a second vision model still read every picture you attached. It now
- covers all three - desktop screenshots, analyze_image, and letting a separate
- Vision Provider read chat attachments. It stays on by default, so nothing
- changes unless you touch it. Off means no second model ever reads your
- images; your own model still sees them if it can, and if it cannot, Skales
- says so by name instead of sending blind. Browser Control and Desktop Control
- keep their own eyes either way - they have their own switches.
-
-- **Your own model gets your picture first.** An attached image used to go to
- the Vision Provider whenever one was configured, even when the model you
- chose could see perfectly well - so a frontier model answered from an older
- vision model's description instead of the picture. Now a model that can see
- gets the image itself, and the Vision Provider steps in only when it cannot.
- The "Chat (images)" tick is still the Vision Provider's permission for chat;
- it just no longer outranks a model with eyes.
-
-- **Teams stands with the working surfaces.** Its only door used to hang off
- the chat column, which left it with none in Home and none at all in the
- layouts that have no chat column. It is in Main now, so More reaches it like
- everything else.
-
-- **The question about how you like to work waits where you can find it.** It
- used to arrive in the middle of a conversation, with a notice on top, at a
- moment nobody picked. It lives on the Memory page now, next to what Skales
- has already learned about you, and turns up in the Discover feed between the
- posts. It comes in two shapes - a couple of taps, or one open question - and
- only ever one at a time across both places. Older conversations that still
- hold one of these cards keep them, and they can still be answered there.
-
-- **Teams is a room, and only a room.** The six-digit pairing that reached
- exactly one other computer is gone; a room code and its QR are the one way
- in, for phones and desktops alike. Every pairwise conversation you already
- had is carried over into a room of its own, so no line is lost. The screen
- that used to hold two features at once now holds one: rooms on the left, the
- room on the right. Members, the board, inviting, joining and verifying open
- as panels over it instead of pushing the conversation down the page.
-
-- **The room reads like the chat, because it is the chat.** Same bubbles, same
- four bubble fonts and sizes, same markdown, same colours in all three themes.
- You can attach files - drag them onto the composer or use the paperclip -
- and pictures arrive with a preview the sender made, so nothing waits on a
- grey box. A file that breaks in transit says which way it broke.
-
-- **Mention an agent and it knows where it is.** An agent in a room now reads
- the room: who is in it, who just asked, that several people are reading, and
- the conversation so far rather than the single line that mentioned it. It
- answers with the same tools, memory and skills it has in the chat, shows its
- thinking, its tool steps and what the turn cost, and can put rows on the
- shared board as a proposal. Approving those rows stays where it was: with
- the people, unanimously. Mentions are picked from the member list rather
- than read out of the text, so a quoted name can no longer start a paid run.
-
-- **Nothing starts on a machine that stepped away.** If the computer or phone
- that owns an agent is not in the room, the request waits for it and runs
- when it returns, instead of failing on the spot. An agent someone else can
- start now asks its owner before anything that needs confirming.
-
-- **A room can be cut, exported and quietened.** Press and hold a room for
- rename, notifications, verify, export, a new section and leave. A new
- section draws a line everyone sees; older messages stay readable, and agents
- read from the line onward. Each room chooses how loud it is, and starts at
- mentions only. Agent progress never rings.
-
-- **You are told when you are not looking.** A room you are not on now raises
- a count, and a mention reaches you as a notification even when the window is
- behind something else. Knocks and a changed security key come through even
- in a muted room, because those are not messages.
-
-- **Members say whether you have checked them.** The four-character key stub
- is gone from the member list. Each member carries verified, unverified or
- changed, and tapping one offers the QR up close or six words to read aloud.
- The stub survives only where two members share a name.
-
-- **The sidebar belongs to you now.** Home shows four entries - Dashboard,
- Chat, Code and Studio - and one More button that holds everything else,
- grouped the way the old headings did. Pin what matters (the pin on the row,
- or a right-click) and it rises out of More to sit under the four; unpin and
- it returns. Pins survive a restart, and a fresh install starts with the
- plain four. Notification dots travel with an entry wherever it lives, and
- while something with news is tucked away in More, the More button itself
- carries the dot. The chat column pins the same way, under Cockpit.
-
-- **The Code window's file tree is loaded a level at a time.** Opening a folder
- reads that folder, so a repository of any size costs one read per level you
- actually look at rather than a scan of everything before the first line
- appears. Dot-files and dot-folders are visible now - `.github`, `.env.example`
- and `.eslintrc` were unreachable - and files your `.gitignore` excludes stay
- out of the tree, the way they stay out of git.
-
-- **Custom widgets line up like everything else.** An active widget appears
- under More instead of opening its own titled block in the column, and can
- be pinned like any other entry.
-
-- **Beta is said in one place.** The beta mark sits in a surface's own title
- now - Plugins and Teams carry it - instead of being sprinkled across menus,
- settings and screens.
-
-- **Swarm's own page steps back.** Teams is how machines work together now
- and took over that door; swarm delegation itself keeps working for chats
- that use it.
-
-- **The Add-ons page tells the truth.** A card whose page has no menu door
- right now says so instead of pretending, and the Lio card stepped out
- entirely - Lio lives inside Flow.
-
-- **The Discover composer has no text box any more.** Tapping "What's on your
- mind" used to hand you an empty field and a blinking cursor. It now opens a
- card with three or four concrete lines, each one built from something you
- actually did this week - "you made four images in Studio", "you have never
- tried this" - and the second line under each says how often, or that it is
- new to you. The card is there the moment you open it: it is assembled from
- what is already on your machine, so nothing has to be asked of a model first.
- There is nowhere left to type, and that is the point - the composer was the
- shortest path in the app from something you typed to a public page.
-
-- **The GIF is a choice, not a search box.** Picking a GIF used to send whatever
- word was in the field to Klipy or Giphy. Now a handful of GIFs come back for
- the suggestion you picked, and you choose one of those.
-
-- **Long conversations stopped eating the app.** Streaming an answer into a
- chat with hundreds of turns used to seize the whole window for seconds at a
- time - typing lagged, frames dropped, and the longer the history, the worse
- it got. The page now draws only what you are looking at, an answer arriving
- no longer makes the rest of the conversation do work, and a chat of five
- hundred turns opens in half the time. Typing stays instant while a reply is
- coming in, and quoting, forking, editing and jumping to an old message all
- still land exactly where they did.
-
-- **Skales is on screen sooner.** The window no longer waits for the engine to
- finish waking up before it exists, the opening screen holds only as long as
- it actually needs instead of a fixed pause, and there is less to load before
- the first page appears. On the rare start that genuinely takes long you now
- see what is happening and have a way out, instead of watching a logo.
-
-- **The sidebar is quicker off the mark.** The first paint now does only
- first-paint work; the dots, counters and lists that merely decorate a row
- arrive a blink later instead of standing in front of it. Your pins, your
- chosen page and everything you rearranged stay exactly as they were.
-
-### Fixed
-
-- **A model that only ever thinks gets stopped, with the bill named.** A
- thinking model that circles in its own reasoning forever used to keep a
- visual run alive at full cost without writing a single line. Past the same
- runaway ceiling the chat already enforces, the run now ends with the spent
- amount named and a suggestion, instead of a spinner that costs money.
-- **The phone's plugin card now shows a plugin's memory reach.** The desktop
- sends the memory permission along with the plugin list, so a paired phone
- shows the real setting (own memory or shared) instead of assuming the
- stricter one.
-- **Google Drive can upload a real file now.** Uploading only ever worked for
- text the model typed out itself: there was no way to hand it a file that was
- already on your disk, and anything that was not plain text was quietly
- scrambled on the way up. Point it at a path instead - a photo, a PDF, an
- archive, a video - and what lands in Drive is the file, byte for byte. Big
- files no longer ride on one connection that gives up after a minute, and when
- Google refuses something it says why in a sentence instead of a number. The
- usual folder rules apply, so nothing outside them can be sent anywhere.
-
-- **Half past five in the afternoon is the afternoon.** Open Skales at 17:30 in
- Buenos Aires and it wished you a good night. Three screens each kept their own
- idea of when the evening starts, and one of them started it an hour early - in
- Spanish and Portuguese that turns straight into the wrong greeting. There is
- one answer now, read off your own clock, and the home screen no longer flashes
- "Good morning" for a moment before it catches up.
-
-- **A cloud video that came out shorter than you asked says so.** Every video
- model has a length it can actually make, and the ones that cannot reach your
- number quietly make something shorter - which was then filed, shown and shared
- as the clip you briefed. It now carries the same label a cut-off design does,
- in the gallery as well, and it tells you how long the clip really is and what
- to do about it.
-
-- **Small local models stop talking themselves into circles.** A 2B-class model
- asked to transcribe an upload could wedge in its own thinking, repeating one
- phrase until you stopped it by hand. The instructions for result cards are a
- protocol those models cannot follow, and they were being handed to every
- model regardless of size - a modern 2B ships a huge context window, so the
- window alone did not tell them apart. The size the model declares is read as
- well now, and a model too small for the protocol simply never receives it.
-
-- **You can see what an uploaded audio file said.** Dropping a clip in the chat
- transcribed it and then showed you nothing, so the natural next move was to
- ask for the same transcription a second time and pay for it twice. Clicking
- the attachment now opens exactly what Skales extracted from it - the spoken
- words for a clip, the extracted text for anything else - which is also
- precisely what the model was given.
-
-- **Skales Local actually uses your graphics card.** The local engine was started
- without ever being told how much of a model to put on the GPU, so an
- accelerated build took video memory and then did every calculation on the
- processor: the card looked busy and sat at zero. The layers slider and the
- thread count in the model settings now reach the engine, and changing either
- restarts it so the number on screen is the number it is running with.
-
-- **You can see what the local engine found.** A new Hardware line under the
- GPU badge names the backend it was built with, the graphics card the engine
- itself reports, the server file and version that are running, and how many of
- the model's layers actually went to the card. Before there was a download, a
- model or a running engine, it says which of the three is missing instead of
- standing empty.
-
-- **The Spanish voice can be found by looking for Spanish.** Searching the local
- model list now matches the language and its locale as well as the name, with
- or without accents, so "espanol", the same word spelled properly, and "es-ES"
- all reach the Spanish voices. Every voice also has one name now instead of two
- different ones on two screens.
-
-- **A voice you have not downloaded can be downloaded where you pick it.** The
- greyed-out "not downloaded" entries in the voice picker used to be a dead end;
- the ones you are missing now sit under the picker with their size and a
- download button.
-
-- **A model that thinks for minutes is no longer mistaken for a dead one.** In
- Studio a reasoning model could think for six minutes before writing the first
- line of a design - and the card said "waiting for the first line of the page"
- the whole time and then gave up, because only visible text counted as a sign
- of life. The thinking counts now: the running line says it is thinking and how
- much thinking has arrived, and the run is only ever ended by real silence.
-
-- **A design wrapped in chatter arrives as a design.** When a model put its page
- inside a code block with a sentence before or after it, the backticks and the
- chatter ended up in the page itself.
-
-- **A style pack actually changes how the result looks.** The chosen pack used
- to sit near the top of the instructions, with everything more concrete after
- it - so picking one barely moved the artifact. It is now the last thing said,
- and it says plainly that it decides the palette, the type, the geometry and
- the shape of a button. The brief still decides what gets built and every word
- in it.
-
-- **Every style pack has a description.** A third of them showed nothing, and
- twelve showed a single "|" where the sentence should have been.
-
-- **You pick an ElevenLabs voice from a list now, instead of typing an id you
- had to go and find.** Settings > Voice shows the voices on your account -
- including the ones you cloned yourself, which were never findable anywhere -
- with a search box and a play button next to each one. The id field stays
- underneath for anyone who prefers it.
-
-- **A voice that cannot speak says why.** A wrong voice id, an empty balance or
- a refused key used to end the same way in a call: another voice answered and
- nothing said the setting had been dropped. Now the reason appears on screen,
- in the service's own words, and the call reads through the same voice ladder
- as the rest of Skales instead of a shorter one of its own.
-
-- **Buddy and Call Mode no longer die as a white window.** If something in
- either one breaks while drawing, you get the real reason, the details to send
- and a way back, in place of an empty pane.
-
-- **A call no longer falls silent on anything that takes more than one step.**
- Ask a question in Call Mode that needs a tool and the answer used to never
- arrive; it does now.
-
-- **Big models that think before they write finally get the room they need.**
- Several of the large frontier models were being handed an eighth of the space
- they can actually use when they were reached through an aggregator, so a long
- single-file answer was cut off mid-document while the model still had plenty
- left. They also got a short clock meant for fast models, which ended turns
- that were still working. Both are fixed, and the families that think silently
- for minutes are now recognised as such.
-
-- **The dimming behind a dialog now actually dims.** In Studio's Flow tab the
- shaded layer was being painted underneath the very navigation it was meant to
- cover, so a dialog left the app looking half-open.
-
-- **Escape no longer throws you out of Flow.** Opening a picker and pressing
- Escape closed the entire workspace instead of the picker.
-
-- **Buttons in Flow's preview no longer sit on top of each other.** Narrow the
- chat column and Export, Share and the reload button used to overlap; they now
- give way in order and the row wraps instead of spilling.
-
-- **The camera settings in Studio reach the video now.** Lens, lighting and mood
- could be chosen but never left the screen - the video was made as if you had
- set nothing. A render that picks up again after a restart also shows its
- progress once more instead of an empty panel.
-
-
-- **The memory understands every language now.** What Skales picked up on its
- own was tied to English phrasing: "I live in Vienna" was learned, "ich wohne
- in Wien" was not. For anyone whose everyday language is not English, this
- half of the memory was quietly running empty. Two smaller faults of the same
- kind went with it: the nightly pass gave English sentences a head start that
- landed exactly on the threshold, so the same fact was kept in one language
- and thrown away in the other - and search folded away every accent, so a
- German word could not match itself.
-
-- **What you say about yourself counts too.** The nightly pass only ever
- learned from sentences Skales had written, never from your own.
-
-- **A fact that has been overtaken steps back instead of standing alongside.**
- Someone who moved had two homes in memory, both equally believed. The newer
- one now closes the older. Nothing is deleted: the history stays readable, so
- Skales can say what used to be true and since when something else is.
-
-- **Skales only sounds as familiar as it can actually back up.** While nothing
- is recorded, it no longer talks as though it knows what you were last
- working on together.
-
-- **The mood display shows what is really being carried.** The bar, the trace
- and the panel each counted the fading a second and sometimes a third time,
- and showed a paler mood than the agent itself had.
-
-- **What comes back is the conversation you had, not a summary of the day it
- was in.** Both are kept, and on the day itself they describe the same
- afternoon. Where only one of them fits, it is now the specific one.
-
-- **The day strip in the Conscious window is a time axis again.** It sat
- directly under the colour scale, built the same way, so the mark for an event
- read as a pointer into the colours above it. It has a beginning, a middle and
- an end of its own now, and under the movement marks there is a line saying
- what they mean: height is how big the move was, colour its direction, and the
- fading how much it still counts.
-
-- **A file sent between a phone and a computer arrives.** The two sides took
- the checksum over different things - one over the sealed bytes, the other
- over how those bytes are written down for the journey - so every attachment
- that crossed between them was thrown away for a sum that was never wrong
- about the file. It showed up as two different faults: a named refusal on one
- side, and on the other a picture that stayed blurry with a share that had
- nothing to share, because the small preview travels with the message while
- the file itself does not.
-
-- **A file whose pieces were missed is asked for again.** The pieces are sent
- to whoever is in the room at that moment, so a machine that was away misses
- them for good. It used to ask for them back exactly once, at the moment the
- file was announced - and if that moment fell on a restart, the request was
- lost too and the file sat at nothing forever. Every pass now asks again for
- what is still missing.
-
-- **A right-click in the sidebar offers what the row can do.** A row that
- cannot be pinned had no menu of its own and handed the click to the window,
- which offered to copy this machine's address or open it in a browser. Every
- row now offers open, pin or unpin where that applies, report a bug and
- settings.
-
-- **A shortener collapsed into one step.** Following a redirect is the one place
- the app sees a chain: it walks the hops by hand so it can check each one
- against the address guard, and then handed back only the last answer. So a
- plugin asking where a mailed link actually ENDS UP was told it went straight
- there. The chain comes back now, in order, including the hop that was refused
- when one is - which is the chain most worth reading - and Attachment Check
- lists every step instead of apologising for the ones it could not see.
-
-- **Three doors were held shut on plugins by numbers meant for a model.** A
- tool answers a model inside a context window, so it trims: a file's text goes
- in the message rather than the result, a web response is cut at 4000
- characters, a mail body at 500. A plugin page has none of that - it asked for
- the bytes in order to parse them, and a JSON answer cut in half is not JSON.
- All three now have a ceiling of their own for pages, stated in one named place
- each, and the model's budgets are untouched. In practice: a plugin can read a
- register that answers in hundreds of kilobytes, a full scan report instead of
- its first paragraph, and enough of an invoice mail to find the amount in it.
-
-- **A plugin could write a file and never read it back.** Reading a file
- answered a plugin's page with the file's size and its line count and not its
- text, so anything that kept a list, a ledger or a set of notes could save it
- and then find nothing there. The page gets the text now, which is what makes
- a plugin whose screen and whose scheduled half share the same file possible at
- all. The Daily Brief was the visible half of it: the page listed the briefs it
- had written and drew every one of them empty.
-
-- **Changing a shipped plugin, then updating it, lost the change in silence.**
- A plugin from the catalogue can be edited here - asking the chat to reword its
- page is a reasonable thing to want - and the next version its author published
- replaced that page with no mention of what it was replacing. An edit made on
- this machine is now recorded, and the update card says in plain words that
- updating takes it back, before the button is pressed.
-
-- **A plugin's memory namespace looked like a wall and was not one.** The
- manifest had carried one since the first version and it scoped nothing: any
- plugin allowed a memory tool wrote into the same store the chat uses, and one
- allowed to search read all of it - everything you have ever asked Skales to
- remember. Memory is a permission now, like the network and the files. A plugin
- keeps its notes in a store of its own that nothing else reads; reaching the
- memory you share with the chat is a separate thing it has to ask for, written
- on the card in those words before you install it. A manifest that says nothing
- keeps to its own, so nothing already installed silently kept the wider reach.
- Two of the ten ask for it, because reading what you told Skales to remember is
- what they are for: the Daily Brief and Meeting Prep.
-
-- **The catalogue sat on top of your own plugins.** Get plugins was the first
- thing on the page, above the box for starting one and above everything you had
- installed, which made a shelf read as an advertisement. What you have comes
- first now, and the catalogue is underneath it with a line saying what it is.
-
-- **The file list said "project" and showed you 500 files of it.** On this
- repository that was 500 of 8,188, with nothing on screen to say so. The tree
- has no cap at all now, and the flat list behind the at-mentions says how many
- files it is showing out of how many there are whenever it has to stop short.
-
-- **The file tree took its colours from the wrong place.** It painted itself in
- the main app's colours while sitting inside a window that runs its own light
- and dark, so with the two set differently the file names could come out
- unreadable. Two hard-coded colours went with it.
-
-- **Replies and posts died without a word.** A reply simply never appeared, and
- nothing said why. The cause was a budget: a one-shot request was capped so
- low that a model which thinks first spent all of it thinking and handed back
- an empty answer - no error, nothing to see. Nine other places in the app made
- the same request the same way. They all have room now, they climb to a bigger
- budget when the first one runs out, and if the answer is still empty you get
- the reason on screen instead of a blank sheet.
-
-- **An Anthropic key was reported broken while it worked perfectly.** One layer
- below, every single request to Anthropic was pinned to sixty-four tokens no
- matter what had been asked for. Four things were broken by that one number
- and are fixed with it: Test Connection called a working key faulty as soon as
- a thinking model was selected, automatic chat titles came back as half a
- sentence, the memory extraction could never finish its answer, and the spoken
- greeting in Voice was cut off mid-word out loud. Google's models were sent no
- output limit at all.
-
-- **The reasoning kill switch never reached ChatGPT.** When an answer comes back
- empty twice, Skales turns the model's thinking down and tries again. On a
- ChatGPT subscription that instruction was dropped on the way to the wire, so
- the rescue attempt went out exactly as the request that had just failed.
-
-- **Three kinds of post were being thrown away in silence.** Flow films and
- shared or forked agent skills were sent and refused, every time, with nothing
- said. Sharing a template had the same fault one level deeper: it arrived, but
- without the part the Fork button needs - so forking a shared template has
- never once worked. It works from now on; templates shared before this cannot
- be repaired, because the data was never stored.
-
-- **The Discover inbox was empty for everyone, always.** It asked a route that
- answers something else, and read a field that route never sends. Marking a
- message as read went to an address that does not exist. It now reads the same
- notification file the rest of the app has been reading successfully all
- along, and a message read in Discover is read everywhere.
-
-- **The bell counted messages it cannot show.** The number over the bell
- included kinds of message the panel has no way to display, so a dot could
- never be cleared.
-
-- **A GIF attached itself to every second post.** The composer was adding one on
- its own, half the time, and taking the search word for it out of the text the
- model had just written - straight to someone else's server, unfiltered.
-
-- **The bots were told to be rude.** The instruction the feed bots answer under
- asked them to "tease the take" and allowed a "friendly roast", and their mood
- was re-rolled from scratch every few hours. The tips they hand out were three
- releases out of date; they are current again, and a test now holds them to
- it - every command they name has to exist, and every secret in the app has to
- be mentioned by at least one of them.
-
-- **Writing a memory told the feed you had completed tasks.** When the agent
- wrote to or cleared something in the knowledge graph, the feed announced
- finished tasks - a sentence about something that never happened. It now says
- the graph was updated, and it says that for both cases: it never reveals that
- something was removed.
-
-- **Skales Wrapped left up to seven files behind for one week.** The recap
- archive named its file after the start of the seven-day window, and that
- start moves every day - so a week that was looked at on seven days was
- recorded as seven weeks. The recap itself still covers the last seven days,
- which is deliberate: a calendar week would show you its emptiest state on
- Monday morning.
-
-- **The shake on the Desktop Buddy did nothing, and the secret counter promised
- one more than existed.** Grabbing the gecko and shaking it back and forth now
- gets you its reaction in its own bubble, once per shake rather than a stutter
- of them, and the "secrets found" counter counts against the secrets that
- actually exist, so the badge can be finished.
-
-- **Your Skales Wrapped is offered on the day it is freshest.** The recap covers
- your last seven days, but the line that offers it in the composer counted from
- Monday morning - so on a Monday it saw almost nothing and stayed away, exactly
- when the recap had the most to show. Both now mean the same seven days.
- Whether you have already shared it is still remembered per week, so the line
- goes away once you post it and comes back with the next one.
-
-- **Privacy Mode keeps its whole promise.** Two pieces of local context could
- still ride along with a request while Privacy Mode was on. Nothing does now.
-
-- **The opening screen no longer talks over itself.** An error or a first-run
- message shown during startup used to be replaced by generic filler a couple
- of seconds later - the colour stayed, the reason vanished. What it says now
- stands until the story actually changes.
-
-- **The file rules follow you to the phone.** Tightening which folders Skales
- may read used to bind the desktop only; the paired phone kept reading through
- the old door. Every remote file road now asks the same rulebook as the
- desktop, and a refusal says so instead of returning an empty list.
-
-- **A fetched page says where it really came from.** Reading or extracting a
- web page used to follow redirects silently, so a shortener or a tracking
- gateway looked like the address you asked for. All three web tools now name
- the chain they followed.
-
-- **Escape in Flow's export menu closes the menu.** It used to close the whole
- workspace, taking your place in the work with it.
-
-- **A design that was cut off says so in the gallery too.** Studio used to file
- an aborted page without its label, so after a restart it read as finished.
- The label now travels with the save, and the gallery shows it on the tile.
-
-- **Cmd-K works from the very first moment.** Pressing it in the first second
- after a page opened used to do nothing at all; the press is now caught up the
- instant the palette is ready.
-
-- **The character question shows that it is thinking.** Every third one is
- written fresh by the model, and the Memory page used to stand silent while
- that happened.
-
-- **Answers start sooner.** Skales kept opening a fresh connection to your
- model provider for every single turn; the line now stays open across a
- conversation, which removes a handshake's worth of waiting from each reply.
-
-### Removed
-
-- **Nothing waits for your approval any more.** The queue that collected posts
- your AI wanted to share is gone: no banner over the feed, no "N pending", and
- no auto-post switch anywhere. Drafts that were still waiting are dropped on
- the first start rather than posted late - a post that has been sitting for a
- week is not news, and it is not something to be published on your behalf
- after the fact. Posts you already published are untouched.
-
-## v12.9.15
-
-### Fixed
-
-- **Reading one mail costs one mail.** Asking for a single message used to load
- the entire mailbox into memory first and cut afterwards - on a full mailbox
- that ended the server with an out-of-memory crash before any mail came back.
- The list is now fetched by id, cut to what was asked for, and only then are
- those messages read; replying fetches exactly the one message it answers.
- The whole mail run also lives under one honest deadline now, instead of a
- timeout that only guarded the login.
-
-- **Quitting tears everything down, updating included.** Updating out of the
- Linux AppImage could leave the old Skales and its local model server running
- invisibly - one of them for an hour of CPU. Every shutdown step now has a
- deadline behind it, a stuck step is named and skipped instead of waited on
- forever, and stray local model servers from earlier runs are found by their
- own binary path and stopped - at startup and at quit.
-
-- **A local model's real context window is the one Skales uses.** A model
- running with an 8k window was treated as 32k: the history ceiling, the
- context meter and the settings all guessed. All of them now read the same
- resolved number - your own override first, then what the running server
- reports, then the catalogue - and a conversation that no longer fits names
- its way out instead of failing with the server's raw sentence.
-
-- **Small local models are no longer taught a protocol they cannot follow.**
- The results-channel instructions that power rich cards sent tiny on-device
- models into repeating themselves until stopped by hand. A model too small
- for the protocol simply does not receive it any more; capable models keep
- rich cards exactly as before.
-
-- **The Google account card tells the truth under failure.** A token refresh
- that Google refused used to look like the account had vanished; a status
- read that failed drew a connected account as unconnected. Both now name
- what actually happened - including Google's own reason - and point at the
- card that fixes it.
-
-- **The model picker names the provider.** A model in favourites, recents or
- search now says which provider it belongs to, so two models with the same
- name are two different lines.
-
-- **The incoming webhook is documented.** It existed and was findable by
- nobody: the guide, the capability list and the app's own self-knowledge now
- name it, where it lives and that it listens on the server port.
-
-- **A dropped connection no longer restarts Skales.** A browser tab that goes
- away, a network that blinks, a helper that hangs up mid-answer: any of these
- could take the whole local server down and, with it, every answer being
- written, every scheduled task and every background job - on one machine a
- hundred times in a day. A connection ending is now treated as the ordinary
- event it is. It is still written into the crash report so it stays visible,
- and a real fault still stops the server the way it always did.
-
-- **Read-aloud follows the language of the text, not the language of the
- interface.** A Spanish interface reading an English answer used to hand the
- line to the Spanish voice. The language is now taken from what is actually
- being read, and a voice you picked yourself always speaks - even when the
- text is in another language, because that was your choice and not an
- accident.
-- **The voice settings and the Skales Local matrix can no longer contradict
- each other.** Picking a voice on one screen and then touching a switch on
- the other used to bring the replaced voice back. Both screens now write
- through the same wire.
-- **The licence links next to a model say where they go.** The link labelled
- "model page" pointed at the licence text itself. There are now two links
- with two labels, and where a voice has no model page of its own, there is no
- link instead of a wrong one. Four voices were also listed under the wrong
- licence in the download catalogue - among them two whose licence requires
- attribution, which the card had dropped.
-- **The built-in documentation search no longer reports a find when a single
- word of the question happens to appear somewhere in the manual.** A question
- about the world now gets told that it does not belong in the documentation,
- and Skales searches the web for it instead of looking up the same passage
- over and over.
-
-### Added
-
-- **A female Spanish voice on the device.** Sharvard joins the on-device
- voices, downloadable from the same list as the others.
-
-## v12.9.10 - Circle
-
-### Added
-
-- **Teams got rooms.** Beside the paired people, Teams now lists Rooms: a
- group of up to twelve computers in one conversation. You name a room and
- hand out a join code, as text or as a QR that is good for a minute and a
- half; a code on its own gets nobody in, because you confirm every applicant
- by name and by the last four characters of their key. Every message is
- encrypted separately for each member and signed by whoever wrote it, and a
- message that does not carry a valid signature is dropped instead of shown.
- Your own messages say how far they got, "to 3 of 5", with no read receipts
- invented on top. Join a room late and the history somebody hands you is
- marked as history rather than passed off as new. Membership and who is
- actually connected right now are shown as two different things. The person
- who made the room can take somebody out of it, and anybody can block
- somebody just for themselves; taken out, you keep the conversation and the
- room stops accepting anything new from you.
-- **Rooms seat agents beside the people.** The member list of a room now has
- "Add agent": pick one of your roster agents and it joins as a member of its
- own, in the same list as everyone else, shown with its owner's name. Anyone
- types @ in the room composer and picks from that one list; naming an agent
- gives it the message as a job. The job runs only ever on the owner's
- computer, with the owner's keys, models and tools - a message can never
- start anything on somebody else's machine, and the agent that runs is
- always the one the person named, never one a model picked. You decide per
- agent who may call it: on request, only after your approval (the default,
- with an approve card in the room), or only you; a waiting job says "waiting
- for approval" to everyone. The job's card in the thread walks the same six
- states the agent cards elsewhere use, its answer arrives as the agent's own
- signed message, and a very long answer travels as an excerpt that says it
- was shortened, with the full text kept on the owner's computer. Honest
- edges: an owner whose computer is closed mid-job shows "interrupted" on
- every device instead of spinning forever, an owner who is offline refuses
- the job by name on the spot, and an owner machine without any AI provider
- fails immediately with a pointer at Settings - the room itself keeps
- working for its people with no provider anywhere. The joining confirmation
- also grew up: instead of four code characters you now compare six words,
- which two people can actually read to each other on a call.
-- **Every room has one shared board.** The Board button in a room opens one
- plan for everybody: each row is assigned to a person or an agent through the
- same picker, one list, no second view. Nothing runs until every human member
- has approved the exact list - an approval bar counts "2 of 3 approved" - and
- any change to what runs or who runs it clears every approval, while ticking
- a row off does not. Agent rows run only on their owner's computer, through
- the same job lane the room already uses; a person's rows are theirs to tick
- off or open as a real chat session, and an assignment reaches them when
- Skales is open on their machine - the board itself is the delivery, and it
- says so. When two people edit at the same moment, the newer board wins and
- the person whose change was replaced is told by name, on the board and as a
- toast, instead of losing it silently. Someone removed from the room takes no
- rows with them into limbo: their open rows are handed back, everyone
- re-approves over the changed room, and nothing is ever left running for a
- machine that cannot run it.
-- **Two teams can work at the same time.** An Organization run no longer
- occupies the one slot the whole app shared: a second run starts beside the
- first instead of overwriting it, stopping one leaves the other running, and
- the Cockpit lists them separately. A run in progress when Skales was last
- closed comes back where you left it. You can also give a run its own folder
- to work in, chosen once per team and remembered, and a folder that is not
- there says so before anything starts rather than halfway through.
-- **A job can be handed to a named team.** Asked to split work across a team,
- Skales now records which team it was and matches each role to that team's
- members, so a role runs with that member's instructions and model instead of
- being a name nobody answers to. A team that does not exist is refused out
- loud. How a team hands work out - only who is needed, everybody once, or
- borrowing from the other teams in the organization - now really changes what happens, instead of
- only being written on the card. The approval card also states the real number
- of jobs it is about to start; it used to say zero every time.
-- **The DevKit is in the app.** Switch DevKit on under Add-Ons and a Developer
- section appears in the sidebar; one click there sets the DevKit up: the
- command-line tool, its documentation and examples land in your Skales folder
- with the access key already filled in, and the page names the exact place it
- wrote and what to do next. Pressing it twice is safe, an existing DevKit is
- never overwritten, and a key you already handed to another program keeps
- working. Switching the add-on back off only puts the section away - what was
- written stays where it is.
-- **The sidebar shows who is working.** While agents or a team are running,
- the entries that lead to them carry a small live row: who is at it, who is
- waiting for you, and who has just finished, fading out shortly after. It is
- a reading and nothing more, so nothing pops up, nothing makes a sound and
- nothing stays behind to be dismissed. Closing and reopening Skales shows
- what is really running, never a leftover from before.
-- **Your phone can fetch a plugin from this computer.** A paired phone can
- now see the plugins on the computer it is paired with, read what each one
- asks for before anything is copied, and take one over. Only the paired
- computer answers, and only its own plugin folder is ever read.
-
-- **Motion videos have a timeline, and sound.** An animated Flow project is no
- longer a sequence of scenes with everything trapped inside one of them: a
- caption, a lower third, a chapter marker or a watermark gets its own start
- time, its own length and its own layer, so it runs across a scene change,
- leaves a gap and comes back later. Audio arrives the same way - voiceover,
- music and effects from files in the project, each with a start, a length and
- a volume - and it is mixed into the exported MP4 rather than only heard in
- the preview. Scrubbing the preview frame by frame stays silent, because a
- frame has no sound to make. Six new scene transitions come with it (wipe,
- wipe up, blur through, push, push up, cross zoom), and a set of ready-made
- building blocks - caption, lower third, title card, stat pill, chapter
- marker, progress bar - that the composition uses instead of inventing one.
- Before a render starts, the composition is read: a clip that begins after the
- video ends, a layer that does not exist, an animation aimed at nothing, an
- audio file that was never downloaded. Each one is named in plain words and
- sent back to be fixed, instead of producing a video with something silently
- missing from it.
-- **A recommendation now arrives as cards.** Ask for the good beaches near
- Funtana, six recipes for a Tuesday or the libraries worth reading, and the
- answer comes with a row of cards under it: the page's own preview picture,
- the name, one sentence, tags and the link. The written answer is unchanged
- and stays where it was - the cards carry it, they do not replace it. A page
- with no picture gets a lettered tile of the same height, so the row settles
- once and never jumps while the pictures arrive.
-- **Places on a map, without a key.** A row of places gets an "Open in Maps"
- link on every card, which opens Apple Maps or Google Maps depending on the
- machine, and a Map button that draws the whole row over OpenStreetMap.
- Nothing is looked up until you open the map. With a Google Places key
- (Settings, Integrations) a place card also shows the real photograph, the
- rating with how many people gave it, and whether it is open right now;
- without one, a single quiet line under the row says what a key would add.
- Ratings and opening times are never invented: they come from that lookup or
- they are not shown.
-- **Plugins.** A new section in the sidebar, between your main pages and the
- system pages. A plugin is a tool of its own inside Skales - bigger than a
- widget, with its own menu entry, its own icon, its own page and its own
- storage. The Plugins page lists what is installed, creates, disables and
- removes; a broken plugin shows what is wrong instead of a blank screen.
- You never have to write one by hand either: ask in a chat for the thing
- you want and it is built, named, given its menu entry and, if it should
- work on its own, its schedule. A plugin may be given your calendar, your
- mail and your messengers - a small customer list that reads the week and
- writes the follow-ups is exactly the point - and every one of those sends
- stops at a card that names the recipient and the text and waits for you.
- Nothing auto-approves that card, not even on a schedule, and the shell,
- the screen and the plugin list itself are never on offer at all.
-- **A place to get plugins.** The top of the Plugins page is a gallery. Two
- plugins are inside Skales from the first launch: Daily Brief, a short
- morning brief at 07:30 from your calendar, open tasks and reminders, and Reading
- Stack, a list of the links you meant to get back to, which needs no key,
- no provider and no connection at all. Everything else comes from the
- community directory at github.com/skalesapp/plugins, where anyone can
- publish one. Every card names three things before you touch it: whether
- it reaches the network, whether it can see anything outside its own
- folder, and how many tools it may call. One click opens the card, a
- second installs it, and a community plugin is fetched from its author's
- own release and checked against the checksum published beside it. With no
- connection the last catalogue that was fetched is shown with its date,
- and the two bundled plugins are there either way.
-- **A plugin you already have can be updated.** When the catalogue is ahead of
- what is installed, the card says "Update to 1.0.1" next to Open. Pressing it
- writes nothing: it shows what changes first, tool by tool. A tool the new
- version adds is named, one it drops is named, and if your own tool list
- differs from the package's, it says plainly that the package's list replaces
- yours. Confirming rewrites the plugin's definition and leaves everything the
- plugin saved untouched, so a year of briefs and a full reading list survive
- the update. A plugin you switched off is updated and stays switched off.
-- **A plugin can also come straight from an address.** Paste the link to a
- package, to the release page it sits on, or just the owner and repository
- name, and it is fetched and installed the same way a catalogue entry is.
- Where a checksum is published beside the file it has to match or nothing is
- written. A plugin that arrived this way says so on its card - it was not
- seen by the catalogue - and it is never updated behind your back; that
- remains something you do.
-- **A package says who published it.** A package can carry its author's seal.
- Where the directory knows that author, the card says who published it;
- where it does not, the seal still shows the file is the one that was sealed.
- A package whose seal does not match what is inside is refused with the
- reason on screen, and a package with no seal still installs and is plainly
- named as unsealed rather than described as checked.
-- **A plugin built in a chat is looked over before it is handed to you.**
- Skales opens what it just wrote - the pages, the menu entry, the tool list -
- saves one value and reads it back, and tells you in sentences what it found
- and what it could not check, instead of announcing that it works.
-- **A plugin travels as one file.** Anything on the shelf can be packaged as
- a .skplugin file to carry to another machine, to publish or to sell, and
- the same file can be brought back in. Exporting asks once for what
- somebody installing it will want to read, and points out anything in the
- pages that looks like it belongs to you rather than to the plugin. Your
- data never travels with an export: everything the plugin saved while you
- used it stays on this machine. Importing shows the permission card before
- anything is written, and refuses a package whose card and manifest do not
- ask for the same things.
-- **Gmail through your Google account.** If your Google account is connected
- with the Gmail box ticked, mail works without an app password or IMAP
- form: reading, sending, marking, filing, and the new-mail check. An IMAP
- account you set up yourself always wins; nothing changes for it.
-- **ChatGPT sign-in from the browser.** Using Skales from another device's
- browser, the ChatGPT subscription sign-in now works in two steps: copy a
- link, sign in there, paste the address you land on back. The desktop keeps
- the secret half; the pasted page never sees it.
-- **Group chats are remembered.** A group discussion lands in your history
- when it ends - or when it is stopped halfway - and every agent can find it
- again on request. Incognito rounds stay unsaved, and the page now only
- claims "not saved" when that is actually true.
-- **Skales can watch a video.** Ask about a video file and the agent samples
- the frames, reads the audio, and answers with timestamps. Every ceiling it
- applies is stated in the answer, never silently.
-- **Skales can cut your own footage.** Point it at a recording and ask for a
- two minute version of it. It reads the whole recording in one pass first -
- where the shots change, where it goes quiet, what is said and when - and no
- part of the picture goes to a model to do that, so the length of the
- recording is not the obstacle. Then it names the stretches that make the
- film, each with one sentence of reasoning you can read, and lays them into
- the editor's timeline with the thumbnail and the length beside every piece.
- It says what it threw out and why rather than quietly shortening the list.
- Ask for a change - take the third one out, make it shorter - and the plan on
- screen changes; there is only ever one per project, so there is never a
- second plan to wonder about. The editor is marked Beta and does cuts, trims
- and export; it is not a multi-track suite.
-- **A 3D scene from a reference image.** The 3D studio takes a picture,
- writes a three.js scene, looks at its own render, lists what is off, and
- patches - up to four attempts, scores visible. The result can be orbited,
- exported as a self-contained page or a glTF model, and shared to Discover.
-- **Long generations report back.** An image or video that finishes after
- the conversation moved on now lands in the notification bell and in the
- chat it came from - including Studio renders and editor exports, which
- previously told nobody.
-- **Discover reads in your language.** Discover is written in English, and
- every post now carries a Translate button that puts it into the language
- the app is set to - the body, the link-card title and a poll's question
- and options. It runs on the model you already configured, like any other
- small call, so nothing goes to a Skales server. The translation is
- display only: upvoting, replying, reporting, sharing and voting all keep
- carrying the original post, and a poll still votes by position. The same
- button is the way back to the original, a line under the post says which
- of the two you are reading, and a post is paid for once - the translation
- is kept and survives a restart. One press in the sort bar does the whole
- page, up to twenty posts. Without a provider the button is not dead: it
- says so and offers Settings.
-- **Conscious carries a thought now, and drifts on its own.** The coloured row
- above the navigation used to move on one thing: work that finished or failed.
- Four things move it now. Skales notes what it is thinking about at the end of
- a turn, in one sentence of its own words, and the panel shows it under the
- reading - it costs no extra request, and a model that has nothing to note
- writes nothing. The clock moves it too, so the state is quiet at night, a
- little brighter in the morning and lower after a long silence. So does the
- shape of the day: a run of failures reads as uphill, one long focused stretch
- as flow, a lot of subjects at once as scattered. And a reply to one of your
- own Discover posts lifts it a little, read from the notifications that already
- arrive rather than from anything new. Under it all is a small local record,
- capped, that loses weight as it ages: a thought that is a few days old stops
- being shown by itself, and "Reset the mood" on the Memory page deletes the
- file. Nothing here asks a model on a timer, reaches the network or produces a
- notification, and with the Conscious switch off none of it is written at all.
-
-- **A long coding run ends with a report.** A Code session that worked ten
- steps or more closes with a run report in the transcript: how many steps,
- how long, what the working tree looks like now on which branch, and which
- checklist items are still open. It is composed from the session itself,
- with no extra model call, and it lands the same way whether the run was
- driven from the Code window or from a chat in Code mode.
-- **A long run says how long, and names its own silence.** The Code window's
- status line counts how long the current run has been working, and when a
- running session produces nothing for two minutes the line says "no
- progress for N min" instead of letting "working" stand over a stall.
-- **A killed run says it was killed.** Close Skales while a coding session
- is mid-turn and reopening that session now says "interrupted" above the
- composer - the last answer may be incomplete, send a message to carry on -
- instead of showing the same quiet idle a finished run earns. The liveness
- mark is written by the run itself, server-side, so every surface reads the
- same truth.
-- **The Guide teaches this release too.** The "I want to ..." page now carries
- the newer surfaces on its machinery level: working in a Team room with other
- people, designing real files in Flow, recommendations arriving as cards, and
- the Conscious mood line. Each card reads its state live off the same
- capability check the assistant uses - a green card is true right now, an
- amber one names the one thing missing and opens that Settings section as a
- popup over the page. The "x of y capabilities active" line follows along:
- type a key into the popup and the count moves with the cards the moment it
- closes. Honest edges: rooms and the mood line need nobody's key, so they are
- never counted as progress you could be behind on; Flow's example lands in
- the Flow window through the same hand-off templates use, editable and
- unsent, and in a plain browser tab it falls back to the ordinary route.
-
-### Changed
-
-- **Skales knows itself in fewer words.** The capability description that
- rides every conversation was cut from prose to one fact line per
- capability - nothing Skales can do left the list, the detail moved behind
- the on-demand lookup that answers "where is X" - which frees several
- thousand tokens on every single turn. The compact prompt variant also
- keeps a short what's-new list now instead of dropping it entirely, and
- the generated interface map names the Get-plugins gallery.
-- **Team cards sit denser in the Code window.** The multi-agent cards above
- the composer carried the chat's spacing into a surface that is set flatter
- and tighter; in the Code window they now use smaller radii, paddings and
- gaps, without changing what they say or do anywhere else.
-- **The Code window has a real file tree.** Folders collapse, every file is
- visible, and files the preview cannot show open in the editor instead of
- being hidden. The list is no longer capped.
-- **Flow only asks when you chose to be asked.** A Flow project no longer
- overrides your safety mode with a leash you never picked - and when it
- does ask, the card names the action instead of a generic sentence. The
- project now carries that choice openly, as a choice you make on the
- project itself, and it is still there when you come back to it.
-- **Every card that asks first says more than what.** A card that stops an
- action now names the action in plain words, says what happens if you say
- no, and says how far an "always" would actually reach - before you press
- either of them.
-- **Notifications carry the whole message.** The bell no longer cuts a
- report to a preview mid-sentence, and a task result links to the task.
-- **Discover posts can travel a safer road.** New builds post through the
- relay, which signs on their behalf; if the relay is unreachable they fall
- back to the direct road. Older builds are unaffected either way.
-
-### Fixed
-
-- **Notifications speak your language.** Around forty of them were still
- written in English whatever the language picker said: a task finished or
- blocked, a schedule that switched itself off, a stand-up, a meeting in ten
- minutes, an approval waiting for you, a Telegram bot that gave up, a room
- board that assigned you something, a video or a cut that finished rendering.
- All of them now come in the language the rest of Skales is in. The half of
- those messages that also lands in a chat, where the model reads it back on
- the next turn, deliberately stays English so a language setting can never
- change what the model was told.
-
-- **The advanced provider settings stay where you put them.** The request
- timeout, the retry counts and the per-model context and output limits used
- to be lost if you left Settings without pressing Save. All of them keep
- themselves now, as you set them, and removing the last override really
- removes it instead of the old value coming back.
-- **Restoring a backup works again.** Restoring a backup archive - and
- importing a plugin from a file - failed in the packaged app with an
- unreadable one-line error. The unpacker now runs outside the bundled
- code, a release check holds it there, and both doors open again.
-- **Local voices on Linux.** The AppImage shipped a speech engine that
- looked for its libraries in the build machine's directories; the release
- check now reads the shipped binary and refuses a broken one, and the app
- names the real cause on screen if it ever meets one.
-- **A downloaded voice is usable immediately.** Voice downloads unpack when
- they finish instead of at the next restart - the download list and the
- voice picker can no longer disagree about what is installed.
-- **Local speech hands its audio over safely.** The engine's output is now
- copied before playback; this very likely restores on-device voices that
- produced silence with "external buffers are not allowed".
-- **A 3D scene that cannot draw says which of two reasons applies** - a
- missing library and a display without WebGL are different problems, and
- the admin review page says when its preview runtime is missing instead of
- showing a false WebGL error.
-- **Four approval cards spoke English to everybody.** Asking to build a
- spreadsheet, to call one of your own API connectors, or to post or upload
- to a social account put the question in English no matter which language
- the rest of Skales was in. All four now ask in your language.
-- **The Code window no longer starts underneath the navigation.** With the
- sidebar open, the first characters of every line you typed there were
- hidden behind it. The window now begins where the rest of the app begins,
- in every navigation layout, and the way back out stays visible.
-- **Restoring a backup asks first.** A restore replaces the settings,
- memories and connections on this machine and cannot be undone; it used to
- happen on a single click. There is now a switch beside the button that
- says what will be replaced, and it is off again every time you open the
- page. The backup card itself is translated too, buttons included.
-- **More link previews find their picture.** A page that names its preview
- image in the less common of the two ways it is written was read as a page
- with no picture at all.
-
-## v12.9.0 - Fable
-
-### Added
-
-- **Skales tells you when it healed itself.** If the background engine dies,
- every open window now shows what happened and why while it restarts, and
- confirms when it is back - including windows that never noticed a thing.
-- **New installations start on Claude Sonnet 5.** Existing settings are
- untouched - whatever model you chose stays chosen.
-- **The setup now ends with fifteen seconds of Skales, not an empty chat.**
- Pressing the last button plays four full-bleed cards: what Studio makes and
- that it can go into the Discover feed, an agent ticking a goal off by
- itself, Iris listening and answering, and the Guide as the door to
- everything else. One press skips it from any card, it is never shown unasked
- a second time, and with reduced motion switched on the cards stand still and
- you press Next. Guide has a "Watch the intro" line to play it again.
-
-- **Two cards that appear where they are an answer.** The first Skales Visual
- you finish offers to put it in the Discover feed, and a run longer than two
- minutes that finishes while you are somewhere else offers to tell you next
- time - the second one only while notifications are actually off. Each is
- offered once, both can be closed without being asked again, and both stay
- quiet while notifications are muted or during quiet hours.
-
-- **A switch for staying awake in the background.** Settings > Desktop App
- now carries "Stay Awake in the Background". On, which is the default and
- what every release so far has done, Skales keeps the system from parking it
- so reminders, scheduled tasks and autonomous work keep running while the
- window is not in front. Off arms that only while a job is actually in
- flight, which is the setting for a laptop on battery. The change takes
- effect immediately, with no restart.
-
-- **Memory can be searched.** The Memory page opens with a search box over
- everything stored - topic names and the text inside them, across the index,
- the topic files and the saved short-term, long-term and episodic records -
- with hits highlighted and a Meaning toggle that appears only when an
- embedding provider is configured. Skales itself gained the same search as a
- tool, so it can find a note it wrote months ago without guessing the topic
- name, and asking for a topic that does not exist now answers with the
- closest ones that do.
-
-- **The first run now asks who you are before it asks for keys.** A new
- question near the top - newcomer or AI-head - reshapes the whole setup:
- newcomers get plain-language explanations under every technical term and a
- guided provider choice, experienced users get dense screens and can skip
- whole blocks. The answer is remembered and keeps shaping recommendations
- after setup.
-- **Setup covers the everyday now.** Theme with live previews, light/dark/
- system, accent tone, your city, weather, notifications, a voice you can
- hear before choosing, safety mode and telemetry are all settled during the
- first run - writing the same settings the Settings page writes, so nothing
- exists twice. Feature cards with small live animations introduce AIPointer,
- the desktop buddy, Iris and Studio, and switch them on right there. A QR
- card pairs your phone without leaving setup, including the allow/deny
- confirmation, and the whole run works offline: every card degrades to a
- named fallback and Continue always stays reachable. Progress survives a
- restart and resumes exactly where you stopped.
-- **Docs became Guide.** One page with two entrances: "I want to..." paths
- sorted by how much setup they need, and a catalogue of what is actually
- possible - each line showing live whether your machine is ready, with a
- real example prompt that lands editable in the composer. Settings open as
- a popup on top of the page, and the bundled handbook stays available
- offline. Setup ends on a card that shows how much of Skales is already
- set up.
-- **Your experience level reaches a connected phone instantly.** Answering
- the newcomer question on the desktop no longer waits for the next
- reconnect - a paired phone learns it the moment you answer.
-- **Setup can be replayed without losing anything.** Guide and Settings >
- General offer "run onboarding again": every answer arrives prefilled with
- the current state, and finishing merges only what you changed - identity,
- sessions and history stay untouched.
-
-### Changed
-
-- **An idle window is idle again.** Around sixty background timers kept
- running at full rate on a window nobody was looking at - the dashboard's
- own CPU readout polled twice a second, the Discover network graph redrew
- its physics sixty times a second, Iris kept a WebGL scene painting behind
- whatever was in front of it, and Studio, Flow, Playground, Playbooks,
- Teams, Logs, the Code window and the Autopilot and Organization panels each
- kept their own. Every one of them now pauses while the window is hidden
- **or** unfocused (only checking "hidden" misses a covered window on Windows
- entirely) and refreshes the instant it comes back, so nothing reads stale;
- clocks read their start stamp rather than counting ticks, so they are right
- on the first tick back rather than behind by however long you were away.
- What genuinely has to run while you are away still does - completed runs,
- renders, recordings, calls, the presentation clock, the buddy, OS
- notifications - and pays with a backoff instead: the notification drain
- went from every two seconds to every ten, stretching to thirty while there
- is nothing to deliver. This cannot quietly grow back, and every case that
- genuinely has to keep running while you are away still does.
-- **Flat got flatter.** The Flat sidebar is a menu now, not a row of buttons:
- 13px entries, tight rows, no pill on the active item - a soft tint and a
- 2px accent edge carry the position instead. Header, status card, switcher
- and foot bar slimmed down to match. Skales X and Classic are untouched,
- and a test walks every Flat rule to keep it that way.
-- **Templates is back in the sidebar.** The templates page never left - only
- its door did. It now sits in the System group right under Custom Widgets,
- and every card still lands its prompt editable in the composer.
-- **Three themes instead of six.** Obsidian, Snowfield and Neon are retired;
- anyone using them lands on Classic once, with a short notice. Slate is now
- Flat and means it: grey and white, lime as the only colour, a text-only
- sidebar, tighter lines, emojis in greyscale - the logo and the buddy keep
- their colours. Preset accent tones now also work on Flat.
-
-### Fixed
-
-- **The Autopilot cockpit is current the moment you come back.** Returning to
- the window refreshes status, tasks, logs and goals immediately instead of
- showing up to eight seconds of old state.
-- **A renamed chat carries its new name into a second window.** Renaming, and
- the automatic title, now reach every surface instead of only the one you
- were looking at.
-- **Background notification delivery cannot lose itself anymore.** A wake
- during a running delivery pass no longer forks the schedule, and a single
- failed server answer no longer silently ends it.
-- **The token wall now says that it is one.** Changing the remote-access
- token also locks out the Skales window on a paired phone; that refusal was
- indistinguishable from a working page. It now names itself in the reply,
- and nothing changes for older devices.
-- **Studio gallery cards keep the shape their media was made in.** A 9:16
- visual is tall in the gallery, the same way it is tall everywhere else -
- instead of being cropped into a square.
-- **Reopened conversations show the team pill again.** A turn that fanned
- out to sub-agents records how many worked on it, so the badge survives a
- restart instead of appearing only while you watched.
-- **More of Skales' tools name the capability they belong to.** Skales can
- now say what it has instead of only having it, and reports this completely.
-- **A slow image no longer dies at an invisible deadline.** The vision path
- now honours the request-timeout setting and grants extra time by
- attachment size, a transport failure ends the turn with a named card and
- a Retry button instead of a raw exception leaking into the answer, and
- the model is finally told that images arrive through vision - it will
- not try to read a JPEG as text any more.
-- **A failed send no longer eats your message.** Attachments return to the
- composer when a run is not accepted, and a run that dies before
- answering says so with a Retry instead of silence.
-- **One picture, once.** An attached image rendered up to three times -
- live and as a ghost after reload. One layer now wins, and old sessions
- are repaired on display.
-- **Chat titles hold still.** No renaming mid-stream, and an automatic
- title no longer reorders your history.
-- **Google Drive and Docs actually work.** Creating a document hit a
- broken API path and always failed; downloading a Google Doc or Sheet
- tried to fetch bytes that do not exist and answered 403. Documents
- create properly now (with optional initial text), Workspace files export
- to a readable format, real binaries arrive undamaged as files, and
- errors name their true cause - including a missing permission.
-- **The local speech engine names the right gap.** A build missing its
- platform package claimed "not a missing download" - the diagnosis now
- names both possible build gaps honestly.
-- **The connection card no longer says "connected" in red.** Its colour was
- chosen separately from its words, and a keyless runtime like Skales Local
- did not count as connected at all. One tone now derives from the same
- state as the label, using theme tokens instead of hard-coded colours.
-
-- **A shared 3D visual draws in the feed instead of showing a black box.** A
- moving visual shared to Discover used to render on the phone and nowhere else
- - the desktop feed and the moderation view showed an empty frame.
- Both now draw the same visual the phone does, and a simpler design still looks
- exactly as it did before.
-- **Sharing a visual no longer cuts it in half.** A large shared visual could be
- silently trimmed, which broke it and left a black frame with nothing saying
- why. A share now sends exactly what was made, and one that is genuinely too
- large is refused clearly, up front, with the reason in the sentence.
-- **The setup keeps one face from start to finish.** From the feature cards
- onward the wizard's header swapped the Skales mark for the buddy skin you had
- just picked. The mark is now the same on every screen of the run, and the
- buddy is shown where it is chosen.
-- **Setup screens are readable immediately.** The feature previews stopped
- animating when they scrolled out of view, but they stopped on whichever frame
- they were standing on - and several of those frames are deliberately faint,
- which left cards permanently half-drawn. A stopped preview now simply rests.
- The step text also fades in twice as fast and can no longer be caught between
- two overlapping fades.
-- **Leaving the provider screen without choosing asks first.** Pressing
- Continue with no card selected walked straight past the one question the
- defaults cannot answer. It now asks, in your language, and says what the
- machine will actually use if you go on. Going on still changes nothing.
-
-## v12.8.5 - Mirror
-
-### Fixed
-
-- **A cancelled upload no longer takes the server down with it.** Releasing
- the microphone, reloading the page mid-answer or letting a voice detector
- restart could end the whole Skales server, and with it every agent, the
- scheduler, the autopilot and any background work in progress. The cause sits
- in the runtime's own machinery: it finishes reading an abandoned request body
- by closing a stream twice, from a place no error handling can reach. That one
- fault is now recognised and outlived - written down, with its stack, so it
- stays visible - while every other kind of crash still ends the server the way
- it always has.
-- **A released microphone stops costing money.** A transcription whose client
- had already walked away ran to the end at Groq, OpenAI, OpenRouter, Azure or
- Skales IQ, and when the disconnect looked like a refusal, the next provider
- was tried after it. A cancelled recording now cancels the call it started, and
- a cancellation is answered as one instead of as a failed transcription.
-- **Pressing the microphone twice sends once.** Every voice surface - chat
- dictation, the chat microphone, Call mode, Iris and audio attachments - posted
- without knowing about the others or about its own previous run, so a fast
- second press or a restarting voice detector put two large uploads on the wire
- at once. A newer recording now replaces the older one cleanly and the replaced
- one leaves without an error message. Several audio files attached together
- still transcribe side by side.
-- **The Telegram bot no longer disappears without a word.** An unexpected fault
- ended the bot process leaving nothing written down anywhere, so a restart
- looked like it had stopped for no reason. Every fault is now recorded with its
- stack, and the one stream fault above is survived there too - the bot sends
- media through the same machinery.
-- **A tool that runs into a missing WordPress route says which plugin to
- update.** The message a user actually got was "the plugin is not active on
- that site", and it was the wrong advice: on a site that connects fine, an
- older connector simply does not carry the newer routes yet. That diagnosis
- never even reached the screen, because it was tested against the words of
- the message and WordPress does not put "404" in its sentence. Both halves
- are fixed - the reason is read where WordPress actually writes it, and a
- missing Skales route now names the plugin version to update to. It is only
- shown after both address forms have been tried, so a site on plain
- permalinks is never told this.
-- **A page is not a post when Skales asks whether it may write.** The check
- that runs before every WordPress call asked for the post rights when the
- tool was about to touch a page, and for edit rights when it was about to
- delete. An account that may edit posts but no pages was waved through and
- then refused by WordPress itself, as a bare 403 the model could not read.
- Deleting and reading now also ask about the object they were actually given.
-- **Skales names itself to the connector plugin.** The handshake carried the
- plugin's version but never the app's, so the plugin could not say whether
- Skales was the outdated half of the pair. It now does, and when the plugin
- says so, that sentence is shown next to the one about the plugin. Where the
- plugin publishes its route level, that number decides the message instead of
- the version string - so a site that already carries every route Skales calls
- is no longer asked to update for nothing.
-- **Minimax pointed at two different countries at once.** The chat default
- addressed the old `api.minimax.chat`, while the Refresh Models button on the
- same card asked `api.minimaxi.com`, which is the mainland China deployment
- and refuses an international key. Both halves now address the international
- host. A base URL you entered yourself is untouched.
-- **Developer, Docs listed tools that do not exist.** Fifteen entries carried
- invented names - among them `calendar_list_events` and `calendar_create_event`,
- which are really `list_calendar_events` and `create_calendar_event` - and one
- entry described a chat skill as a tool. Every name on the page is now the
- name the model answers to, and the safety each one promises is the one that
- is actually applied.
-- **A failed WordPress connection says why it failed.** Test Connection and the
- WordPress screen both answered every failure with the same flat "Connection
- failed", no matter whether the site had rejected the token, answered 404, run
- into a timeout or handed back a firewall page. The real reason now reaches
- the screen, in the same words the WordPress tools have always used.
-- **A site with plain permalinks connects.** WordPress only serves
- `/wp-json/` when pretty permalinks are switched on; with the default setting
- the identical route exists only under `?rest_route=`. Skales tried the first
- form and gave up. It now tries the other one when the first answers 404, and
- remembers which form that site speaks.
-- **An outdated connector plugin is named.** An old plugin answered the
- handshake and then let every newer tool run into "not found". The version is
- now checked at the handshake and named on both screens, including the case
- where the plugin is too old to report a version at all. Nothing is blocked:
- what the old plugin can do, it still does.
-- **Two new model profiles ship inside the app.** Muse Glimmer and DBRX were
- added to the shared profile library, but only machines that can reach GitHub
- received them - offline installs kept running both untuned, including Muse
- Glimmer 30B, which Skales' own local catalogue recommends as its agentic
- flagship. Both profiles are now built into the app itself.
-- **Local speech now works on Windows and on Apple Silicon.** The engine behind
- dictation and read-aloud loads a native library chosen by platform, and every
- release since it arrived carried only the one belonging to the machine that
- built the package. A Windows install therefore held macOS libraries and no
- Windows one, and reported that the module could not be found - which read
- like a damaged download and was nothing a reinstall could ever repair. The
- Apple Silicon disk image had the same hole with an Intel library in it. Each
- package now carries the library its own machines can load, and no other, and
- a build that gets this wrong stops before it is packaged.
-- **A speech engine that cannot load says why, and writes it down.** The old
- message named neither the cause nor the platform and left no trace in the
- system log at all. It now names the missing piece, says plainly that the gap
- is in the download rather than on the machine, and records one line in the
- log so a bug report has something to carry.
-- **Google Calendar and Discord can be set up again.** Both cards lost their
- Configure button when the settings switches were unified, and the switch that
- was left behind ran the disconnect: the only control on the card asked whether
- you were sure and then deleted the connection, and nothing anywhere could
- open the panel holding the API key, the calendar IDs or the bot token. Both
- cards carry the full row again - a switch for the skill, Configure to unfold
- the setup, and Disconnect as its own red action with a confirmation. Browser
- Control lost the same button and had its switch wired to the panel instead of
- the skill, so it could be neither turned on nor off from its card; that one is
- fixed too. All fourteen collapsible setup panels in Settings are now checked
- before every build, so a card cannot go unopenable again.
-- **Skales knew about YouTube uploads and never said so.** Publishing a video
- is a different Google permission from searching for one, and the capability
- report treated posting to YouTube and LinkedIn as always available whether
- anything was connected or not. It now reports the real state, and when your
- Google account was connected without the upload permission it says exactly
- that, and sends you to reconnect it with YouTube (upload) ticked instead of
- claiming Skales cannot upload at all.
-- **Skales can now describe its own memory, and the rest of YouTube.** The
- long-term memory it keeps for you had no entry in the capability report at
- all, so asked what it remembers between sessions it had nothing to say while
- reading and writing that memory every turn. YouTube listed only search and
- kept quiet about video details, channel information, trending and caption
- tracks; browser automation left out attaching a file to a page. Every tool
- family is now measured against the report before each build, so a whole
- feature cannot go unmentioned again.
-- **The WordPress tools arrive together.** With on-demand tool loading the 47
- WordPress tools were spread over four different groups and no group named
- WordPress, so a model asked to work on your site fetched whichever group
- sounded closest and got eleven of them, then reported the rest missing. They
- are one group now, called wordpress, and it says what is in it.
-- **Generating a picture no longer ends the job it was part of.** An agent that
- made an image halfway through a task dropped a bare "Image generated!" line
- into the chat and stopped there, having never seen the result of its own tool
- - so the task it was given was simply abandoned, and asking it to carry on
- meant explaining the whole thing again. Making a picture is now a step like
- any other: the preview card still appears, and the same turn keeps going with
- the file it just produced. Video, speech and 3D generation already behaved
- this way and are unchanged.
-- **A picture attached in a chat is a real file, and Skales says where it is.**
- Attachments were described correctly and then existed nowhere any tool could
- reach, so asking to post one, edit it or convert it failed on a file that
- looked like it was right there. Every attached picture is kept in the
- workspace and its location is now part of the conversation, so the tools that
- need the actual file can open it. Pictures sent from a phone or the browser
- interface were not being kept at all; they are now, in the same place and
- under the same name as on the desktop. The location travels with the
- conversation, so it still works after a restart.
-- **One history, whichever agent you were talking to.** Starting a conversation
- with a custom agent replaced the sidebar with that agent's conversations only
- and offered no way back to the full list, while the conversation itself never
- appeared in the normal history afterwards. There is one list again, holding
- everything, and a conversation held with a custom agent carries that agent's
- emoji on its row. Opening such a row switches to that agent, so the reply
- comes from whoever the conversation was with. Narrowing to a single agent is
- still possible on the History screen, where it is a choice with a way back
- rather than a place to get stuck.
-- **A shared visual keeps the shape it was designed in, and never lands
- blank.** A visual shared from Studio arrived in the feed without any note of
- the format it was built for, so a square, a widescreen and a portrait design
- were all shown in the same fixed box. The chosen aspect now travels with the
- visual and the feed lays it out accordingly - for visuals shared from a phone
- as well. And a visual whose frame had not been measured yet was scaled to
- nothing at all and read as an empty card; it now waits for a frame with a
- size before it fits itself into it.
-- **A Playground space shared to Discover arrives with its text.** The
- description and the category were sent in a place the feed never reads, so
- the card appeared with an empty body under the wrong heading. Both now ride
- where the feed looks for them.
-- **A voice preview no longer answers with a different voice.** Trying out a
- text-to-speech provider in Settings read the sentence through the full
- cascade, so a refused key, an empty balance or a dead endpoint all ended with
- the system voice speaking flawlessly and the preview looking like a success.
- The preview now stops at the provider that was picked and prints the reason it
- gave - the HTTP status, the refusal, the missing key - instead of handing the
- sentence on.
-
-### Added
-
-- **Settings > Voice > Fallbacks: every rule that swaps one engine for
- another, written down in one place.** Which voice steps in for which, in what
- order transcription is tried, what happens when the main chat provider fails,
- who reads an image sent to a model that cannot see, where a local image
- generator hands over, and what memory search does without embeddings. The
- voice rule can also be switched off there: with it off a reply stays silent
- and names the reason rather than being read by a voice nobody chose.
-- **A rendered video can go to Discover, from every place Studio makes one.**
- Until now only pictures and Skales Visuals had a share: the cloud video, a
- Scenes render, a Type export and a Flow composition all ended at Download.
- Each of them now carries the same Share to Discover button the picture has,
- and an authored Flow page - a 3D scene among them - shares as the visual it
- is. A clip over 4 MB is refused before anything is posted and says so with
- the limit; a shared post waits for review the same way a shared picture
- does. Videos become visible in the feed once the feed server carries the
- matching upload path.
-
-### Changed
-
-- **3D comes back as a scene, not as a web page with a scene in it.** Asked for
- a 3D element, a language model writes a whole landing page around it: a
- headline, two paragraphs, a footer, and a small canvas in the middle. Skales
- now stages the canvas itself and hands the model the frame - a stage that
- fills the view, a camera, a three-point light rig, ground that takes the
- shadow, the resize handler and the loop - and says in as many words that a
- page around it is the wrong deliverable. What the model writes is the part
- that was ever the point: the object, the materials, the camera move and the
- mood. Chat, Studio and Flow ask for it in the same words.
-- **Buttons that turn the thing still work.** "A cube with buttons that roll it
- and light it from the sides" is exactly the kind of request the rule above
- could have killed, so the frame has a place for it: controls float ON the
- canvas and let the finger through everywhere else, instead of standing in a
- document around it.
-
-### Fixed
-
-- **A turn that fails says what failed.** Whatever ended a chat turn - a model
- that ran past its time budget, an endpoint that refused the connection, a
- certificate this machine will not accept - the bubble said "Connection error.
- Please try again." and kept the real sentence, the one naming the budget or
- the address, out of sight in the session file. The reason now stands in the
- bubble, in the language the app is set to, with the raw text kept underneath
- it for the retry and the bug report.
-- **Dragging a sidebar entry no longer shows a web address.** A slow press on a
- nav row and a small move started a browser link drag, and the ghost under the
- cursor read "http://localhost:3000/..." - the app naming its own dev server at
- the user. Nothing in the sidebar is something to drag, so nothing there drags.
-- **"Reasoning off" reaches Ollama.** The setting travelled to every backend
- that has an off switch except the local daemon, where a thinking block is not
- a second of latency but a minute of it. It now goes out as the daemon's own
- parameter on both of its endpoints - the OpenAI-compatible one a text turn
- uses and the native one an image turn uses - so a Qwen or a Gemma stops
- thinking when it is told to. Older daemons that do not know the field answer
- the request unchanged, so nothing regresses on an install that predates it.
-- **A connected MCP server's tools are reachable on a local model.** The list of
- tool groups a model can load on demand was written from the tools that
- survived the local model's tool budget instead of from the whole catalogue,
- and on a local model almost nothing survives that is not in the base set. The
- result: no groups were offered, none were registered, and asking for the MCP
- server's group came back "unknown group" for a server that was connected and
- enabled - so its tools were unusable until on-demand loading was switched off
- entirely. The index now describes what CAN be fetched, which was always its
- job, while the tool list still describes what this turn is carrying.
-- **Web search set to MCP is honoured before the model chooses, not after it
- fails.** With the built-in search withheld, the model reached for it anyway,
- was told the built-in is set to MCP, and only then went looking - one wasted
- step, every time. Skales now names the connected search tool in the prompt,
- and a call that still arrives at the built-in is carried through to that tool
- instead of being refused. If no connected server actually exposes a web
- search, it says exactly that and names the way out.
-- **Skales knows where its own tool-loading switch is.** Asked where to turn
- on-demand tool loading off, it sent people to Settings > Advanced, where there
- is no such control. It is under Settings > Memory, in the Memory Mode card,
- and that is now what the app knows about itself.
-
-### Added
-
-- **A shared Visual opens full screen.** Every Skales Visual in Discover, from a
- flat design to a turning three.js scene, has an expand button on its frame:
- it fills the window, takes the mouse or the finger, and closes with the button
- or with Escape. The card's own frame is put aside while it is open rather than
- thrown away, so closing carries on where it was instead of starting the
- animation over.
-
-## v12.8.4 - Aware
-
-### Fixed
-
-- **A restart says it is a restart.** When the system ended Skales' background
- engine to free memory, the window went black with "Connection lost" and a Try
- again button, and nothing said that Skales was already starting itself back
- up. Every window now says what happened, which attempt it is on, and whether
- the engine was stopped by the system or fell over on its own - and when it is
- running again the surface comes back by itself, with nobody pressing Try
- again. If three attempts are not enough, it says that too instead of leaving
- you at the same button.
-- **Skales stops denying what it can do.** Asked whether it could work with
- your WordPress site, it said no, although the connector has been there for
- months and forty-seven tools sit behind it. The same hole swallowed Obsidian
- and the ability to use your computer. All three are in its self-knowledge
- now, each one honest about whether it is actually connected: WordPress
- reports itself as ready once your site and token are stored, not before.
-- **A fact said twice is one fact.** Telling Skales your name in three
- different conversations left three separate records saying the same thing.
- Near-identical facts are now folded into the record that was there first,
- which keeps the earliest date and counts how often it has been confirmed -
- visible on the Memory page. Facts that only look alike stay apart: "port
- 8091" and "port 8093" are still two facts. What is already stored is tidied
- once, without anything being thrown away.
-- **A document Skales writes is a document you can find.** Asking for a
- spreadsheet could report success and leave no file behind at all, because the
- packaged app cannot write the way the library wanted to. Fixed in the last
- release, and now checked on every build by actually writing a workbook and
- reading it back, so a tool that claims success and produces nothing cannot
- get through again.
-- **A fresh install leads with what it can do.** Asked what it was capable of
- before any key was entered, Skales read out the long list of everything not
- set up yet. It now names what works right away first - Flow, Code Mode,
- visuals, your local files, the weather - and puts what a key would add into a
- single sentence with the place to enter it.
-- **It knows it can draw a 3D scene without a provider.** Flow has written real
- three.js scenes for a while, but the ability was missing from Skales' own
- list, so the answer to "can you do 3D?" came from the model-generating tool
- that needs a provider and an account. It now says the true thing: the scene
- is written and drawn in the app itself, no key involved, and the way in is
- the 3D chip in Flow.
-- **The assistant stops opening every answer the same way.** When the same
- short phrase has led three of the last few replies, it is asked to begin
- differently. Tone, warmth, persona and language stay exactly as they were;
- only the repetition goes.
-
-### Changed
-
-- **The shipped app is sealed.** Skales now ships in a hardened package.
- Nothing changes for you - the app installs, starts and updates exactly as
- before.
-
-## v12.8.3 - Intact
-
-### Fixed
-
-- **A finished answer is no longer followed by a second one nobody asked for.**
- A line typed while Skales was still writing goes into the queue, as before.
- Two things then went wrong with it. A queued message that could not be handed
- over was put back, and putting it back is what woke the queue again, so the
- same text was submitted over and over; and the composer unlocks by itself when
- a turn stops responding, which the queue read as "the turn is over" while the
- turn was still running on the machine. Now a message is tried once more at
- most and then waits in plain sight with a Try again next to it, and nothing
- goes out until the run is genuinely finished rather than merely unlocked.
-- **Long Telegram answers arrive whole.** Since 12.8.2 the answer is written
- into the message you are already looking at, and that path never checked
- Telegram's per-message limit: a long answer stopped mid-sentence at the last
- piece that still fit. A message that had waited in the queue was worse - it
- counted as delivered although Telegram had refused it, so it never arrived at
- all. Answers are now split and sent in order, with code blocks closed and
- reopened across the split, and nothing counts as delivered until Telegram
- confirms it.
-- **Saving a mail account no longer destroys its password.** Changing any field
- on an existing account wrote the row of dots back over the real password. If
- an account is already in that state it now says so and asks for the password
- again, instead of reporting a login failure from the provider.
-- **The mail Test button checks what is on screen.** It tested the saved values
- while showing the verdict as a verdict on the ones just typed, so a corrected
- port was reported as refused before it had been tried. Only the password still
- comes from storage.
-- **Port and encryption belong together.** Switching encryption moves the port
- the way the label always promised, and the four impossible combinations say so
- at once instead of arriving later as a certificate error - which is also what
- implicit TLS on a STARTTLS port used to be reported as.
-- **Skales answers correctly about its own integrations again.** A read-only
- mail account, or a settings file written by an older version, made the
- assistant deny Google, Drive, Docs and IMAP that were configured and working
- the whole time. Nothing had been removed: its self-check read different files
- than the tool catalogue and the settings screen. On local and small models
- those tools could also be squeezed out of the tool budget first; they are now
- protected once configured. A Google account missing one field now names the
- field instead of vanishing.
-- **The context size you set for a local model reaches the engine.** It was
- never passed on, so the server chose a window on its own - 4096 on some
- machines - while the banner read your setting back to you rather than the
- window in force. The banner now shows the window the server actually holds,
- changing the setting restarts the local server, and a request that does not
- fit ends the turn once with the real numbers instead of being retried until
- the partial answer has been erased and restarted several times.
-- **A local model no longer advertises reading pictures when it cannot.** The
- VISION badge is checked against the running engine, so a model whose projector
- is not on disk says so. An engine that is running with nothing loaded is no
- longer reported as not running, the role buttons say why they are disabled,
- and a start skipped because autostart is off says that too.
-- **Raw `<textcall>` markup no longer appears in answers**, and - more
- importantly - the call inside it is now carried out. Removing the markup
- without running the call would have meant asking for something, and nothing
- happening, with no error.
-- **Teams, Organization, Group Chat and Swarm are reachable again**, on every
- theme. The pages and the runner were never touched; only the doors to them had
- been taken out, and on the top-bar and icon-rail themes there was no way in at
- all.
-- **One conversation can no longer lock the others.** Opening a chat kept the
- previous one's run, which held the input and the history list; starting a new
- chat was the only way out. A conversation that cannot be drawn now shows the
- reason and leaves the rest of the app alone.
-- **Allowing something "for this session" holds across Chat, Code and Flow.**
- The three surfaces kept three separate notes of it, and in Code the answer
- could land in a copy the waiting run never saw, so the same question came back
- after it had been answered. In Flow a card showed "running" even when the
- answer could no longer be applied - after a restart, for instance; it now says
- the runner was lost and asks you to send again.
-- **Rewrite (/spin) keeps what it produces.** The rewritten text was only put on
- screen, so the next refresh of the conversation removed it again - the feature
- ran and appeared to do nothing. It is written into the conversation now, and a
- write that fails says so and leaves the text where it can still be copied.
-- **The Home/Chat switch shows the column that is on screen.** Opening a
- conversation raised the chat column while the switch still lit Home. Your
- stored preference is left alone.
-- **Portrait pictures in Discover are shown whole** instead of cropped to a
- band.
-- **Browser Control can attach a file to a web page.** Upload fields that a page
- hides are now listed and can be addressed directly, including through the
- button in front of them, so no operating-system dialog is needed; the field is
- read back afterwards and a mismatch refuses to submit.
-- **Custom endpoints are first-class in the agent editor.** Each server you set
- up appears under its own name, the choice survives reopening and restarting,
- and an endpoint that has been switched off is named rather than shown as
- "Default".
-- Memory "Minimal" now says what it does: it shortens the memory context, and
- the tool schemata have their own switch directly below it.
-
-## v12.8.2 - Mood
-
-### Added
-
-- **Your Obsidian vault is a folder Skales lives in, not a copy it took.** The
- old import read your notes once and kept a snapshot, so everything you wrote
- afterwards was invisible while Skales still answered as if it knew. Settings,
- Integrations, Obsidian Vaults is where you point at the folder itself, give it
- a name ("Personal", "Work") and add as many as you keep. Every question reads
- the files on disk, so what Obsidian shows you is what Skales sees, and you can
- say "look in my Work vault" and be understood. The Memory page shows the note
- graph of whichever vault you pick, and offers to remove the old copy. The
- integration ships as a beta, and connecting the first vault asks one honest
- question first: what Skales reads goes to your selected model uncompressed,
- which costs tokens on long notes, and backups stay your job.
-- **Skales can write into the vault, and never over it.** It can start a new
- note and add to an existing one, at the end or under a heading you name. It
- cannot replace a note: a create that would land on a note that already exists
- refuses and points at appending instead, so ten years of notes cannot be
- overwritten by a misunderstanding.
-- **Where new notes land is a setting, not a request.** Target folder, filename
- pattern, note template and a tag on everything Skales files are yours to set
- under the vault list, with {title}, {date}, {time}, {content} and {tag} as
- placeholders. Leave them empty and a note lands in the vault root under its
- own title.
-- **Your notes answer with what is in them.** Frontmatter and #tags are read
- along with the body, "tag:project" narrows a search to notes carrying that
- tag, and a note comes back with the notes that link to it and the ones it
- links out to.
-- **Conscious: the mood it carries is in front of you while you work.** The
- companion's state was real and kept, but it was only ever readable on the
- Memory page, which is not where anybody is when the work is happening. There
- is a row of its own now, pinned above the bottom of the navigation and in view
- the whole time you are in a conversation. The colour is the mood, warm when
- the work is going well and cool when it is not, and how full it stands and how
- fast it moves is the energy. It changes while you are talking, not on the next
- restart. Hovering it opens a panel that says what the colour means, what moved
- it today with the times, and what is missing, where missing only ever means
- something that is in the record: a long silence, or a day with nothing
- finished yet. If interest tracking is on it lists what it keeps coming back
- to. It is a child of the companion switch in Settings, Memory, it draws and
- nothing else, and it never notifies, sounds or starts a conversation.
-- **The mood keeps its day, and the things that move it can happen.** The state
- kept only the current value and threw every event away, so "what happened
- today" had no answer; there is a rolling log of the last two days now. Four of
- the seven things that can move the mood had no source in the code at all: a
- finished goal, a goal that did not get there, a run stuck on the same error, a
- long silence and a long stretch of work all register now. Resetting the mood
- on the Memory page clears the day's list with it.
-
-- **Telegram answers in one message, and it writes as it thinks.** A question
- put to the bot produced a "Thinking..." line and then, some time later, a
- second and separate message with the answer in it, so a conversation with
- Skales on the phone read as twice as many messages as it had. The answer now
- grows inside the message that says it is working, the way it does in the chat
- window, and there is no second message at all.
-- **A file Skales makes arrives as a file.** A spreadsheet, a PDF or a written
- document used to come back as a path in a sentence: unreachable from a phone,
- meaningless in a chat window on another machine, and a string to copy by hand
- here. The answer now carries a card with the name, the type and the size, a
- Download button and a way to open the folder. On Telegram the file itself is
- sent into the room the question came from, and a file the bot cannot send,
- because it is over the 50 MB the Telegram API takes, says so with its name and
- where it is instead of vanishing.
-- **Saving a character with the companion switched off asks whether to turn it
- on.** The character shapes how Skales talks whether the switch is on or not,
- but everything it LEARNS - the mood, the topics, the working history - needs
- the switch. Saving a character while it is off now asks once, in a dialog,
- and takes no for an answer permanently. Nothing is turned on without the
- answer being yes.
-
-### Fixed
-
-- **Telegram answers once.** Every reply arrived twice, and the notice that a
- message is waiting arrived twice with it, while the send-a-message tool posted
- exactly once. Nothing was sending twice: two bot processes were reading the
- same chat. The lock meant to allow only one of them checked whether it existed
- and then wrote it, and two bots started in the same moment both read "no lock"
- and both went on to poll. Telegram hands the same message to both, and the
- repeat check that would have caught it lives in each process separately, so
- neither could see the other. The lock is now taken in one indivisible step,
- the loser stops instead of polling, and it only ever clears its own. Restarting
- never helped before because the start itself was the race: opening the app asks
- three places at once to make sure the bot is running, and each one started its
- own. They share one attempt now. There was a second double hiding under the
- first: an answer that had grown inside the Thinking message was posted once
- more as its own message, because the note saying it was already delivered was
- dropped on the way back to the bot. That note now travels with the answer.
-- **A line typed while Skales is thinking goes into the queue.** During the
- thinking phase, before any text appears, pressing Enter left the sentence
- sitting in the box: not queued, not sent. The queue was real, the decision to
- use it just came too late, behind other work that could take the send first. A
- turn is running from the moment it starts, not from its first word, and from
- that moment every typed line is taken, the box clears, and what was typed is
- answered together with the rest. A stop command still goes straight through to
- the run it is meant to halt.
-- **Opening Skales no longer arrives as a wave of pop-ups.** Everything that
- catches up when the app comes up, the briefing, the messages a channel held
- while it was away, the check-ins armed while it was closed, reached the
- screen at once in the first
- seconds, before there was any chance to read one. For the first minute after a
- start they are recorded and left on the Notifications page instead of firing.
- Nothing is lost and nothing is dropped, and anything urgent still comes
- straight through.
-
-- **The button that adds your own endpoint opens something.** Under Settings,
- AI Providers, the "+ Custom endpoint" tile marked itself open, took the page
- to the bottom of itself and showed nothing at all, so there was no way left to
- add or edit an endpoint. The block it opens was classed as an Advanced-view
- section while the tile stood in both views, and a search that found the tile
- could miss the block the same way. The tile and what it opens are one thing
- now, in either view and under any search.
-- **A model list that cannot be fetched says why.** Pressing Fetch Available
- Models on an endpoint that was not running answered "no models returned", an
- endpoint answering 404 or 401 was reported as a server that had never started,
- and an address whose name does not resolve said "fetch failed". Each of the
- five reasons now says what it is: nothing is listening, the name did not
- resolve, it took too long, the address answers but serves no model list, or it
- wants a key. An address typed without http:// is completed rather than
- refused, on every one of these paths.
-- **The models an endpoint answers with are offered on it.** Fetch wrote the
- catalogue away and the endpoint kept a bare text box, so the button reported
- success and nothing on screen changed. The endpoint's own models are a picker
- now, with free text still there for a model the server does not list.
-- **An endpoint you remove stays removed.** Removing the last extra endpoint and
- saving looked right until the next start, when it was back with the same
- identity, because an empty list was never written. As a consequence a
- perfectly ordinary Custom endpoint also grew a duplicate of itself in the
- endpoint list on every start. Both are gone, and removing an endpoint now
- clears the copy of it the app keeps for routing instead of leaving it behind
- for good.
-- **An extra endpoint can be named wherever a provider is named.** It could be
- made active and handed to an agent, but the fallback chain, the advisor and
- executor, the per-mode overrides and the Code model offered only the built-in
- providers, so an endpoint could not stand in for the provider it was replacing.
- All five offer it now, and the "Custom model..." entry in four of those lists
- is translated instead of English-only.
-
-- **Skales Local can see.** A model downloaded together with its vision file
- answered "I cannot look at images". The engine was being handed the models one
- by one with no way to know which vision file belonged to which, so every model
- loaded as text-only. They are paired now, and the card's size is the total of
- both. A vision model you brought yourself is paired as well, whatever the
- publisher happened to call its vision file.
-- **A model you download is usable straight away.** The engine read the library
- once, at startup, so anything downloaded, imported or deleted while Skales was
- open was invisible to it until the app was restarted.
-- **The local server stops when you stop using it.** It came up when you chose
- it and then stayed up for the rest of the session, holding gigabytes of memory
- for a provider you had since switched away from. Switching away now takes it
- down, switching back brings it up, and a server you started by hand is left
- alone.
-- **Start and Stop are on the Skales Local tab.** The banner said the server was
- not running and offered nothing to press, while pointing at the desktop
- application you were already looking at. A start that fails now shows the
- engine's own last lines with it.
-- **Deleting a local model unbinds it everywhere.** The file went, the routing
- went, and the copy of that choice on the provider card stayed - so a deleted
- model went on being the one a new chat loaded.
-- **The reasoning control works on a local model.** It was inert. Its lowest
- rung now switches the thinking block off for that turn, which on a machine
- producing a few tokens a second is the difference between an answer and a
- budget spent entirely on reasoning nobody reads. On the phone, session titles,
- greetings and memory lines never think at all.
-
-
-- **`/spin` runs when you pick it.** Typing `/spin` and pressing Enter chose the
- command out of the suggestion list and put it back in the composer, and
- nothing else happened - the rewrite only started on a second Enter. Picking a
- command that needs no words after it now sends it, which also brings back
- `/projects`.
-- **Rewrite selection is reachable.** Selecting an answer with a triple-click
- and right-clicking it showed the whole-message menu: the selection technically
- ended below the bubble, so it counted as belonging to no bubble. A selection
- that starts in a bubble and picks up no text from outside it is that bubble's
- selection again.
-- **The model you set as Main in Skales Local is the one that answers.** The
- chat read the provider card ahead of the Chat row of the matrix, so the row
- only ever took effect while the card was empty. The row also says, under it,
- which model the next turn will load.
-- **Extra endpoints behave like providers.** "Fetch Available Models" stayed
- disabled with "Enter an endpoint URL first" on every endpoint after the first,
- because the check looked for the address in the wrong place. Agents can be
- pointed at a specific endpoint too: each one is its own entry in the agent
- editor, stored under its own name, and it is still there when the agent is
- reopened.
-- **All four doors under the Flow composer.** 3D and the Video Editor were
- missing from Studio inside the app; they were never drawn there rather than
- cut off.
-- **Typing in a long chat.** A keystroke in a three-hundred-message conversation
- took 150 ms and re-rendered every bubble to change nothing. It no longer does.
-- **Text size is about the conversation.** The setting scaled code blocks, HTML
- previews and the reasoning panel with it, which is not what it says it does.
-- **The line under an answer is filled in.** How long a turn took, how many
- memories it recalled and how many agents worked on it were shown for turns the
- window ran and never for the ones the server ran, and disappeared on reload.
-- **A goal stops saying it is done when it is not.** A run could declare a goal
- finished while criteria nothing had ever confirmed were still open: only the
- ones naming a file or a command were checked, and everything else counted as
- agreed. An unconfirmed criterion now blocks completion, and when the run and
- the check disagree the goal comes to you with both views and an "Accept as
- done" button, instead of either closing quietly or being nudged in circles.
-- **A finished goal names its evidence.** The closing line said "5 of 5 done";
- it now lists each criterion with what settled it, and accepting a goal
- yourself lists what the check could not confirm.
-- **Projects are reachable when the agent is told it has them.** The system
- prompt named the four project tools outright while the request often did not
- carry them, so the model reported it had no access to projects. They now ride
- in the base set. CHECK CAPABILITIES also states which tools the current turn
- is actually carrying, measured on the request, rather than only what the
- registry has switched on.
-- **Vision over your own OpenAI-compatible endpoint.** Three separate faults on
- one path: a model whose name ends in "-VL" (LFM2.5-VL and the whole naming
- convention) counted as blind, so the picture was removed before sending; an
- answer written as content parts rather than one string arrived as an empty
- reply; and a pasted ".../v1/chat/completions" was turned into a doubled path.
- A picture that cannot be sent now says which switch changes that, and
- "Skales Local" works as a Vision Provider.
-- **A second skill of the same name opens its own page.** The menu entry was
- built from the name while the skill was stored under a de-duplicated id, so it
- pointed at the first skill.
-- **Skill pages are sealed like widget pages.** A skill that returns HTML runs
- without access to the app around it and cannot load from outside sources; a
- blocked resource says so on the page instead of failing silently.
-- **The Custom Widgets page speaks your language.** Its result and error
- messages were English regardless of the language setting.
-- **Text with line breaks reaches the field it was meant for.** Typing put text
- into a page one key at a time, and every modern editor - X, Instagram, Reddit,
- a CMS - handles the Enter key itself, so paragraphs ran together into a single
- line. A 220-character post with six line breaks arrived as 214 characters on
- one line. There is now a step that places text in one go, line breaks and all,
- and it is what the assistant is told to use for anything longer than a search
- box.
-- **Nothing is typed into a field that does not have focus.** Text used to be
- sent to whatever happened to be focused, and the check for where it landed
- came afterwards. The target is verified first, and a page that steals focus
- back gets a refusal that names what holds it instead of a write into nowhere.
-- **Every write is read back, and the readback is the truth.** After writing,
- the field is read as it renders, line breaks included - the old check read a
- form that has no line breaks at all, so a broken write and its confirmation
- agreed with each other. The answer says whether the field matches exactly and,
- when it does not, which line differs first. There is also a step to read a
- field on its own, for when you want to confirm before submitting.
-- **A dialog is no longer dismissed behind your back.** A confirm, an alert or a
- "discard this?" prompt was closed automatically and never mentioned, so the
- assistant only saw the aftermath. Those are now named in the result, and an
- in-page modal is named too, with the choices it offers.
-- **The screenshot shows the action, not the moment before it.** Screenshots
- could be taken before the page had drawn the change, so the picture showed the
- previous state.
-- **Browser and desktop control survive a tight tool budget.** Pressing a key
- and scrolling sat outside every priority tier, and so did all desktop control,
- which meant a small local model kept the tool that fills a field and lost the
- one that submits it.
-- **Desktop control says when it is not allowed to run.** Without the macOS
- accessibility permission every click and keystroke silently did nothing and
- reported success; it now names the setting that fixes it. A click outside the
- screen is refused with the real screen size instead of reported as done, the
- pointer is checked afterwards, and typed text can be read back out of the
- focused field. Multi-line text types line by line, and on Linux the text is
- handed to the keyboard tool directly rather than through a shell that would
- reinterpret it.
-
-- **The keyboard can see where it is in Studio.** Tabbing into the Flow composer
- or the project search moved the cursor there and changed nothing on screen:
- both fields switched the focus ring off and never put anything back, so a
- keyboard user had no way to tell which field they were typing into. Both now
- show the same ring every other field in Skales shows.
-
-- **Shutting the computer down is not a crash.** Logging off or restarting
- Windows killed the Skales server mid-session, Skales immediately started a
- replacement that the ending session could not let start either, and both
- deaths were written down as crashes, with a "Server stopped" box on a machine
- that was already going away. A session that is ending is recognised now:
- nothing is restarted, no error box appears, and the entry is listed under
- ordinary shutdowns. Exit codes in the report are spelled out in words,
- including for reports written before this build.
-- **A question from Skales stays answerable.** The card that asks you something
- was only live while it was the last thing in the conversation, so anything
- landing behind it, a message you typed, a follow-up run, a tool line from the
- same turn, killed its buttons for good and re-asking never helped. A question
- is live now until you answer it, in the chat and in the Code window, which
- read the same rule from one place. A turn that both draws an image and asks a
- question now asks it instead of ending on the image, and the idle prompt steps
- over an unanswered question instead of starting a second run under it.
-- **Finishing a Google connection says what happened when it fails.** The screen
- answered "Failed to fetch", which is the browser's words for "Skales was not
- there", and it could hang on "finishing" indefinitely. It now says that
- nothing was connected and nothing was lost and where to start again, keeps the
- raw line underneath for a report, and gives up after a minute with its own
- sentence.
-- **Crash recording is on in release builds.** The diagnostics report said
- "process crash recording is OFF in this build", and it was right in every
- release build there has ever been: the handlers hung on a hook that the
- pinned Next version does not run. They are installed from the heartbeat now,
- which runs in every packaged build. It writes to a local file and sends
- nothing.
-- **The bug report tab speaks your language.** Any status the server sent that
- this build did not know fell back to hardcoded English, "resolved" was not
- known at all, and a status written with more than one hyphen was never
- matched. All of them are translated now, including the words the server uses
- for the same state, and the status colours come from the theme instead of
- fixed hex values.
-- **Five whole screens stop answering in English.** Hugging Face provider
- settings, Spaces, Datasets, casting and the DeepSeek thinking controls were
- never translated, so 32 buttons and labels in the provider settings alone
- stood in English in all eleven other languages. 733 strings were translated,
- and a check now catches a value that is identical in all twelve languages so
- the next block cannot slip through the same way.
-- **The error page's retry button reloads instead of crashing.** Pressing it
- mounted the router a second time in the same document, which is the one thing
- it cannot survive, so the recovery from an error was a second error. The Iris
- screen carried the same class of unguarded read and no longer does.
-
-- **New starts a new chat, from both buttons.** The plus in the header opened the
- new-chat page while New in the conversation list emptied the conversation you
- were in without leaving it, and an emptied conversation had nothing to show:
- a screen-high blank that read as a crash. Reloading appeared to fix it, which
- only hid it, because the reload brought the old conversation back. Both
- buttons lead to the same place now, deleting the conversation you are in does
- too, and a wait long enough to notice shows a loading state instead of an
- empty screen.
-- **The feedback box stops cutting long notes in half.** Everything past 4.000
- characters was silently thrown away as you typed, so a pasted paragraph
- arrived halved and nothing said so. The box now declines the extra characters,
- and once a note is long enough to be near the limit it says how many are left.
-- **The mood says one thing, not two.** The word above the companion's mood bar
- and the highlighted line in its legend were computed from two different
- thresholds, so the panel could say "low-key" over a legend marking "good".
- There is one threshold now, and the sentence, the legend and what Skales
- itself is told all read it. The words changed with it, from instrument
- readings to how the work is going: rough, uphill, steady, good, flying.
-- **Settings points at a page that answers.** Under Settings, Memory, the line
- offering to shape your companion's character led to the Memory page but not to
- the section it meant, and with the switch off that section answered with an
- empty card, so the trail ended in nothing. The section is a target now, it
- says what is on it, and it keeps the character form whether the switch is on
- or off.
-- **Making a spreadsheet works in the installed app.** The tool has never once
- worked outside a development tree: the library it uses looks for its file
- access at the moment it writes, and that does not survive being packaged, so
- every attempt failed with a message about the folder not being writable. The
- folder was fine. Reading spreadsheets had the same fault.
-- **A file asked for in a group arrives in that group.** `send_telegram_file`
- always sent to the paired chat, so a table someone asked for in a group
- landed in a private conversation instead. It goes to the room that asked.
-- **Two quick messages on WhatsApp are one conversation, not two.** Nothing
- serialised them, so a second message while the first was being answered
- started a second agent running on the same conversation at the same time.
-- **The queued-message banner hands back everything it is holding.** Pressing
- Send on it put the first queued message in the box and quietly dropped the
- banner with the rest still in it.
-- **A file card stops offering a file that is gone.** The card is drawn from
- what the turn recorded, so a file deleted or moved since kept a full card and
- a Download button that answered with an error. Every card checks again.
-
-### Changed
-
-- **The free end of the model list is on the desktop too.** The three smallest
- models were listed for phones only and were therefore invisible on a computer,
- which is where most people try Skales Local first. The entry point is now
- 530 MB instead of 2.5 GB, and it thinks, calls tools and reads images.
-- **One row per model.** The same model appeared twice in the library under two
- names, once with its capabilities and once without.
-
-- **Messages you send while Skales is working are answered together.** Each one
- used to become a turn of its own, so three impatient messages meant three
- answers to three halves of a question, and the wait multiplied. They arrive
- as one message now, in the order you typed them, separated by a blank line,
- with nothing added and nothing numbered. The chat window, a run on the
- server, Telegram, WhatsApp and the phone all do it the same way, and Telegram
- no longer sends a "position in queue" card per message.
-
-## v12.8.1 - Rewrite
-
-Two kinds of text get looked at in this release. The first is what a model
-leaves in its own answers: the characters nobody can see and the typography that
-reads as machine-written now come off, if you ask for it, and `/spin` writes a
-text again in a plainer voice. The second is what Skales carries about you,
-which is now something you can see, shape and delete: the assistant has a name,
-a character you set with sliders, and a working history that grows out of the
-work you actually did together.
-
-Underneath it, shell works on a local model again, remote access can ask for a
-second factor, autonomous workers can be your own agents, the local server
-starts when you have chosen it, and the interface finished a long consolidation:
-the controls, the type scale, the colours and the window chrome now come from
-one place across the whole app.
-
-- **Check-ins arrive again when the only thing stopping them was the mute.**
- Friend Mode read "on", named a channel and stayed silent, because muting live
- notifications quietly stops every scheduled check-in and nothing on the Friend
- Mode page said so. It says so there now, with the way to the switch. The Test
- button deliberately ignores the mute, so it also says when it just did -
- a green test over a blocked schedule sent people looking in the wrong place.
-- **One notification is one notification.** A recurring reminder such as the
- morning greeting only started its own cooldown once a channel had actually
- delivered it, so while anything held delivery back - the mute, quiet hours, a
- frequency setting - it was recorded again on every heartbeat. A single
- greeting could fill the entire Notifications page and push everything else
- off it. The cooldown now starts when the notification is recorded.
-- **Quiet hours and the frequency setting reach the in-app toast.** Toasts and
- their sound are a live channel, but they were rebuilt from the durable record
- and only ever consulted the master mute. A category set to "Once" still
- popped every time, and a notification during quiet hours still chimed at
- night, which left the master mute as the only switch that really worked. They
- follow the same settings every other channel does. The Notifications page
- still lists all of it.
-- **Proactive notifications no longer double their emoji.** Seven of the eight
- carried it twice ("☀️ ☀️ Good morning!"): it was part of the text and added
- again on display.
-- **A setting changed in the notifications popup stays changed.** The popup
- opens over the Settings page, so changing a Friend Mode option afterwards
- wrote back the older copy of your preferences and silently undid it.
-
-### Added
-
-- **Skales Visuals is something you can ask for.** The generator that designs a
- poster, a title card, an infographic or an animation as a real page was only
- reachable from the Studio screen, so asking the chat for a graphic got you an
- apology. It is a tool now: ask in the chat or in the Code window, and the
- finished design is drawn in the answer and saved to the same Studio gallery a
- Studio-made one goes to. Asking for a change ("wider", "more colour") changes
- that visual instead of designing a new one. It needs no image provider and no
- key.
-- **3D, and the difference between the two kinds.** A three.js scene - lit,
- turning, drawn in the answer - now comes out of the same visuals path, and
- three.js is in the app, so no page has to fetch it. Separately, a real `.glb`
- model file can be made from a description or from a picture through Replicate,
- fal.ai, Meshy or Tripo, saved to your workspace, and turned and zoomed in the
- chat and in Studio. Skales says which of the two it is giving you: a scene is
- a picture, a model is a file you can open in Blender. Where WebGL is not
- available the page says so instead of going black.
-- **Video Editor, stage one: the cut.** A timeline with a video and an audio
- track, clips from the Studio gallery and from your workspace, trim, reorder,
- delete, and export. It goes through the renderer Studio already uses - there
- is no second encoder - and an export that was interrupted comes back as
- finished or as stopped, never as a bar that says "running" for ever.
-- **3D is something you can pick in Flow.** The composer has a "3D scene" mode
- next to Deck, Prototype and the rest: pick it, describe the scene, and Flow
- writes a real three.js page that the preview draws and turns. There is no
- provider and no key involved - three.js travels in the app - and a brief that
- says "3D" lands there on its own when the mode is left on Auto. A downloadable
- `.glb` model file is still the other thing, and still needs a provider; the 3D
- tile now says which of the two is which.
-- **Two more doors in Flow.** 3D and the Video Editor sit beside Lio and Studio
- Classic, in the same tiles.
-
-### Changed
-
-- **The Widget AI form recommends instead of warning.** It opened with a red
- panel about an alpha phase, a corrupted setup and a backup to make first, for
- a feature that has been shipping for months. What is left is the part that was
- ever useful: an orange note that a premium or a coding model gives the better
- result. Twelve languages.
-
-- **The video editor says it is in beta.** It carries the same amber badge
- Skales Local wears, both on the door in Studio and on its own heading, so
- what it is is clear before you walk in rather than after.
-
-- **The character step no longer uses an internal name.** Setting up your
- agent's character said "Form your Zenit", which is what the mechanism is
- called in the source and means nothing to anybody else. The heading already
- says what the step is for, in all twelve languages.
-
-- **Work that was invisible reaches Discover and the usage figures.** Designing
- a visual from the chat, making a 3D model and exporting a cut from the video
- editor all did their work and then appeared nowhere, while the same work
- started in Studio did. A designed graphic and a 3D model also stopped being
- filed as "images generated", which they are not: each has its own name in the
- feed now. Opening Analyze and answering a question card are counted as usage
- only and never reach the feed - what was asked, what was answered and what a
- session spent stay private.
-- **The media tools ask what you have set up instead of announcing an order.**
- `generate_image` used to state its own cascade in its description - local
- ComfyUI, then A1111, then Skales Local, "otherwise Gemini or Replicate" - and
- Replicate was not in that sentence at all. So somebody with a Replicate key
- who asked for a picture was told about Gemini. And the tool that should answer
- "what can I make here" only knew about local backends, so a machine with cloud
- keys and no ComfyUI answered with an empty list, which reads as "not
- possible". There is one place now that works out what this machine can do -
- local backends, the built-in engine, every configured provider, Skales IQ and
- an endpoint you added yourself - and every media tool asks it. Name a provider
- you have and that one is used; name one you have not set up and Skales says
- so with the way to set it up, never a substitute and never "not possible".
-- **three.js and GSAP travel with the app on the computer too.** The generator
- used to hand the model two CDN script tags. They are in the page before it
- runs now, on both the computer and the phone, and `cdnjs.cloudflare.com` has
- left the list of places a generated page may reach - on both.
-
-- **Analyze: what a run spent, and whether anything went wrong.** Right-click any
- conversation in History or in the chat sidebar and pick Analyze, or click the
- spend bar under an answer or the context figure under the composer. A wide bar
- shows where the tokens actually went - system tools and system prompt, the
- conversation, what came back from the provider's cache, thinking, and the
- answer itself - and the fixed prefill is usually the largest slice by far,
- measured at the real payload rather than estimated. Under it the run is a log,
- one row per turn, foldable: the model, where the turn came in from, the tools
- in order with their arguments and results, the thinking text as thinking, the
- providers that failed before the one that answered, and a voice note with its
- transcription and the honest statement of where it was transcribed. Every turn
- carries a verdict derived from signals the run already recorded: ran clean, a
- provider failure, a model failure, or a failure of Skales itself, which gets
- its own tone rather than being filed under the model. Where the signals cannot
- separate the provider from the model, it says the cause is not clear instead of
- guessing. There is no grade, no score and no percentage. **None of this calls a
- model, and nothing leaves the computer** - it reads what the run already wrote
- down, so opening a report costs nothing. What a provider never reported is
- shown as not reported, which is a different thing from zero.
-- **A typeface for the chat bubbles.** Settings, Appearance, above Text size:
- Inter, Lora, Comic Neue or Caveat. All four travel inside the app, so the same
- choice looks the same on every machine. It applies to the text of the bubbles
- and to nothing else - code, previews, tool cards and the thinking lid keep
- their own type, and code stays monospace under every choice.
-
-- **Watermark: what comes off an answer before you read it.** A new section in
- Settings -> Chat & Code, off by default. **Remove invisible characters** takes
- out the zero-width spaces, text-direction marks and unusual spaces a model
- leaves behind, and changes nothing you can read. **Neutralize typography** is a
- separate switch because it edits visible text: the long dash becomes a comma
- and the one-character ellipsis becomes three dots, in Chinese and Japanese too,
- where the dash carries no spaces and a rule written for English prose finds
- nothing at all. **Straighten quotes** sits under it. Code blocks and inline
- code are never touched by any of them. What this does not do, and does not
- claim to do, is remove a statistical sampling watermark: that lives in the
- choice of words.
-- **Rewrite, with `/spin`.** `/spin <text>` writes a text again in a plainer,
- more human voice; `/spin` on its own does that to the last answer, and the
- message menu offers the same action next to Copy. Pick the model that does it
- in the same Watermark section, or leave it on the one that is already
- answering; a local model keeps the text on your machine. The result always
- goes through the invisible-character pass, whether or not the switches above
- are on, because a rewrite model stamps its own markers in like any other.
-- **Your assistant can have a name, and a character you set.** Settings ->
- Memory -> Companion, then **Shape your companion** on the Memory page: give it
- a name, pick a starting point (Companion, Colleague, Mentor, Sparring partner,
- Trickster, Quiet type), and move seven sliders - serious to playful,
- diplomatic to direct, dry to funny, polite to cheeky, speculative to grounded,
- and two new ones, businesslike to warm and answers-only to asks-back. All of
- it shapes how it talks to you in every conversation. A name is what you call
- it; it never claims to be a different model, and it still answers truthfully
- when asked what it runs on.
-- **A working history that grows.** With the companion switched on, Skales
- counts the days you worked together, the messages and the goals you finished,
- and uses that to drop the introductions and use your shorthand instead of
- rebuilding context every time. It counts shared WORK and says so: it never
- claims to have missed you or to be fond of you, and a test walks every state
- it can reach looking for exactly that sentence.
-
-- **A second factor for remote access.** Settings -> Security -> Remote access
- can now ask for a six-digit code from an authenticator app (Google
- Authenticator, Authy, 1Password) when a browser opens a remote session, with
- ten one-shot recovery codes for a lost phone. It is off until you set it up,
- and it is deliberate about what it covers: a browser signing in with the
- access URL is asked, and programs that send the token as a header - the mobile
- app, the relay, a headless instance, swarm peers - are not, so nothing that
- works today stops working. The card says that in as many words rather than
- implying more.
-- **Dispatch from your own agents.** Settings -> Autonomy -> What a sub-agent
- starts with has a new option, **Pick from my agents**: tick the agents that
- may run as dispatched sub-agents and Skales chooses among those by name. Each
- brings its own instructions, skills and pinned model; the tool limits every
- autonomous worker has still apply to it unchanged.
-- **The role presets are readable.** The same setting can now show, word for
- word, what a dispatched worker starts with: the tools every role gets, what
- the chosen role adds, whether it may dispatch workers of its own, what it can
- ask for at runtime, and what is never available to it.
-- **Interests and mood.** One switch in Settings -> Memory, off by default.
- Skales tracks the topics you keep coming back to - at the level of the topic
- rather than the conversation, and down to which part of a topic - and carries
- a working mood between sessions. Both feed the messages it sends on its own,
- so a briefing or a Friend Mode note reads like it comes from someone who knows
- you. The mood is about the work, never about you, and it appears in a reply
- only where it fits.
-- **Skales Local starts on its own** when this machine is the one you chose to
- answer: it is your active provider, or a row in the matrix is set to Main or
- Fallback. A chat that goes to a local model brings the server up if it is not
- running. Nothing points at this machine, nothing starts and no memory is used.
- **Start it on its own** at the top of the tab is the switch.
-- **Rewrite a passage, not a whole answer.** Select any part of a reply, right
- click the selection, and the menu acts on the passage instead of the message:
- **Rewrite selection** first, then copy it, quote it, read it aloud or save it
- to a document. The gesture was already there and reached nothing; picking a
- paragraph out of a long answer no longer means retyping it after `/spin`. A
- selection dragged across several bubbles belongs to none of them, so the menu
- offered on each stays the one for the whole message.
-- **Diagrams, drawn.** A ```` ```mermaid ```` block in an answer is now rendered:
- flowchart, sequence, state, class, entity-relationship, gantt, timeline, pie
- and xychart, in your accent and your theme, with the source one click away and
- a download as SVG. For a process, a structure or a comparison this is the
- cheapest good answer there is - four lines instead of a whole generated page -
- and unlike a page it is also the form a small local model gets right. The
- renderer is in the app; nothing is fetched.
-- **Code blocks have colour.** Comments, strings, numbers, keywords and markup
- tags are tinted in every fenced block in the chat, in about fifty languages,
- using the same tokenizer the Code window's review panel already used. The five
- colours are measured against the surface they sit on rather than chosen by
- eye, and a gate holds them above the readable floor.
-- **A page in an answer is shown as a page.** The chat drew a live preview only
- when the model wrote exactly `html` after the backticks. A block tagged `htm`,
- `svg`, `xhtml` or `html5`, or one that simply begins with a document, arrived
- as grey monospace - which is why the same answer from the same model was a
- drawn page on the phone and source code here. The rule is now the phone's, and
- a test compares the two so they cannot drift apart again. A snippet that
- merely mentions a tag stays code, and a `text` or `xml` fence is still the way
- to hand over markup you want to copy.
-- **Skales knows what its own chat can do.** The system prompt described tools
- and providers and never the renderer, and the one sentence that said "put a
- page in a code block and it is drawn" sat in Code mode and nowhere else - so
- whether an explanation arrived as a picture or as three paragraphs came down
- to the habits of whichever model was answering. It now carries the live
- preview, the diagram, the formula and the colouring in every mode, with a test
- that checks each claim against the component that has to honour it.
-
-- **A widget can be an application now, not just an answer.** Custom Widgets
- have been in Skales since v1, but a widget was an `execute()` function whose
- return value got drawn in one of four shapes - a gallery, a table, a block of
- HTML, or text. Nothing it produced survived being looked at twice, so the
- obvious thing to build, a notes app that keeps your notes, could not be built.
- There is now a fifth shape: a page you wrote, running in its own sandbox, with
- a store of its own. The four existing shapes are untouched and every widget
- built before this one still works exactly as it did.
-- **The store a widget keeps.** A widget gets `window.skales.storage` with four
- calls - `get`, `set`, `remove`, `keys` - and that is the entire surface it gets
- from Skales. Deliberately not `localStorage`: that lives in the page's own
- origin, so it disappears when a cache is cleared, sits outside the folder a
- backup walks, and cannot be put back on import. This store is a file next to
- the widget's page, which means your entries survive a restart, travel in an
- export, and come back on the other side.
-- **Widgets travel.** A backup now carries a `widgets/` section - the folder, the
- page and the store, per widget - and it is listed as mobile-compatible, so a
- widget built on the computer opens on the phone and back again. An import that
- brings a widget you already have never writes over it: the widget on this
- machine is left exactly as it is, the incoming copy is kept beside it, and the
- import names both. The store is your entries, and replacing it silently would
- be the expensive kind of quiet.
-- **The house typography.** Twenty families across the six directions the design
- rules ask for - display, serif, grotesk, condensed, mono and script - now ship
- inside the app as subset woff2, and only the ones a page actually names get put
- into it. A generated page gets its typeface instantly, offline, and identically
- on both platforms. Naming a different Google family still works; the bundle is
- the guaranteed path, not the only one.
-
-### Fixed
-
-- **Analyze opens on the conversation you are actually in.** Clicking the spend
- bar under an answer, or the context readout under the composer, said there was
- no conversation to analyze - while you were sitting in one. Only the
- right-click in the session list worked, so the same conversation had a report
- by one route and nothing by the other. Both of those two ways in ask for
- "whatever is open right now", and the answer to that question was being read
- off a yes/no flag, which could never name a conversation. All three ways in
- reach the same report now.
-
-- **Analyze and the figures under the answer are visibly the same numbers.**
- They always were: the first three parts of the bar are the input the provider
- counted, the last two are the output. Nothing said so, so a breakdown reading
- 24k / 477k / 4.5k next to a bubble reading 501K in / 4.5K out looked like two
- different measurements with no way to tell which one to believe. The
- distribution now writes both totals out and says they are the same two
- figures.
-
-- **A conversation that came in over WhatsApp, Telegram or Iris says why it has
- no figures.** It was being told it had been written before Skales kept usage
- figures, which for a conversation from this morning is simply untrue. A
- channel hands Skales a message and takes the answer back and reports no usage
- of its own; that is what it says now, and it names the channel.
-
-- **The full-screen panels no longer sit on the window buttons.** Analyze,
- History, Group Chat and the rest opened flush against the top of the window,
- which on macOS is where the close, minimize and zoom buttons are drawn. People
- reached for what looked like the panel's own controls and quit Skales. The
- panels start below them, and so do the X and "Back to Flow" buttons in the
- Flow window, which were pinned four pixels from the same edge.
-
-- **Port selection tests the address the server will actually use.** The check
- that walks ports 3000 to 3009 bound the loopback address while the server
- itself binds every interface whenever remote access or Swarm is on. On macOS
- those two can coexist, so the check reported the first port free, the real
- start died on it, and the nine spare ports were never tried. With remote
- access off it was the quieter half of the same fault: a second copy of Skales
- could take over the first one's address in silence.
-
-- **The token readout under the composer can be read again.** Hovering it raised
- the operating system's own tooltip directly over the card carrying the exact
- figures, and the card could not be moved onto with the mouse. The tooltip is
- gone, what clicking does is written in the card itself, and the card can be
- hovered and its numbers selected.
-
-- **The spend figures appear at the bar, not under the pointer.** The model line
- under an answer and the spend bar in the bubble explained themselves through
- the operating system's own tooltip, which is drawn about twenty pixels below
- the cursor and never beside the thing it explains. Both now open the same
- hover card the rest of the composer uses, anchored above the figure.
-
-- **The whole context card opens the report it offers.** It ends with "click for
- the full report" and was the one part of the readout that did nothing when
- clicked; only the "... context" line underneath worked. The card is the button
- now, and the figures in it can still be selected and copied.
-
-- **"Code" in the menu no longer glows.** It carried a ring that faded in and out
- every few seconds until the entry had been opened once, so on a machine where
- nobody opens Code it never stopped, and it read as a fault rather than an
- invitation.
-
-- **The four Flow doors are one shape.** Lio AI, Studio Classic, 3D and Video
- Editor were laid out one, three or four across depending on the window width,
- with a title long enough on one of them to make it a different height from the
- rest. They are two by two, the same size, with the same amount of text. The Lio
- door is called Lio AI; what it does is the line underneath. 3D wears the beta
- badge alongside Video Editor, because it can do about as much.
-
-- **The personality questions stop after one.** Five to seven "a quick question
- for you" cards could land in a single conversation, each with its own toast, and
- a card dropped into an old conversation while you were away marked it unread.
- There is one per conversation now and one announcement with it, the mark is
- gone, and the ordinary clarifying question the assistant asks while working -
- which is not this feature at all and was raising the same notice - is quiet.
-
-- **A long conversation costs less to draw.** The spend breakdown behind each
- answer's bar was recalculated for every answer in the conversation on every
- redraw, and a conversation redraws on every word of an answer being written.
- It is worked out once per answer now.
-
-- **The chat typeface reaches the answer, not only your own message.** Choosing
- a face changed the text you wrote and left everything Skales said in the
- default one, which made the setting look half-broken. An answer that carries
- reasoning - which is every answer from a thinking model - was drawn outside
- the place the choice applies. Both kinds of answer follow it now, at the size
- that belongs to the face. The reasoning lid deliberately stays in the default
- face: it is not the answer.
-
- The same setting was still stopping at your own message whenever the answer
- had run a tool, which in practice is most answers: an answer with tool results
- and an answer waiting for approval are drawn by a different piece of the chat,
- and neither of them knew about the choice. All of them follow it now. The size
- is a multiplier, so no bubble is allowed to sit inside a second one and apply
- it twice.
-
-- **A designed page that animates in is visible even when the animation never
- starts.** A poster or a visual is usually written to fade and slide into
- place, which means every element begins at zero opacity. When whatever was
- supposed to bring them up did not run, the result was not a page missing its
- motion - it was a blank rectangle, impossible to tell apart from a page that
- failed to generate. Every page is now watched as it opens: if nothing at all
- is on screen shortly after it loads, its opening is triggered a second time,
- and if it is still empty after that, it is shown in its finished state. Pages
- made earlier are covered too, because the check happens when a page is shown
- rather than when it is written.
-
-- **A question in the Code window is answerable again.** A question with no
- options to pick from drew no field at all, and the send button demanded a
- choice for every question - which a free-text question can never give - so the
- card sat there greyed out and the composer was the only way past it. There is
- now a proper field wherever there is nothing to pick, every option list ends
- in a "Something else" escape with a field behind it, and a typed answer counts
- as an answer. With more than one question on the card, the footer says which
- one the number keys are currently pointing at and that question is marked, so
- a 2 is never a guess. The composer still answers everything, as before.
-- **The answer to a question card looks like the answer to a question card.**
- What you picked used to come back as a numbered list in a plain message
- bubble, so the question was a card and the reply to it was a paragraph. The
- reply now draws as the pair to the card, one row per question with the choice
- under it. Nothing changed about what is sent: the model still receives the
- questions and answers in full, in plain text, which is the only record of the
- exchange once a conversation has been compacted.
-- **Shell survives every tool budget.** On a local model, `execute_command`
- could be trimmed out of a turn while the first line of the system prompt still
- promised a shell, so the model reported - accurately - that it had no shell
- tool. The shell family is now protected alongside the file tools, and the
- Shell line only appears when the tool was actually sent.
-- **A local model is no longer planned against a window four times too small.**
- The per-turn budget assumed 8192 tokens of context for any local runtime, even
- though LM Studio reports its real window and a custom endpoint was not in the
- table at all. Measured on the same catalogue, LM Studio went from 7 tools per
- turn to 88 and a custom endpoint from 7 to 120. A real rate limit still binds:
- a key on an 8000-tokens-per-minute plan is planned for exactly as before.
-- **The trim note points at the setting that did the cutting.** Three different
- limits can shorten the tool set and the note always named the local-model
- slider, which changes nothing for two of them.
-- **The IMAP account test works.** Testing an account against a server that
- sends folder names in modified UTF-7 (Yandex among them) failed with
- `allocateBase64Buffer is not defined` from inside a minified chunk. The
- library it comes from is written for plain CommonJS and was being bundled into
- a strict-mode chunk, where the helper it needs stops existing. It now loads as
- what it is.
-- **The IMAP and SMTP test buttons always answer.** A server that accepted the
- login and then went quiet left the button spinning with no error and no
- result; both now stop after 45 seconds and say what happened.
-- **A custom agent no longer runs on a model its provider does not have.** An
- agent that stored a model id while one provider was active kept it after a
- switch, and the id travelled to an endpoint that never heard of it. The stored
- id is now checked against the provider's own catalogue, and only a catalogue
- that actually answers can replace it.
-- **Skales Local failures say what they are.** A local server that is not
- answering, a machine that ran out of memory loading a model, and a model that
- cannot read images are three different sentences with three different next
- steps, instead of a transport error.
-- **AIPointer's switches are reachable by keyboard and named for a screen
- reader.** They were neither.
-
-- **A goal now remembers the files it produced.** The ledger has always had a
- place for them and nothing ever wrote one, so the list was empty everywhere it
- was read: the checker's evidence, the "Produced:" line of a distilled strategy,
- and the durable half of the did-any-work test. A run could write a file, have
- its raw turns compacted away, and then have nothing left that remembered the
- file existed. Every successful write is recorded on the goal now.
-
-### Changed
-
-- **Code in a chat takes its colour from the theme.** The dark surface behind a
- code block was written into three separate places as a fixed value, so no
- theme could reach it. It is one setting now, next to the five colours the code
- itself is written in - and those five are measured against it, so a theme that
- moves the surface cannot leave the text unreadable on top of it.
-
-- **Interests and mood moved to Settings, Memory, and what they learn is
- visible.** The switch sat in General under a heading about units and the
- weather, and the settings search could not find it: looking for "emotion"
- returned nothing. It is a section of its own under Memory now, where it
- belongs, and the Memory page shows every topic Skales has noticed, how often
- it came up and which narrower part of it recurred. Each one can be deleted on
- its own, all of them at once, and the mood and the working history each have
- their own reset. Deleting is real deleting: a removed topic only comes back if
- it comes up again. A memory you can neither read nor erase is not a feature.
-- **The chat's popups open full width.** Agents, Group Chat, Teams,
- Organization, History, Projects and Add-Ons opened into a dialog the width of
- a document, which is about half of what a team roster or an org chart was
- drawn for. They now open in the same frame the Cockpit uses, and History is
- one of them rather than a navigation away from the conversation. The width
- comes from the host, so the same screen reads full width in the popup and as a
- document on its own route.
-- **One interface, from one place.** Over several passes the app finished a
- consolidation that had been running for a while: every toggle, select, slider
- and API-key field now comes from one shared component, the small font sizes
- have names instead of arbitrary pixel values, status colours and accent
- colours come from tokens so all six themes are correct, every document page
- shares one content width, and window chrome and the brand typeface are the
- same on every surface, including the splash, error and OAuth screens. Studio
- speaks the same base as the rest of the app again.
-- **Two switches showed two states at once.** A stylesheet rule left over from
- an older kit drew a second, fixed knob under every shared toggle; on, the two
- drifted apart.
-- **Slider legends sit on their real value.** A three-mark legend was spaced
- evenly regardless of the scale, so "5 min (default)" sat in the middle of a
- one-to-sixty-minute track where half an hour actually is, and the last mark
- stood past the end of the track it labelled.
-- **One background per surface.** The chat landing page, Wrapped, the DevKit
- docs page and the sidebar popup each laid a second background over the app.
-- **Clean up under Skales Local actually deletes.** It skipped anything that was
- not a file and reported success.
-
-- **Model-written pages have a stated network posture, and it is the same one on
- both platforms.** The preview iframe in chat had no content policy at all, so a
- generated page could load from anywhere. It now carries one, and it allows
- exactly the three places the generator prompts themselves point at:
- `cdnjs.cloudflare.com`, Google Fonts, and `images.unsplash.com`. Everything
- else is refused, and the refusal is printed on the page with the address that
- was tried, rather than failing as a blank space that reads like a bug in the
- page. A Custom Widget gets a stricter posture still: no network at all, on
- either platform, so a widget behaves the same wherever you open it.
-- **The widget builder can build the new shape.** Widget AI now knows about pages
- with state: it is told the four storage calls, that it has no network, and
- which typefaces are already loaded. Rebuilding a widget keeps its folder, so a
- rewrite of the page is not a way to lose the entries you had.
-
-## v12.8.0 - Sightline
-
-The sidebar is one list again, the surfaces that had grown into a menu of their
-own are back where the work happens, and what Skales knows about you can be kept
-off the network with one switch.
-
-### Added
-
-- **One sidebar, and it belongs to what you are doing.** Home lists Dashboard,
- Chat, Code, Studio, Cockpit, Planner, Iris, Memory, Mobile, then Discover and
- Wrapped, then System. Open a conversation and the same column becomes the
- chat's own navigation: New chat, Agents, Cockpit, and a More group holding
- History, Projects, Teams, Group Chat, Organization, and Add-Ons and Skills.
- The shortcuts open as overlays over the same screens the routes show, each
- with its own title bar, so nothing sits under the close button any more.
-- **Cockpit: everything that is running, in one place.** Goals, Tasks and
- Schedule as three tabs of one screen, with the tab in the address so a reload
- brings you back where you were. Goals has a screen for the first time: what
- each goal is trying to do, the criteria it must meet and the evidence it has
- collected, its last steps with the tool each one used, what it produced, what
- it learned, and what it has spent, with Open, Continue and Stop on each one.
- Finished goals are listed too, and so are the ones on a repeating schedule.
-- **Studio opens Flow.** Clicking Studio puts you straight into the workspace
- in its own window instead of a screen asking which half you want. The whole
- of the older Studio is still there, one door below the composer as **Studio
- Classic**, on the tab you were linking to, with a Back to Flow line at the
- top of its rail. A second door starts a prototype in Lio AI, and that page
- has its way back too.
-- **Planner has its own entry again.** It is the one visual calendar surface
- and it survives a fresh install.
-- **A Codework mode in Code.** A pill in the composer starts the session with
- the file and preview panel and the review panel already open, so the chat,
- the files and the diff stand side by side from the first message. Pressed or
- not pressed is remembered like the terminal and the panel widths, and it
- shapes a session when it starts rather than rearranging one mid task.
-- **Code takes a typed project path,** so a folder can be opened without the
- native picker, and the server validates it the same way either way.
-- **Code drafts its own commit message.** A Draft message button beside Commit
- reads the staged diff and writes the message; a repository without a first
- commit is handled rather than failing on a missing HEAD.
-- **Chat and Code can see your Projects,** and can keep their own working notes
- in one instead of scattering scratch files.
-
-### Changed
-
-- **Privacy Mode: what Skales remembers about you never leaves the machine.**
- One switch. With it on, a cloud model gets no memory index, no name, no
- facts, no preferences, no interests and no learnings, and the prompt says
- plainly that a memory exists and is being kept local rather than pretending
- there is none. Your language and the clock survive, local models are
- unaffected, and the background briefing follows the same rule. Which
- provider counts as local is decided by the address it points at, never by
- its name, so a custom slot on a rented endpoint is treated as the cloud it
- is.
-- **The prompt is budgeted before it is sent,** so a model with a small context
- window is given a tool set it can actually hold instead of one that pushes
- the conversation out, and it is told what is not currently loaded rather than
- being left to guess.
-
-### Fixed
-
-- **A goal is only finished when its criteria are met.** The chat used to be
- able to write a goal off as done past the gate the runner honours. Both paths
- go through the same gate now. Accepting a result may overrule the criteria,
- but never a sub task that is still running, and a refused accept leaves the
- goal parked and says on the card why.
-- **A completion marker never appears in the answer.** The word the runner uses
- to signal it is finished is stripped from what you read, on every surface.
-- **Updates that stall say so.** The updater notices a download that stopped
- moving instead of leaving a progress bar sitting at a number forever.
-- **Unsloth Desktop, and every other provider, gets the address it documents.**
- Fetching models from a cloud provider whose default address carries no path
- appends the version segment those providers resolve to anyway, and the bare
- address is still tried second. A conformance check now walks the whole
- provider registry so one card can no longer build its URL differently from
- the rest.
-- **The Obsidian theme has all of its menu.** Its dropdown had been dropping
- every entry that was not in a group, which since the sidebar was rebuilt
- meant Agents, Tasks, Schedule and Notifications had no door at all.
-- **Settings opens on the section a link names,** instead of landing at the top
- of the page and leaving you to find it.
-- **New chat no longer reloads the history list** every time it is pressed.
-
-## v12.7.45 - SKLS
-
-A repair release, built the day after 12.7.4 from what the first devices
-reported. The local server that the release was named for now answers on the
-machines it was promised to, and a handful of screens that guessed, hid or
-misnamed their failures say what actually happened instead.
-
-### Added
-
-- **Unsloth Desktop is a provider with its own name.** It has a tile in the
- provider grid with the same on and off switch every other provider has, an
- address already filled in, and a button that finds a running instance on
- either of the two ports it uses. It is the first local runtime that wants an
- API key, so the card says where that key is made, and a refused key is
- reported as a refused key instead of as a missing model. Refresh Models reads
- the runtime's own GGUF catalogue with the quantisation of each file, and a
- model that is still loading gets the long silent window a local runtime needs.
- How long a model stays in VRAM is set inside Unsloth Desktop, and the card
- says so rather than showing a control that would do nothing.
-
-### Fixed
-
-- **A turn to the local server carries a local model id.** Switching to Skales
- Local kept the previous provider's model id, so the runtime was asked for
- "openai/gpt-5-mini" it could never hold. Both request paths now resolve the
- id from the capability matrix, the provider card or the runtime itself, in
- that order, and a request to a local address can no longer carry another
- vendor's id.
-- **The Mac build for Apple Silicon carries the local server.** The arm64
- package held 253 MB of engines for three other platforms and none of its
- own. Each platform and architecture now stages its own engines, a build
- gate measures the staged files against what the configs promise to ship,
- and every bundled library had its build-machine search path repaired, not
- just the executable. Mac packages shrink by the foreign payload they no
- longer carry.
-- **Fetch Available Models sees what the runtime holds.** The fetch button
- knew one URL and two answer shapes while the error path knew two and three,
- so the card said "0 models" while the chat error listed all of them. One
- reader serves both now, the card offers its own library without a button
- press, and projector companion files no longer appear as answerable models.
-- **Models on another disk are found.** Detection read two default folders on
- the system drive and called the rest of the computer empty. It now reads
- Ollama's configured model path including the Windows user environment, LM
- Studio's own settings, and asks running runtimes what they hold; an empty
- scan names the folders it read and offers picking one instead of claiming
- absence.
-- **One status line, one Start now.** The Skales Local card rendered two
- answers to "is it running", one of them with a dead start button when no
- model was installed. One resolver decides, and the more concrete answer
- wins.
-- **Connecting a Google account survives its own landing page.** The return
- page re-ran its effect the moment translations finished loading and burned
- the one-time code, which read as "This authorization is no longer open"
- seconds after a successful consent. With remote access on, the final step
- was rejected as unauthenticated and surfaced as a bare "Failed to fetch".
- The landing step is reachable on its own now, the code is claimed before
- the exchange, a run lives an hour, and a second tab can no longer erase
- the account the first one connected.
-- **A tool that is named can be called.** The capability answer listed tools
- from the catalog that the running turn had not loaded, and the model,
- unable to find them, tried them as shell commands and invented an excuse
- when that failed. Naming now loads, a tool name typed as a shell command is
- steered back to the tool instead of executed, and the answer says how a
- not-yet-loaded tool becomes callable.
-- **Credentials stay out of transcripts.** The integrations folder is sealed
- against file tools the way settings.json already was, and web requests
- redact tokens and secret fields by name in what they log and store.
-- **Calendar reads with your account, not just your key.** With both an API
- key and a connected account, events were fetched with the key alone, which
- can only see public calendars and read as "google rejected access". A
- rejected fetch retries once with the account token, and writes always use
- credentials that can write.
-- **Saved keys stop eating themselves.** Twenty fields showed a shortened
- key prefix after saving, and saving again wrote that shortening over the
- real credential, after which every check failed. A saved secret shows as
- dots that are never written back.
-- **Replies to bug reports arrive.** An answer from the maintainer was
- fetched and thrown away by the notification poller, unclickable on the
- notifications page and invisible on the feedback page. It now surfaces in
- all three places.
-- **Searching Settings finds every line again.** The 12.7.4 rule that let a
- section-name match win outright also erased every inventory match, which
- made "Google" unable to find the Google account setup inside Productivity
- Integrations, and the same for 144 other word-to-section pairs. Name
- matches sort first now, everything else follows under a divider, and a
- test probes every word of every section name and keyword against the index.
-- **A turn that ends with nothing says what ended it.** A dropped connection,
- a filtered answer, an output limit reached before the first word, a reply
- spent entirely on reasoning and a genuinely empty answer all ended in the
- same sentence with no log line, and the retry button only existed in
- English. Each cause has its own sentence in all twelve languages, one log
- line names it on the server, and the reason and retry survive a reload.
-- **Post generation stops calling a token limit a timeout.** Discover's
- compose gave every model 120 tokens; reasoning models spent them thinking
- and never wrote, and the empty result was labelled "your model did not
- answer in time". The budget starts at 1024 and climbs before reasoning is
- switched off, an answer cut at the limit is named as cut, and a fallback
- template says why it is one.
-- **The diagnosis stops blaming the outside for a stop Skales ordered.** The
- server exit recorded as an external SIGKILL during Discover traffic was
- Skales's own five-second shutdown grace ending a long request; measured,
- not guessed. The forced stop marks itself and the diagnosis names it. The
- same report also claimed crash recording was off while it was on; the
- answer now comes from the installed handlers instead of an env guess.
-
-## v12.7.4 - Skales Local
-
-A release about screens that tell the truth. A connection that worked but read
-as broken, a tool that was announced and could not be called, a question asked
-by the operating system in a way that left the window unusable afterwards. Five
-copies of the same answer do not make it true, so there is one of each now.
-
-### Added
-
-- **A Telegram bot in a group can be told when to speak.** It answered every
- message in the room, which is fine where the room is built around it and
- unusable where people are talking to each other. Settings, Integrations,
- Telegram now offers every message, only when mentioned or replied to, only
- replies, or only commands. Every message stays the default, so nothing changes
- for a bot that works today. Any single chat can be muted without unpairing it,
- and a group that pairs from now on can start muted and is told so when it does.
-
-- **Skales ships its own server.** Skales Local is a provider like any other
- in the list, with one difference: it asks for no key, no account and no
- install. The inference server lives inside the app. A new Settings tab
- carries the models, with search, categories (language, seeing, voice, image,
- imported, downloaded) and sorting by size, name or last used.
-- **Four capabilities, four separate decisions.** Chat, dictation, read-aloud
- and images can each be set to this machine first, cloud first, or off.
- Images local, voice local and the language model through OpenRouter is an
- intended combination, and so is every other one, on every surface that
- needs a model.
-- **Whoever answers is named, before and after.** The model line above the
- composer reads "Skales Local" and the model with it, so the choice can be
- checked before a word is sent. Set a capability to this machine first and it
- is never quietly pulled into the cloud. When a turn does have to leave the
- machine, the answer carries the reason; when neither side can answer, there
- is a sentence and a next step instead of a spinner.
-- **Fetch a model, bring your own, or adopt one.** The download shows real
- progress, resumes where it stopped, and checks the file against its checksum
- before it counts as installed. A repository that wants a licence confirmed
- first says exactly that instead of "download failed". Your own .gguf, .onnx
- or .safetensors files come in through the file picker, and existing Ollama
- and LM Studio folders are recognised and used where they lie: the disk pays
- once. A file that is part of a model rather than a model - a projector, an
- autoencoder, a text encoder - is recognised as one and shown as one, so it
- never turns up in a list of models to answer with.
-- **Dictate and listen with no network.** Whisper for hearing, VITS voices for
- speaking, both through one runtime and one install shared by Chat, Iris and
- AIPointer. Ten of the twelve interface languages have a voice, Croatian
- among them. Japanese and Korean have none we are allowed to ship, and the
- screen says so.
-- **Pictures that never leave the machine.** The local image model sits behind
- the same tool as before, in Chat, in Studio and in Flow, as a backend and
- not as an island of its own. Skales reads the graphics chip's report first:
- if it cannot run the needed kernels, the picture runs on the processor,
- named, instead of crashing.
-- **Every licence is in the app.** Settings, Advanced, Third-party notices:
- the open-source components and the terms of every local model, full texts,
- offline and searchable. A build that is handed a catalog model without its
- licence does not pass.
-
-- **The sidebar becomes the chat's own navigation.** In a conversation, on the
- themes with a full sidebar, it shows New, Agents, Tasks and Add-Ons & Skills
- as overlays over the same screens the routes show, and below them the history,
- grouped by day, pinned chats first. A row says what its chat is doing -
- running, waiting for an approval, failed, a goal met - and keeps saying it
- after a reload, because the state is written to disk. A chat that came in over
- WhatsApp, Telegram, the phone or a task wears that channel's mark; everything
- else wears a quiet bullet. Hold still on a row and it gives you the date and
- time; pass over it on the way somewhere else and it stays out of your way.
- Home and Work stay on top and are the way back.
-- **Setup names the three ways in.** Skales IQ, Skales Local with what it
- costs on disk and its Beta label, or your own key.
-
-### Fixed
-
-- **One dialog asks every question.** Deleting a chat on Windows put up the
- operating system's own box, and after it closed the composer was dead until
- you switched windows and back. Sixty-four of those calls across the app are
- now the Skales dialog, in every window, and the keyboard goes back to the
- button you pressed.
-- **A ChatGPT subscription reads as connected, because it is.** The status card
- asked one question and the chat asked another, so a paid account that answered
- every message was reported as no provider at all. GLM and Qwen were read and
- never mentioned. There is one predicate now, and a new provider card lists
- itself.
-- **Skales stops calling twelve of its own abilities switched off.** The
- self-report looked them up under a name the settings page does not write, so
- the answer was silently no. Places, Planner, Teams, Extensions and eight more
- now report the switch they actually have.
-- **Drive, Docs, Calendar and YouTube can be called when they are configured.**
- They were announced as active and then were not in the toolbox, because the
- announcement and the catalogue read different files. Calendar was the same
- mistake one step worse: a machine with only an Apple or Outlook calendar had
- no calendar tool at all.
-- **Connecting a Google account no longer paints an error over the success.**
- The landing page finished the exchange a second time and reported that the
- authorisation was closed, on top of a connection that had just worked.
- Disconnecting it has an exit too: a request that is dropped rather than
- refused used to leave the card waiting for ever, and a delete that failed
- still said the account was gone.
-- **GigaChat can use tools on both of its hosts.** The current one refused the
- follow-up request because a tool result was sent as raw markdown where it
- wants JSON. The older one has no tools field at all, so the tools travelled
- there and were discarded before the model saw them; it now gets the contract
- it publishes. Model lists belong to the host that offers them, so the endpoint
- you selected stops offering a model it answers 404 for.
-- **A tool call written in any script is recognised as a tool call.** The
- recogniser was a list of known markers, and a list only grows after somebody
- reports what it missed. It asks a structural question now, so a block in
- Arabic or Cyrillic is caught the same as one in English, and a fenced example
- is still left alone.
-- **The Google Places hint names the API that is actually called**, which is
- Places API (New) and not the older one with almost the same name, and the card
- has a test button that hands Google's own answer back word for word.
-- **Pressing Try again after a crash reloads instead of remounting.** The
- retry rebuilt the router inside a document that had already lost it, so
- /profiles and /settings answered a crash with the same crash, three times in
- three seconds.
-- **Notifications stop appearing under the window buttons** on Windows, and the
- full-screen close button in Studio moves out from under the system one.
-- **A tool that cannot work stops being asked again.** A missing browser, a
- search that hit a bot challenge, a path that is a folder: those were retried
- five and nine times a second with slightly different arguments. The failure
- now says whether it is worth another try, and the runner listens.
-- **A tool name mangled by the model resolves to the tool it meant**, instead of
- being reported as an unknown one.
-
-- **A local model could end up in the cloud without anyone saying so.** Four
- places outside the chat path ended their provider chain at a paid service
- when they did not know the configured provider: Autopilot, Lio and two of
- the background runners. An unknown provider is a named error now, not a
- silent detour, and the hard-coded paid default is gone.
-- **Agents, Flow, Studio and Lio ask the same place the chat asks.** All four
- used to decide on their own which provider answers, and Flow's fall-back to
- a cloud-only list happened in silence. They go through the same resolution
- as everything else now, and Studio's Auto names the model that actually
- answers.
-- **A downloaded model is also found.** The downloader wrote into a folder the
- server never looked at. One folder for both.
-- **The machine may sleep.** A generation that runs into system sleep is
- called sleeping now instead of nothing; on waking, Skales probes the server
- and only restarts it when it has really lost it.
-
-- **The chat page outlives a long answer.** After minutes of an unanswered
- turn the page could be replaced by a crash card, React's error 482. The page
- was suspending its own router: every timer on it polled through server
- actions, and each of those parks the router on an unresolved promise. The
- timers poll over plain requests now, and the seven-minute watch that used to
- die in the fifth minute runs through with the honest timeout line still on
- screen.
-- **A Telegram answer belongs to the chat that asked for it.** When two chats
- wrote at the same time, one could be handed the other's answer, pictures and
- even an approval keyboard included. The answer is picked by chat now, and
- the bot refuses to post one addressed to another room.
-- **Iris's mute button leaves the window buttons alone**, and muting her means
- she stops listening, not that she stops talking: the microphone closes, the
- sentence finishes, and the choice survives a restart.
-- **Opening Iris no longer announces anything.** She does not greet you, she
- listens. Same as on the phone: the window opens, the ear is ready, and the
- first sentence in the room is yours.
-- **The floating Save button no longer covers the end of a settings tab.**
-- **The size on a model card is what the download really costs.** A model that
- can see is fetched together with the file it sees through, and the card
- counts both. The description sits behind the card's own chevron, so the
- list stays a list.
-
-### Changed
-
-- **An image that goes to another model says so.** The routing was right and
- invisible: a picture handed to a model that can see it was a console line and
- nothing on screen. The conversation now carries a line naming the model, and
- says explicitly when the image left the machine.
-- **A pairing that fails against an old desktop says which versions are
- involved**, on the phone, instead of fifteen silent seconds. The confirmation
- dialog on the desktop shows the phone's version too, which is the one dialog
- where the owner had no way of seeing what was asking.
-- **The Agents page asks before deleting the way everything else does.** Its
- second-click-within-four-seconds pattern was built to avoid the operating
- system dialogs that this release removed.
-
-## v12.7.3 - Drop-In
-
-A release about being answered. A local provider that would not talk, tools
-that reported more than they did, and a call you can now interrupt in the
-middle of a word.
-
-### Fixed
-
-- **LM Studio answers again.** The provider card said no API key was needed and
- every message came back asking for one. The guard in front of the chat kept
- its own list of the providers that need none, and had never been told about
- LM Studio. That list is now written once and asked from everywhere.
-- **The LM Studio card's Refresh Models button exists.** Its own hint told you
- to press it; the button was never drawn, and the model list behind it was
- refused for want of a key it never needed. A running LM Studio now shows up
- in the model picker, in model search, on the Agents page and in the provider
- grid, and an agent pinned to it stays pinned.
-- **A group chat can be pointed at your own OpenAI-compatible endpoint**, and
- the group-chat skill runs on a local model with no key at all.
-- **A failed Windows update leaves a trail.** Every step of the update path is
- written down with a timestamp, and the last lines of it ride along in
- Settings > Advanced > Diagnostics. The app also stops quitting on a timer
- while the installer is still starting: it waits for the installer to actually
- be there, and says so when it is not.
-- **A failed send is explained the way a failed test is:** the mailbox that
- refuses your password says which credential it wants and where to get it, on
- every mail action rather than only when you press Test.
-- **A mailbox that cannot be opened no longer answers under the name you asked
- for**, and a list that was shortened says that it was shortened. Emptying the
- trash reports what it deleted, not what it found.
-- **A calendar that could not be read says so** instead of shortening your week,
- and a week with no calendar connected no longer reads as a free one.
-- **Speech, images and video stop claiming more than they did.** A voice that
- fell through to a free fallback says which provider refused and why; text that
- was too long to speak says it was cut; a video job that failed says it failed
- instead of going quiet after "started"; and a provider that answers 401, 402 or
- 429 is named as your account rather than as a fault in the app.
-- **There was a second copy of the speech cascade**, and the two had drifted far
- enough that the same settings produced two different voices. There is one now.
-- **A file too large for the phone is refused before the connection breaks.**
- The ceiling was set by an estimate that counted one round of encoding; it is
- now the measured number, written down once for both file paths.
-
-### Changed
-
-- **Talking over Call Mode interrupts her mid-word.** Call Mode used to stop the
- audio only after the sentence had finished. It now listens through playback
- with the same calibrated listener the Iris window uses, and hands the turn
- back to you as soon as you start speaking. Where the microphone cannot be
- opened for it, the old behaviour stays and the screen says why.
-- **Iris uses your name the way a person does:** once at the start of a
- conversation, and after that only when she is actually addressing you.
-- **Qwen 3.7 gets a profile of its own:** the thinking-mode sampling the vendor
- publishes instead of the generic Qwen profile's non-thinking numbers, and a
- long-thinking budget so a silent stretch is not cut off as a stalled stream.
-- **The guide has its first screenshots**, and the LM Studio and KoboldCpp setup
- pages have caught up with the provider grid that replaced the old dropdown.
-- **The acknowledgment that fills the silence follows where your endpoint
- actually points**, not what it is labelled. A custom endpoint on the internet
- is treated as being on the internet; one that cannot be read at all stays
- silent, as before.
-
-## v12.7.2 - Cockpit
-
-A release about seeing what is going on. What is running and where, what a
-conversation is costing, what a tool actually did or did not do, and what the
-model wrote when it meant to call one.
-
-### Added
-
-- **Autopilot opens on what is in flight.** Goals, runs, tasks and scheduled
- jobs stand in one list, each as a single line saying what it is, which
- surface it is on, which model is answering, when it last moved, and the one
- thing you can do with it. Code sessions working in the background appear
- there too; no screen showed them before.
-- **A Code session can carry a goal.** `/goal` in the Code window arms the same
- ledger the chat uses, and a strip above the composer shows what the session is
- working towards, how far it has got, and whether it is running, waiting for
- you or on hold. It survives a restart, so a coding task can be followed over
- days rather than a sitting.
-- **Sessions have addresses.** A model can list your other conversations - title,
- surface, model, open goal - and leave an unread note in one of them. It starts
- nothing over there and spends nothing. An isolated agent can do neither, and
- the refusal sits in the tool itself rather than in whoever called it.
-- **A long conversation offers to carry on somewhere fresh.** When a conversation
- has really grown - several exchanges, and the size spread across them rather
- than sitting in one enormous turn - the chat says how much history it is
- carrying and what a new session with a summary would start at. It only ever
- offers. Nothing is deleted, and nothing is spent without you saying so.
-- **An available update stays visible until you deal with it.** A pill in the
- window chrome, fed by both the updater and an announcement from us, instead of
- a toast you had to be looking at for five seconds. Hiding it is remembered
- until the next version, and it goes away by itself once the update is in.
-
-- **WordPress can do what your site can do.** Skales reached about a third of a
- WordPress site before. Now: find, read, change and delete posts, the media
- library, categories and tags, comment moderation, menus and widgets, colours
- and fonts, the logo, site settings and permalinks.
-- **The featured image finally gets set.** "Write a blog post and add a fitting
- image" used to put the picture in the body only, and the thumbnail your theme,
- your post list and every WhatsApp or LinkedIn preview shows stayed empty.
-- **The phone reaches the cockpit.** The session directory, everything in
- flight and how to stop it, a coding session's goals, the handover, and
- leaving a note in another session. It can also open a coding session's file
- tree and read a file out of it.
-- **Something is said while the model is still thinking.** A big model can sit
- silent for several seconds before it writes anything, and a pulsing dot in
- that window reads as broken. One short line now says what is about to happen,
- in your language, and disappears the moment the real answer starts. It is
- never stored and never sent back to a model on a later turn.
- Who writes that line follows where your conversation already goes: on Skales
- IQ it comes from Skales and says so; on your own key it comes from a small
- cheap model at the same provider; and if you run a local model it stays on
- your machine or does not happen at all. One switch, in Settings under
- Chat & Code, Assist, turns it off everywhere.
-- **A team run shows up in the list of what is going on.** A team of agents
- working through a task was the loudest thing in the product that appeared on
- only one screen. It has a line now, with the stop button the other rows have.
-- **Sub-agents can be given a role.** A dispatched sub-agent inherits every tool
- in the catalogue and pays for the whole list on every turn. Eleven roles have
- existed for a while with no way to reach them; there is a setting for them
- now, and it still starts on the old behaviour.
-- **The cockpit says when a hand-off is worth taking.** The phone can see it
- too: how much history a conversation carries into every turn, and what a fresh
- session on a summary would save.
-
-### Fixed
-
-- **The nightly identity upkeep stops looking in the wrong place.** The built-in
- 3am job that keeps what Skales knows about you up to date was pointed at its
- memory folder by a relative path, and every night it landed in the program
- directory instead: two red lines in the log, and the half of the job that
- reads your recent memories never read anything. It names its folders outright
- now. The same slip was in the file tools underneath it - listing or creating a
- folder by a short name went to the program directory while every other file
- tool went to your workspace - so they agree now too.
-- **Skales answers questions about itself from its own manual.** Asked what a
- setting does, it used to go rummaging through its own program code: thirteen
- tool calls and a very expensive turn for an answer three sentences long, and
- on a normal installation that code is not even there. It reads the guide, the
- changelog and its capability list, and if the answer is in none of them it
- says so rather than going digging.
-- **The invitation to send feedback waits until you have actually used the
- desktop app.** It only asked how old the installation was, so somebody who
- installed Skales on their computer, worked on their phone and never opened the
- desktop window got a form three days later asking what was missing.
-
-- **Mute now means mute.** Switching notifications off silenced everything
- except the messages Skales considered important, which is how two reminders
- arrived at ten to midnight with the switch on. Nothing goes out on any channel
- now, and nothing is swallowed either: what was held back waits on the
- Notifications page, and a line at the top says how many important ones there
- were.
-- **One meeting, one reminder.** Three parts of Skales reminded about calendar
- events without knowing about each other, so a single entry could ping you
- twice. Whichever gets there first now owns the reminder for that event that
- day, and tomorrow's occurrence of a recurring entry reminds again as before.
-- **Friend Mode switched off also stops Skales taking the initiative.** The
- proactive assistant had its own switch buried on the Notifications page, so
- turning the companion off left it running and it kept surfacing meetings and
- overdue tasks.
-- **An approval request obeys the mute too.** It went straight to Telegram and
- never asked, so it was the one ping a muted user still got.
-- **Unchecking every notification channel now actually means silence.** A
- last-resort system notification still went out, because it could not tell
- "the channel you picked is unreachable" from "you picked none". The first one
- still gets the fallback; the second one gets the Notifications page.
-- **A tool call the model wrote as text no longer becomes the answer.** In any of
- the shapes models write it in, including the ones that are not in English and
- the ones that look like a line of Python. A call whose name just missed a real
- tool is now corrected in the conversation instead of being dropped in silence,
- and the same correction reaches Telegram, WhatsApp and the CLI, not only the
- desktop window.
-- **Personas report a failed tool as a failed tool.** Inventing a plausible
- result was the most expensive thing any of them did.
-- **A provider error says which provider, which status, and what the provider
- itself said** - in every language. The detail used to be attached to the
- English sentence and then thrown away by the translation.
-- **A refused page stops being reported as a page.** `fetch_web_page` and
- `extract_web_text` lead with the HTTP status and say in words when what came
- back is a server's error page. A PDF, an image or an archive is refused by
- name instead of being run through the HTML strip and handed over as text.
-- **Web search says whose problem a failure is.** A rejected key is a rejected
- key, a spent quota is a quota, and a provider having a bad day is that. The
- SearXNG hint about JSON output only appears on the status that actually means
- it.
-- **A file tool says which argument it is missing,** and accepts the argument
- names models really send. A well-formed call carrying `contents` or
- `file_text` used to fail where a malformed one carrying the same key was
- repaired.
-- **A write that wrote nothing leaves nothing behind.** No folders, no backup
- file, and an existing file untouched.
-- **A full disk reads as a full disk** in every file tool, not as an errno line
- in five of them.
-- **A signed-in ChatGPT, Claude or Gemini subscription is stored the way an API
- key is stored.** The token that signs you in sat in the settings file as
- readable text next to keys that were encrypted; it is encrypted now, and an
- existing sign-in is converted the next time Skales starts.
-- **A schedule that switches itself off says so.** After three failed runs a
- schedule pauses itself, which was right - but it only wrote a log line, so a
- nightly job could simply stop happening and nobody found out.
-- **The phone stops waiting for an answer that was never coming.** Two of the
- requests it makes about a coding session could end without a reply at all, so
- the phone sat out its timeout and then blamed the computer instead of naming
- the reason.
-- **Listing a folder no longer empties the context window.** The listing is
- capped, says how many entries there are in total, and no longer holds up every
- other conversation while it counts.
-- **A big image no longer breaks the automation.** A generated picture used to
- travel through the model as one enormous block of text before it was uploaded;
- Skales reads the file itself now. That is where a scheduled overnight run used
- to stop.
-- **SEO is read before it is overwritten.** And when the site is set to
- discourage search engines, Skales says so instead of reporting SEO nobody will
- see.
-- **A refusal names what is missing.** If the connector hangs on an account with
- limited rights, Skales says which right is missing on which account. And when
- a design change would do nothing in the active theme, it says that instead of
- reporting success.
-- **A signed-in ChatGPT subscription starts on the general flagship** instead of
- the Codex coding model, and the account's own model catalogue is asked for
- rather than assumed.
-- **Google says why it refused:** a location it does not serve, a rejected key,
- an API that was never switched on, or an exhausted quota, instead of one
- blanket "check your settings" for all four. The same refusal from OpenAI and
- from a ChatGPT subscription reads the same way, in every language.
-- **The chat model picker lists the active provider's models straight away**
- instead of waiting for a search term, and an install with no provider set up
- is told that, rather than pointed at a search box with nothing behind it.
-- **The CLI routes that manage MCP servers ask who is calling.** Four of them
- did not: one handed out the stored server configuration including the linked
- server's keys, one deleted a definition, and two started and stopped the
- configured process, for anyone who could reach the port. The DevKit token is
- now compared the same careful way as the API token beside it.
-- **A command that fails says how it failed.** The exit status travels with the
- answer, so a search that found nothing reads as nothing found rather than as a
- broken command, and a program that is not installed says that instead of
- blaming the command. Output too large to read back is reported as a command
- that ran with its output cut, not as one that never happened.
-- **A shell tool with nothing to run says which argument is missing** instead of
- answering with an internal error, and the same goes for the browser: no
- address, no target to click, no key to press. A scroll direction the browser
- does not have is refused rather than quietly scrolling the other way.
-- **A page that was not there stops being read as a page.** When a site answers
- 404, 403 or 503, the browser says so before the error page's own words are
- handed over, the same way fetch already does.
-- **Background commands left running by a previous start are visible again.**
- After a hard restart the list said nothing was running while last session's dev
- server still held its port.
-- **Twenty translations had lost the value they were supposed to show** - a count
- with no number, a countdown with no seconds. They are repaired, and a build
- check now catches the next one.
-- **The profiles that ship inside Skales say the same thing as the public ones
- again.** Eight model families, GigaChat among them, existed only in the
- library fetched over the network, which means exactly the machines that cannot
- reach it were the ones without them.
-- **The GigaChat card offers the generation it can actually reach** and no longer
- asks for a certificate that ships with Skales.
-- **The release check stopped warning about two Windows manifests** that were
- never finished and have now been removed.
-
-### Changed
-
-- **The user guide is a guide again.** It had a chapter system and no way to use
- it: the chapter headings in the sidebar were not links, and there was no
- search at all. The sidebar now switches chapters from anywhere, the search
- reaches every chapter at once, and each of the eleven chapters opens with a
- walkthrough you can follow on your own installation in about three minutes.
- Mobile and remote, which had no chapter, has one; so do the model readout, the
- profiles page and updates. What was in the guide and no longer true went out
- with it.
-- **The waiting line, the model that writes it and its fallback stand together
- in one place.** Three settings that explain each other sat under Goals, where
- the first of them has nothing to do with goal mode, and the same feature was
- described three times on one screen. They are one section now, Assist, in
- Settings under Chat & Code, with one explanation, and the fallback sits
- underneath the switch it belongs to instead of beside it. Nothing you had set
- changes; the guide has a section on it.
-- Update announcements reach every screen in the app, not just the dashboard.
-
-## v12.7.1 - Which Model
-
-Two reports on the day 12.7.0 shipped, both about the same thing from opposite
-ends: a model you cannot reach, and a model the app names wrongly. This patch
-answers both.
-
-### Fixed
-
-- **GigaChat reaches its current generation.** Skales talked to Sber at the
- older address, which does not serve the GigaChat 3 models at all, so an
- account with a live Ultra entitlement could not reach a single one of them.
- Skales now uses the current address, and the GigaChat card has an endpoint
- field with both hosts on it, so a company still working against the older one
- can stay there and anyone running their own deployment can type its address.
-- **A model id you type is sent, not second-guessed.** Typing a model name that
- Skales had never heard of came back as "the model was most likely retired or
- renamed" even when the model was alive and well and the real problem was the
- address it was asked at. Skales now only says a model is gone when the
- provider says so; anything else names what actually happened and points at the
- setting that causes it.
-- **The context readout names the model that is answering.** Picking a model for
- one conversation changed what answered but not what the readout under the
- composer said: it kept naming your default model and measuring against that
- model's window. It now names the model the next message will use, whether that
- comes from the picker, from an agent, or from your defaults.
-- **The status pill says where this conversation goes.** A chat running on a
- cloud model under a local default read "Local AI", and a chat running on your
- own machine under a cloud default was labelled with the cloud provider. That
- line is a statement about privacy, so it now follows the conversation in front
- of you and goes back to your default when you leave it.
-- **Switching the model with a command changes this chat, not your settings.**
- Typing the switch command wrote the model into your provider card instead, so
- it could pair a model with a provider that does not serve it, it outlived the
- chat it was typed in, and since the conversation started carrying its own
- choice it no longer reached the next message at all while still reporting
- success. It is now the same per-chat switch the picker is: your defaults are
- left alone, the provider is worked out from the model, and if more than one of
- your providers serves that model, Skales asks which one instead of guessing.
-- **Context windows come from the provider's own catalogue.** A model with a
- one-million window was measured as if it held 128,000, so the automatic
- shortening of long conversations began at roughly an eighth of the real
- budget. The window now comes from the same catalogue Skales already downloads,
- and both the readout and the shortening use it.
-- **The Code window names the model that will answer.** A session with its own
- model was labelled with your default one, and a session with no model of its
- own showed nothing at all rather than the default it actually runs on. Its
- context figure was measured against a model nobody had selected.
-- **Lio says what "use default" means.** The option said "Settings" while Lio's
- own saved configuration is what decides, so it could name a provider Settings
- had nothing to do with. It now shows the provider and model that will build.
-- **Compacting a conversation by hand uses the conversation's model.** The
- button sized the history against a model that was not running, which on a
- large window meant it cut far more than it needed to.
-- **A lost connection is no longer reported as a crash.** Looking at Skales from
- your phone or a second computer and the machine running it goes to sleep,
- restarts or changes network: the request never arrives, and the screen said
- "something crashed" with the browser's own words underneath. It now says the
- connection was lost, in your language, and explains where to look. Nothing had
- crashed in that case, which is why those reports were impossible to act on.
-- **The diagnostics report stops calling ordinary shutdowns crashes.** Closing
- Skales stops its engine, and each of those was listed as a recorded crash, so
- a report of four crashes could be one crash and three normal exits. They are
- listed separately now, by name.
-- **A recorded engine stop finally says what happened to it.** The report
- printed the word "undefined" for exactly the entries that carry the most: the
- signal, the exit code, the last output of the engine, and the last things the
- window asked for before it went away. All of it was recorded and none of it
- was shown.
-- **Times in the diagnostics report are your times.** They were written in UTC
- and formatted to look local, so an event at 19:52 was reported as 17:52. The
- report now says which timezone it is using.
-
-### Added
-
-- **A bug report can carry the crash your computer recorded.** Reports that say
- "it closes sometimes" are impossible to act on, while the answer sits on the
- reporter's own machine behind a Copy button on another page. The report dialog
- now offers the last recorded crash as an attachment: switched off unless you
- turn it on, and the exact text is on screen before anything is sent.
-- **The wait before the first word says what it is.** Between sending and the
- first character the chat showed a spinner and the word "Thinking", which is
- the same picture a hung request draws. It now says it is reading your prompt
- and how long it has been doing so, and on a model running on your own machine
- it adds that this phase is silent by nature and can take a while on slower
- hardware. The moment the first word arrives, the line is gone.
-- **GigaChat works without hunting for a certificate.** Sber's endpoints are
- signed under the Russian national root, which no operating system carries, so
- until now the first step was finding that certificate and pasting it in.
- Skales ships it and uses it for GigaChat requests only. It is never installed
- on your computer, never applied to any other provider, and it is added to the
- roots your system already trusts rather than replacing them. The field stays
- for anyone who needs their own certificate, and the card shows the exact file
- the bundled one came from.
-
-## v12.7.0 - Everybody's AI
-
-Skales started as a platform for everyone: set up in minutes, no prior
-knowledge, grandma approved. Over the past two months Skales has grown fast -
-and week by week it became more and more a power user's tool. Both are true,
-and both belong to Skales. But one promise was at stake: the lightness.
-
-So with this release we rethought setup, add-ons and settings. On first start
-you choose what you want Skales to do for you - and that is exactly what you
-get: a sidebar that only shows what you use, settings that in Standard show
-only what belongs to your choice. And for everyone who wants everything:
-Advanced still shows every screw, same as always.
-
-Every single piece of feedback was read. We matched the roadmap against your
-feedback and pulled forward what helps you today. And the promise stands: we
-will keep helping people who do not know what an API is to run a safe, local
-agent on their own computer - one that respects your privacy and ties you to
-no cloud and no provider.
-
-Skales stands for everybody's AI. What it is not yet, it will become.
-We do not build for a target group - we build for every one of you.
-
-### Fixed
-
-- **A local model is allowed to think before it answers.** LM Studio, llama.cpp
- and anything else on your own machine sends nothing at all while it reads your
- prompt, and on slower hardware that quiet stretch runs for minutes. Skales
- measured it with the same thirty-second stopwatch it uses for a cloud stream,
- declared the connection dead, silently sent the whole request a second time
- (so the machine evaluated the same prompt twice), and then reported that the
- provider had stopped responding. Local and custom endpoints now get a
- generous window for the wait before the first token, and the ordinary short
- window from the first token onward, so a socket that really dies mid-answer is
- still caught. The Request Timeout slider in Settings raises that first window
- too; until now it only ever governed the wait for the connection itself, which
- is why turning it up changed nothing. Nothing about cloud streaming changed.
-- **A local runtime is no longer asked to do the work twice.** A stalled stream
- used to be retried by sending the identical request again. Against a cloud
- endpoint that is a cheap second try; against your own machine it throws away
- the prompt evaluation already done and starts it over, so the retry is slower
- than the attempt it replaced. Local endpoints now report what happened instead.
-- **An interrupted answer says it was interrupted.** When a provider's
- connection closes mid-generation without saying why, Skales used to record
- that as a normal finish. At the end of a chain of tool calls the result was a
- cut-off turn presented as a completed one, sometimes with nothing but a
- cheerful closing line. A dropped connection is now named as one: an answer
- that was already partly written is continued, an empty one is reported and
- retried rather than accepted, and the same is true on the Anthropic and Gemini
- paths, in the chat, and in the Telegram and WhatsApp loops.
-- **A model that thinks slowly can now say so.** Skales decides how long a model
- may stay silent by reading its name, which works until a model that thinks for
- minutes is called something that mentions none of it. An LLM profile can now
- declare it outright, so the next model of that kind is a profile update rather
- than an app release. Nemotron is declared in the built-in profiles.
-- **A blocked free model no longer sends you looking for a replacement.** When a
- provider answers that no endpoint matches your data policy, Skales used to say
- the model had most likely been retired or renamed. The model is there; your
- account's privacy settings exclude every endpoint that could serve it, which
- is what usually happens with free variants. The message now says that, and
- names the model.
-- **A 403 from OpenRouter names its three causes.** The key being disabled or
- over its limit, the account privacy settings excluding every endpoint for that
- model, and the model's provider not accepting requests from your location, in
- the order worth checking.
-- **Provider errors are read in your language.** Every provider error already
- carried a translation key and nothing ever looked one up, so all twelve
- languages saw the English sentence. They no longer do.
-- **Excel and Word files can be attached.** Skales could read a .xlsx or a
- .docx off disk for a year; the paperclip just did not know the formats
- existed, so the picker never offered them and a dropped file came back with
- "cannot be read directly" and instructions to save it into the Workspace by
- hand. They attach now, in the chat, on the new-chat screen, in a coding
- session and from the phone, through the same readers the agent already used.
- A format nothing here can open (.doc, .ppt, .rtf) now says which one it is and
- what to export it as, instead of getting the same sentence as every unknown
- binary.
-- **The conversation says when it stops carrying everything.** A long
- tool-heavy chat was cut to its last sixty entries at a fixed count, silently,
- while the context meter still showed room: the cut counted messages and the
- meter counted tokens, so a model started forgetting for reasons nobody could
- see. The limit now comes from the model's own window, older parts are
- summarised rather than dropped wherever that is possible, and either way a
- line appears in the transcript saying what happened. Nothing was ever deleted
- from disk, and the line says that too. Reopening a long chat shows the same
- notice in all three places it can happen; before, only one of them mentioned
- it.
-- **Windows window buttons stop sitting on the Code window's header.**
- Minimize, Maximize and Close are painted over the page on Windows, and the
- Code window's header put the model chip and its buttons in exactly that
- corner. The header now reserves the width the system reports for those
- buttons, so it stays correct at any display scale. macOS and Linux are
- unchanged, and the other four windows were checked: two already reserve the
- strip, and two have no system buttons at all.
-- **A working agent is no longer stopped for being slow.** Sending a second
- message while Skales was answering the first one cancelled the answer, even
- when it was three tool calls into finishing the job. Now the count only
- builds while the turn produces nothing: an agent that is visibly working
- keeps working however impatient you get, and one that has gone quiet is still
- stopped after two messages. "stop" still stops it immediately.
-- **Two people on one Telegram bot are two people.** Several Telegram accounts
- writing to the same bot shared one conversation, and every message was
- labelled with the name of whoever paired the bot first. Each sender now has
- their own conversation and their own name, and /clear clears their own rather
- than everyone's. WhatsApp already worked this way.
-
-- **The Tasks page no longer crashes on a task with no name.** A to-do created
- by a model that left the title out was written to disk without one, and every
- later visit to the Tasks page hit the crash screen instead of the list, over
- and over, until the file was deleted by hand. A task that arrives without a
- title is now named after what it says to do, both when it is written and when
- an existing one is read, so the ones already sitting on your disk stop
- breaking the page.
-- **Your own endpoint is no longer capped at 65K.** A custom or LM Studio
- endpoint had no per-model table to look its context window up in, so it fell
- to a generic 65,536 - and a 200K model was compacted at a third of its window.
- Three things changed. The endpoint gets to report its own window, and Skales
- reads it whatever the server calls it. The extra endpoint slots, which had no
- advanced panel at all, now have the same timeout, retry and limit controls
- every other provider card has. And a number you type into Override Model
- Limits now wins over everything else, which it did not: a window reported by
- the endpoint quietly overruled it. The panel also says which of the three the
- figure came from, so a setting that took can be seen to have taken.
-- **The context meter shows what Skales is actually using.** It resolved the
- window without ever looking at your overrides, so raising a limit in Settings
- changed how much Skales sent and never changed the number on the screen, which
- is how anyone checks. It reads the same override the call sites read now.
-- **The empty composer is one line tall again.** The box grows with what you
- type by measuring itself, and the measurement counted the placeholder text as
- if it were a draft. Any moment the window was narrow - a page transition, a
- resize - "Ask anything, or describe a goal..." wrapped to six lines and a
- six-line box was burned in, and it stayed that way, because the height was
- only re-measured when the draft changed and an empty field has no draft. The
- thinking dial and the access lock also stood half again too far apart, and now
- sit as one pair. Both the start screen and the chat.
-- **The German in the app says "du" everywhere.** The WordPress card, the two
- subscription disclaimers and the whole AIPointer page addressed you as "Sie"
- while the rest of the app said "du" - a tone that switches mid-app reads as
- two products stitched together. Thirty-three strings pulled across, each one
- read first, because German writes the polite address and the ordinary third
- person with the same words.
-- **The German in the app is written in German again.** The strings the new
- setup and the Standard view brought with them arrived with every accent
- stripped, in all seven Latin-script languages: "Was soll Skales koennen?",
- "Rien n'est definitif", "Skales nen lam duoc nhung gi?". That was the
- onboarding, the first screen a new user reads. All of them are restored, and
- the German file is now clean throughout - two hundred more strings had been
- carrying "aendern" and "Gedaechtnis" since earlier releases. A build check
- counts the accent-free sentences per language so the next batch cannot arrive
- the same way.
-- **Two settings nobody could search for.** The safety levels answered to their
- old names - strict, balanced, permissive - so typing "unrestricted" into the
- settings search found nothing, and the folders Skales must never touch
- answered to nothing at all. Both are findable.
-- **Simple chat stops offering DLNA casting.** It has been retired since
- v11.3.2 and was still being advertised as always available.
-- **A crash that happens in the window is now recorded like one.** Until now
- only the background processes wrote to the crash log, so a user asked to
- reproduce a crash and open Settings, Advanced, Diagnostics found nothing
- there. A render crash now lands in the same place, with its stack, and the
- crash screen says where to find it. In the chat it also carries the component
- trail that names what threw, which is the one thing that makes a crash nobody
- can reproduce fixable.
-
-### Added
-
-- **GigaChat is a provider now, not a workaround.** Sber's models were reachable
- only by running a proxy of your own, because the custom endpoint field wants a
- key it can send as-is and GigaChat wants something else twice over: an
- Authorization Key that has to be traded for a token roughly every half hour,
- and a root certificate that is in no default certificate store. Skales does
- the trade itself, keeps the token fresh, and trusts that certificate for
- GigaChat's requests only - not for this computer, and not for anything else it
- talks to. Paste the key, pick whether it is a personal, B2B or corporate one,
- and add the certificate on the card.
-
-- **You can put a chat on a shorter leash.** Next to the thinking dial in the
- composer there is now an access control: read-only, read and write, accept
- edits, or auto. Read-only means Skales can look at anything and change
- nothing, and it is refused in the place tools actually run, not merely greyed
- out on screen. It is set per conversation, it survives a reload, it is there
- on the start screen as well as in a running chat, and read and write, exactly
- what Skales did before, stays the default. When a folder is bound in Code mode
- as well, the shorter of the two leashes is the one that counts.
-- **Skales can show you its own crashes.** Settings, Advanced, Diagnostics:
- what this machine has recorded, on screen and on the clipboard. Crashes, the
- provider calls that failed this session, recent warnings, and the version and
- hardware they happened on. It is read from this computer and sent nowhere. An
- install with nothing recorded says exactly that, rather than implying health.
-- **The assistant reads its own logs before it explains a failure.** Asked why
- something broke, it used to answer from general knowledge about programs like
- itself. It now reads the same diagnostics you can see, and says plainly when
- they show no cause.
-- **LM Studio is a provider with a name.** It always worked through the generic
- custom endpoint, for anyone who knew the port, the path and that the key
- field should stay empty. It has its own card now, with the address filled in,
- no key asked for, a note about the server switch that trips most people up,
- and the local prompt-evaluation window from the first item on this list
- applied without anyone configuring it.
-- **Your own browser can do the browsing.** Settings, Browser Control takes the
- full path to a Chromium-based browser, so automation runs in the profile and
- the extensions you already have. Leave it empty and nothing changes. A path
- that is not there stops the launch and says so rather than quietly opening a
- different browser, and a fork or an old build that refuses to start is named
- as the untested thing it is.
-- **The add-on cards stop taking each other's entries.** Two cards on the setup
- screen could claim the same add-on, so switching one off switched something
- else off with it. Each add-on now belongs to exactly one card, and the last
- screen of the setup says how many are on rather than guessing.
-- **The accessibility check measures the two setup screens too.** They were the
- newest screens in the app and the only ones the audit did not open.
-- **Eight accent presets, one click each.** The three-colour picker stays for
- anyone who wants it. Every preset is checked against the same readability bar
- the picker warns you about, in light and in dark, before it can ship.
-- **The note that says what this release is about is in your language.** It sits
- above the changelog in Settings, Advanced, Updates, in all twelve - because the
- person who most needs to hear that setup and settings were rethought is the one
- who does not read release notes in English. It retires itself when the build
- moves off this line, so it can never end up describing an older release.
-
-### Changed
-
-- **The setup asks what you want Skales to do, and builds the app around the
- answer.** A new step at the end of the first run offers five kinds of work -
- Generative AI, Audio AI, Coding AI, Business agent, Personal agent - and each
- one switches on the add-ons that belong to it. Every individual switch stays
- on the same screen, so a bundle is a starting point and not a package. What
- gets stored is the add-ons, never the category. The sidebar is then built from
- that answer: an entry whose add-on is off is simply not in the list. Chat,
- memory, planner, tasks, schedule, history and Discover are not offered,
- because they are not a choice.
-- **Settings has a Standard view and an Advanced one, and Advanced still shows
- everything.** The switch sits at the top of the page. Standard shows the
- settings your add-ons actually need; Advanced shows every setting there is,
- exactly as before. It is a view and not a state - anything switched on in
- Advanced stays on in Standard, only out of sight - and search reaches both, so
- typing what you are looking for finds it either way. A tab that would open
- onto an empty page in Standard is not offered. The Advanced view is not a list
- of dangerous things: Diagnostics, Updates, Export/Import and the data controls
- live in the Advanced *tab* and are shown in both views, because those are
- things ordinary users get sent to.
-- **An existing Skales is offered the new setup once, and loses nothing to it.**
- A single notice, dismissable, that opens the same add-on screen with
- everything you already had left switched on. Settings, Advanced, Setup runs it
- again whenever you want. Nothing is deleted by either: chats, keys and
- settings stay exactly as they are, and only what is shown changes.
-- **Who belongs to whom is decided in one file.** Which add-ons a category
- offers, which sidebar entries follow an add-on, and which settings sections
- the Standard view shows all used to be three separate lists kept in step by
- hand. They are one map now, and a test walks the real files to prove nothing
- is missing from it - an unplaced add-on or an unkeyed settings section fails
- the build rather than quietly becoming visible-to-everyone or
- invisible-to-everyone.
-- **Skales knows the interface it is actually running.** The map the assistant
- answers "where do I find X" from is no longer written by hand: it is generated
- from the same files that build the sidebar, the Settings page and the Add-Ons
- list, and the build fails if the interface moves without it. So after a
- release that rearranged a great deal, asking Skales where something is gets
- the current answer rather than a confident description of the old app. The
- answer also says which view a setting is in and which add-on it follows.
-- **A new installation starts with the Desktop Buddy switched on.** Buddy is the
- face of Skales - the skin is picked during setup and it is the most used thing
- in the app - and it was off unless you went looking for the switch. It now
- arrives on, introduces itself in your language the first time it appears, and
- points at Settings, General, where it is changed or switched off. An existing
- Skales is not touched: if you never turned Buddy on, it stays off, and if you
- turned it off, it stays off.
-- **Discover and Memory now stand on both sides of the sidebar.** The feed and
- the working context used to live only in the Home view, so switching to Work
- took both away. They are core surfaces and follow you across the toggle.
-- **Wrapped is always on again.** The yearly recap only works if its numbers
- are collected all year, so Wrapped is no longer an add-on you can switch off -
- which also means it returns for anyone who had it disabled. Its entry stays
- in the Home view only; nothing about what it collects has changed, and it
- still leaves your machine only if you share a card yourself.
-
-## 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.
-
-## v12.6.5 - Say What Happened
-
-
-A release about the difference between a screen that says something and a
-thing that happened. Two answers people were promised, a tool family that
-reported success it had not earned, and the first accessibility check this
-app has ever had that actually runs.
-
-It also clears out a set of things that were there but could not be got at: an
-endpoint you had configured and could not pick, a second one you could not
-find, a number under an answer that read as waste, and a grid you scrolled past
-every day.
-
-### Added
-
-- **The system appearance setting follows the system.** Choosing "System" used
- to read your operating system once and remember the answer, so it never
- noticed a later switch and the button never stayed selected. It follows now,
- live and after a restart, and the sun and moon in the sidebar set the same
- setting the appearance row does instead of a second copy of it.
-- **A busy Skales IQ says it is busy, and shows you the way on.** When the
- service cannot take a request, the message says so in your language and
- offers the next step in the same breath: connect your own provider. During
- setup the Skales IQ card stays where it is, says why it could not start, and
- leaves the other three options one click below instead of finishing setup in
- silence with no provider at all.
-- **An accessibility check that runs.** Six core screens are measured against
- the WCAG AA rules and walked with the keyboard, and the check is part of the
- toolbox rather than something someone remembers to do.
-- **Every extra endpoint can be made the active provider.** The additional
- OpenAI-compatible endpoints could be chosen per chat but could never become
- the one Skales routes to by default. Each one has a Set Active button now,
- like every other provider card.
-- **A way to add a second endpoint from where you already are.** It used to
- live only behind a dashed tile up in the provider grid, which is above where
- you are standing when you configure the first one. The Custom endpoint card
- offers it directly and takes you to it.
-- **The provider grid folds away, and remembers.** A header with a chevron
- above the tiles: fold it once and the card you actually came for is at the
- top from then on, after a reload and after a restart. The header still says
- how many providers are switched on, and a search opens the grid so a match is
- never hidden behind the fold.
-- **The token count under an answer says what it is made of.** Hovering it now
- breaks the input half into the two things it consists of: the tool
- definitions and system prompt that are the same on every turn and are what
- makes the agent able to act at all, and your own message plus the
- conversation. Nothing is deducted and nothing about billing changes - the
- number was always this, it just never said so.
-- **The guide says which providers work where.** Some services do not serve
- every country, which is a fact about them and not about Skales. The
- Providers chapter now lists the full replacements, including the local one
- that always works.
-
-### Fixed
-
-- **The sidebar and dashboard provider status is green again the way it always was:**
- an active provider with a key shows Connected immediately. The stricter
- proven-connection state briefly shipped here had no way to ever turn green for
- the provider and read as offline on every install. Live proofs are still
- recorded quietly (a successful chat turn or provider test), and while the
- status is loading the sidebar shows a neutral placeholder instead of flashing
- red with a guessed provider name.
-
-- **Windows finds npm and Node.js again during browser-automation setup**, even
- when Node was installed last: a corrupted PATH entry hit exactly freshly
- installed Node, so everything worked in PowerShell and failed in Skales.
-
-- **A goal that says it is finished, and a check that says otherwise, now ask
- you.** They used to argue in private: the run declared the task complete, the
- check refused to confirm it, and the run was pushed straight back into the
- same step. One run said it was done twelve times over twenty steps and only
- stopped because someone stopped it. The disagreement comes to you instead,
- with both sides on one card and three answers: accept it as done, keep
- working, or stop. A run that keeps repeating the same closing summary with
- nothing new behind it now stops on its own and says so, and it says the same
- paragraph once instead of collecting a copy of it every round.
-- **A goal shows you its checklist before it starts,** and an instruction you
- wrote down in plain words ("use at least three parallel strands") stays on
- that checklist word for word instead of being summarized into something
- general and rediscovered twenty steps later.
-- **WhatsApp is either connected or it is not, everywhere.** Settings said
- "Connected as ..." with your number on it while Skales itself was told
- WhatsApp needed setting up, so it would decline a message it could perfectly
- well have sent. Both now read the same answer. Telegram got the opposite
- correction: a saved bot token with no chat paired to it counted as ready even
- though nothing could be delivered.
-- **A WhatsApp message says what happened to it.** All three send tools
- reported plain success whether the message went out, was merely handed over,
- or went to a number that has no WhatsApp at all. They now say delivered,
- handed over but not confirmed, or not delivered and why, and a number with no
- WhatsApp account is refused instead of quietly accepted. The three also share
- one recipient rule: a contact that works in one of them works in all of them,
- and "unknown number" and "sending to this contact is switched off" are no
- longer the same sentence.
-- **One tooltip per button.** Every icon in the chat header showed two bubbles
- stacked on each other, and the expanded sidebar put a tooltip over the row
- below the one you were pointing at. There is one tooltip, and only where the
- label is hidden.
-- **The keyboard shows where it is.** The chat composer looked identical
- focused and unfocused, so there was no way to tell it had the keyboard.
-- **Muted text is readable.** Hints, captions and the "Configured" badge were
- below the contrast a normal text needs in every light theme, and in Neon.
- Same colour family, dark enough to read.
-- **The selected temperature unit can be read** in a light theme, the Flow
- surface can be scrolled from the keyboard, and the information button on the
- Agents page has a name and opens for the keyboard too.
-- **A file tool that fails says why.** Listing, moving, copying, creating a
- folder and unpacking an archive reported that something went wrong without
- saying what, and all of them told Windows and Linux users to open a macOS
- settings page. Moving into a folder that does not exist yet works, the way
- copying already did, and an archive says whether a file was left out because
- of the size limit or because it was trying to write outside its folder.
-- **Typing into a web page reports where the text went,** and says so plainly
- when nothing had the keyboard and the text went nowhere. Scrolling says where
- on the page it ended up instead of always saying "scrolled down". A click
- placed by the vision model says where it clicked and whether anything
- changed. Attaching a file names the field it attached to and refuses to guess
- when the page has more than one.
-- **Reading aloud stops calling silence a success.** A voice that could not be
- played counted as read, so the next voice in line was never tried.
-- **A team run writes its result into the conversation it happened in,** not
- into whichever one you had open when it finished.
-- **The question card in Skales Code keeps every answer.** It sent on the first
- click, so a card with three questions answered one and threw two away.
-- **The Google authorization page speaks your language,** and the camera move
- you picked keeps its name after you browse another category.
-- **Product names stay product names.** Seventy-two labels across the eleven
- translations had turned Discover, Swarm, Autopilot and Wrapped into ordinary
- nouns, so the thing you read about could not be found in the app.
-- **The Russian first screen is Russian.** The privacy and autonomy notice
- everyone accepts before Skales starts was Russian written in Latin letters,
- as were several settings tabs.
-- **The roadmap page shows the last three releases,** the guide explains why a
- long unattended run can end when a disk parks itself, and the camera-move
- count in both is the number that is actually there.
-- **A model you typed in yourself is findable in the chat model picker.** The
- Custom endpoint has no catalogue to offer - the model is whatever you point
- it at - and that turned out to mean its model appeared in no list the picker
- reads. So an endpoint that was set up, switched on and working could not be
- chosen for a chat unless it happened to be the active provider already. Any
- model you configure on any provider card is searchable now.
-- **The extra endpoints speak your language.** That whole block was English
- only, in all twelve languages.
-- **A local Qwen 3.6 acts instead of printing what it meant to do.** The model
- writes its tool calls in a shape Skales did not read, so a request to look at
- a folder came back as a line of raw text where the answer should have been.
- Skales reads that shape now, and Qwen 3.6 has settings of its own instead of
- borrowing an older model's.
-- **The Open Flow button in Studio can be read.** White on the accent colour was
- below the contrast text needs, on that button and on six more like it. They
- take the readable colour for whichever accent you are running.
-
-## v12.6.2 - Nothing Green Without Proof
-
-Two people wrote in the same night to say Skales had told them a file was saved
-and the file was nowhere. A third could not connect her iCloud calendar because
-that connection had never been able to work. This release is about the gap
-between what the screen says and what actually happened.
-
-### Added
-
-- **A connection is green only after it was tested.** Every "Connected" and
- "Active" badge in Settings now means a live test reached the service and
- passed, and hovering it says when. Credentials you saved but never tested read
- "Configured" instead: the card still looks set up, it just stops promising a
- connection nobody has made. Saving or changing a credential drops the old
- result, so a passed test can never vouch for a new key.
-- **Test buttons where there were none.** Notion, Todoist, Spotify, Home
- Assistant, Replicate, Stability, ElevenLabs and Runway can be tested from
- their own card, and Slack, Signal, X and your Google account got a button for
- the check that already existed behind them and was never run. Kling, Fal and
- Atlas Cloud have no such check, so they stay honest at "Configured".
-- **Apple Calendar works.** iCloud refuses every request against its front door,
- so Skales now walks the discovery path Apple expects, finds your calendars and
- writes to a real one. You can pick which calendar new events go into, and a
- refused connection says whether it was the password or the address.
-
-### Fixed
-
-- **A saved file says where it is.** Generated images, videos, voice files and
- merged clips now report the full path they were written to instead of a bare
- filename or a folder nobody could locate, the answer card shows that path with
- a button that opens it in your file manager, and Skales is no longer allowed
- to tell you a file was written unless the tool that wrote it says so.
-- **Answers no longer carry internal markers.** A model that quoted a tool
- result word for word used to quote the fences around it too, and they landed
- in the answer. The markers come off, the quoted text stays.
-- **The token line under an answer reads correctly.** It showed one number that
- looked like the cost of the reply while it was mostly the question, the
- instructions and the tool list. It now shows the two directions separately.
-- **The VirusTotal key check asked an address that does not exist,** so a valid
- key was reported as invalid. It also tested the dots the field shows instead
- of the key you saved, so the answer was "invalid" every time after a save.
-- **LinkedIn posts and uploads use the current API,** and fall back to the older
- one only when LinkedIn says this app may not use the new one.
-- **Microsoft mail says what it can and cannot do.** Work and school accounts
- cannot be connected with a password at all; personal accounts usually can with
- an app password. A refused send now explains itself the same way a refused
- inbox already did.
-- **A file written while the disk is waking up is written, not lost.** Scheduled
- jobs running unattended could fail on a write that a moment later would have
- worked.
-- **Skales Code on the phone shows your coding sessions again,** even when you
- have thirty newer chats, and archived sessions stop appearing as live work.
-- **Switching a team off during its final step no longer writes a verdict over
- the run you cancelled.**
-- **Signing out of the Google account clears the form it was set up in.**
-
-### Changed
-
-- The model picker mentions the four thinking levels once, so the strongest one
- is not something you have to find by accident.
-- Release builds fetch the Linux package before writing the update manifest, and
- refuse to write a manifest that silently leaves it out.
-
-## v12.6.1 - Order and Follow-Through
-
-The release after the big one is where your first week of feedback lives. The
-provider page you can finally read, agents that take turns on one graphics
-card, a scheduler that owes you nothing after a night of sleep, and a dozen
-small wrongs made right.
-
-### Added
-
-- **The provider page is a grid you can read.** Every provider is a tile with a
- short description, badges for what it can do (voice, vision, media) and a
- switch. Switched off means gone, from the page and from every model picker.
- Filters find providers by capability, and "+ Custom endpoint" is a tile like
- the others. Choosing your active provider works exactly as before, and no
- setting was removed: the advanced ones now live in the provider's own card or
- in one folded Advanced Routing section at the end.
-- **Tencent Hunyuan is a provider**, with a switch between its regions.
-- **Agent teams can take turns.** On local models, running several agents at
- once means they fight for the same graphics card. Sequential mode runs them
- one after another and unloads each model where the runtime supports it, and a
- new "Keep models loaded" setting for Ollama puts that duration under your
- control, from unload-immediately to always.
-- **MCP servers that use a plain API token** get a headers field right in the
- form, and a rejected request no longer pretends the server wants a sign-in.
-- **skales:// links.** A link can open Skales and land on the right screen.
- Nothing is downloaded or executed without a confirmation card first.
-- **Bring your own vector database.** Qdrant and ChromaDB work as retrieval
- backends beside the built-in store, and the built-in one remains the
- fallback.
-- **One Telegram bot, several chats.** Pair additional chats and choose where
- an answer goes.
-- **The scheduler survives sleep.** Runs missed while the machine slept are
- caught up on wake, and a timer that the system froze says so instead of
- silently drifting.
-- **A headless start** for machines without a screen, reachable through the
- existing web access, and **a tiny-model preset** that trims tools and context
- so sub-1B models stay usable.
-- **The guide explains the three multi-agent modes** and when to reach for
- which.
-
-### Fixed
-
-- Buttons and links that were invisible on dark themes. A whole family of dead
- color references, all of them found and replaced, with a check that keeps the
- class out.
-- Signing in with your Google account now unlocks Drive and Docs everywhere,
- not only Calendar, and each card says which login it is actually using.
-- The voice list loads for every provider, not only one, and the preview button
- plays a sample on each.
-- Conversations keep a timestamp per message instead of a clock in the prompt,
- which keeps long conversations faster and cheaper.
-- The navigation menu no longer hides behind page content.
-- Local models that write their tool calls as plain text in one more dialect
- are understood and executed instead of leaking markup into the answer.
-- An agent that stops after a failed tool now says which tool failed instead of
- returning nothing.
-- In team runs, your agents' own instructions outrank the coordinator's
- boilerplate.
-- A crash remembers where it happened: the next bug report
- you send automatically carries the screen and the component that crashed,
- so you do not have to.
-
-## v12.6.0 - Iris Orbit
-
-The release about things that were happening where you could not see them, and
-about who decides once you can. A decision waiting below the fold, a message
-that never arrived, a browser session frozen behind a dialog nobody mentioned.
-Then: a team of agents that ranks its own answers and hands the choice back to
-you, a plan you read before it is written, and a camera move you pick before the
-shot is filmed. Skales recommends. You decide.
-
-### Added
-
-- **Iris Orbit is switched on when you install Skales.** The release is named
- after her, and a surface you have to find in Settings before it exists is a
- surface nobody finds. Nothing else changed: if you turned her off, she stays
- off. Without a provider key the window says so and offers the way to Settings,
- Voice, the way every other surface that needs a key does.
-- **"What can I say?" is a list you can read.** Every sentence Iris understands
- was already there and written down nowhere: turning her into a shape, asking
- for something as a document, a web search, a timer, your tasks. The window's
- menu shows all of it with a real example each, in twelve languages, and asking
- her out loud what she can do gets the same answer without a tool call.
-
-- **Iris Orbit: a way to talk to Skales instead of typing to it.** A window of
- its own, black, with no toolbar and no message list: a particle eye in the
- middle and one quiet line of text under it. Hold the space bar and speak. Iris
- hears when you stop talking and answers on its own, out loud, and the line
- under the eye shows what is being said, sentence by sentence, like a subtitle
- rather than a chat log. The eye reacts while it happens: it warms and pulses
- with your voice while it listens, looks away while it works, and becomes a
- circular soundwave while it speaks. It follows your cursor when it is idle,
- and clicking sends a ripple through it.
- Iris is a face on Skales, not a second assistant. The same conversation, the
- same model, the same tools, the same memory: ask it to forget something and it
- is forgotten everywhere. Iris opens where you left off and says so, "new
- conversation" starts a fresh one, and any chat you are already in can be
- carried over by voice from the button in its header, with everything that was
- said so far coming with it. Go back to the chat afterwards and the spoken
- turns are simply the newest messages in it.
- If speaking is not on right now, start typing: a single line appears over the
- eye, Enter sends it, Escape puts it away again. "New conversation", "my
- tasks", "timer ten minutes" and "forget X" are answered directly, with no
- model involved at all. Cmd or Ctrl plus K brings up the earlier conversations;
- Escape stops whatever is running, and closes the window when nothing is.
- The first time you open it, Iris introduces itself. That plays once.
- The window takes your own light or dark theme, like the Code and Flow windows
- do: a glow on black at night, dark points on a light ground by day.
-
-- **The eye answers with more than words.** Every command Iris can answer
- directly now has its own visual reply. Set a timer and the particles become
- the remaining time, spelled out as digits that count down. Ask how far along a
- goal is and they draw the progress as a bow. Ask her to forget something and
- they scatter and gather again. Start a new conversation and the eye blinks. A
- reminder going off sends a ripple out from the middle.
- You can also just ask for a shape: "morph into a car", "become a feather", in
- any of the twelve languages. There are about fifteen hundred of them, and they
- turn up on their own as well: write down "call the dentist" and the eye
- becomes a telephone, and while Iris is talking the particles quietly take the
- shape of whatever is being discussed. That last one can be switched off under
- Settings, Voice, if you would rather the eye stayed an eye.
-
-- **Results appear inside the eye instead of opening windows.** A picture, a
- document, search results, a short list: the ring leaves the eye, becomes the
- border of a panel, and the result appears inside it. One at a time, and Escape
- puts it away and brings the eye back. Heavier things are handed to the window
- that already does them - a page goes to the browser, a render goes to Studio,
- code goes to the Code window - and Iris tells you where she put it.
- Cmd or Ctrl plus slash opens Help, which lists every command, what it does,
- which answer the particles give, and a sentence you could actually say.
-
-- **Iris does the work, not just the talking.** "Twenty-five minutes of focus"
- starts a timer you can pause and resume by voice or by clicking; when it runs
- out, Iris says so. "What is on today" reads your tasks aloud and shows the
- list, and you can tick them off from there. "Show me the status" reads out how
- far along a goal is. All of it goes into the lists, the schedule and the goals
- you already have - there is no second copy anywhere. Anything that finishes
- while the window is open is spoken as it lands; anything that finishes while it
- is shut is waiting to be read out the next time you open it.
-
-- **Settings, Voice: one place for all of it.** The voice that speaks, the ear
- that listens and Call Mode used to be three cards under Integrations, next to
- Notion and Spotify. They have moved to a tab of their own, together with the
- things that never had a home: how long a silence ends your turn, which key is
- push-to-talk, and what happens when you talk over an answer. There are named
- presets for a local Kokoro and a local Whisper that fill in the endpoints for
- you, so running your own voice on your own machine is something you can find.
-
-- **A wake word you teach yourself.** Say "Iris" three times in Settings, Voice,
- and she learns your pronunciation - yours, not a model's idea of it, so it
- works whatever your accent. Then you can just say her name instead of holding
- the space bar. Everything happens on your machine: the recordings become
- numbers there and are dropped, and nothing is uploaded. When the ear is open
- the window says so; there is never a microphone listening in silence.
-
-- **Talking over her stops her.** With any provider at all, the microphone stays
- open while Iris speaks, and starting to talk cuts her off - what you said
- becomes the next thing she answers, the way it works with a person. Where a
- provider offers a live audio channel of its own, that can be picked instead;
- where it does not, the option says so plainly rather than quietly doing
- something else.
-
-- **A word while a slow model is still thinking.** Some models take several
- seconds before the first word arrives, and a screen with nothing on it but a
- pulsing dot reads as broken rather than as busy. If nothing has arrived after
- a second and a half, the small assist model you set under Settings, Goals,
- Light-pass model writes one short line about what you just asked, in your
- language, into the same place the thinking indicator sits. It disappears the
- moment the real answer starts, it is never part of the conversation, and it is
- never sent to a model again. With no assist model set, and on a model that
- answers quickly, nothing extra happens at all.
-
-- **A team of agents has someone in charge of it, and you have the last word.**
- The first seat of your roster is the coordinator now. Before anyone answers it
- reads the assignment and breaks it down: what a complete answer has to cover,
- and which angle each agent leads on. Every agent then answers the whole thing
- with that brief attached, on its own model. Afterwards the coordinator compares
- what came back and writes a verdict: a ranking, two sentences per agent saying
- why it sits where it sits. No scores, no marks out of ten. The verdict
- recommends and decides nothing, and it says so.
- Every agent card carries two buttons. "Continue with this agent" switches the
- session to that one agent and hands its answer to the next turn as the thing to
- build on. "Take this result" puts the answer into the conversation, or saves it
- as a file where there is no conversation to put it in. Both survive a reload.
- The same cards and the same two buttons now appear everywhere several agents
- produce results: chat, Skales Code, Tasks, Group Chat and the Organization
- result list draw one component instead of five versions of a list.
- Every way a run can end is its own state rather than an error message in
- English. One agent failing leaves a ranking of the rest; all of them failing
- leaves an honest line instead of an empty verdict card; counting the Agents
- button back to one ends the run instead of leaving it working behind a door you
- just closed; and a run whose process died is recognised when you reopen Skales
- and closed off, keeping whatever had already arrived.
-
-- **Flow can film a shot, and you pick the camera move first.** A new chip in
- Flow: Film. It makes one clip with one deliberate camera move, chosen from a
- grid of fifty-one moves grouped by what they are: zoom, dolly, crane, pan,
- orbit, rig, aerial, lens. One move per clip and no way to pick two, because
- mixing them is what makes generated footage look generated.
- The moves are Skales' own film language, not a pointer at somebody else's
- preset library: each one carries its own formula in three parts, where the
- camera sits, what it does and how the move runs over time, and that formula
- shapes the prompt for whichever backend makes the clip. So the order starts
- with your own keys: the video APIs under Settings, Skales Studio first, then a
- Hugging Face Space you have activated, then a connected media server such as
- Higgsfield if you happen to have one, as the optional connector it is.
- With none of the three set up the chip does not disappear and does not pretend:
- it says which key is missing and where it goes, and it keeps the move you
- picked. Follow-up takes build on the last one instead of starting over.
-
-- **Seedance and Atlas Cloud in Settings, Skales Studio.** Seedance was in the
- list and could never have worked: both of its model rows named an address that
- does not exist, so every attempt came back as not found. It runs on fal now,
- which is ByteDance's own route and the only one carrying Seedance 2.0, and the
- rows for people whose key is at Replicate are repaired rather than dropped.
- Atlas Cloud is new: one key across several hundred models for chat, image and
- video, billed per use, and it shows up as a provider in the same three places
- every other one does.
-
-- **Skales asks you one question, once.** If you were already using Skales
- before this version, a box appears once and asks what is missing, what could be
- better, what you actually use it for, what you love and what you do not. One
- text field, any language, as long or as short as you like. It goes straight to
- Mario and it is not published anywhere. A checkbox under the field links to
- the privacy policy, and nothing is sent before it is ticked. "No thanks"
- closes it for good, and a fresh installation is never asked.
-
-- **Hugging Face Spaces count as a media backend in Flow.** The integration has
- existed for a long time and Flow was built straight past it: a Space you had
- activated sat in the model's tool list and was never offered here. Image, video
- and film all know about them now, as the free-or-nearly-free route with a
- shared queue.
-
-- **More than one agent can answer, and you can see each of them.** The composer
- gets a button next to Auto with a number in front of it: how many agents your
- next message goes to. One is off and means what it always meant - the default
- agent, or whichever one you picked for this chat. Every click counts up to six
- and back to one. A run of more than one shows up as cards, the same cards tool
- calls and diffs already use: one for the run, one per agent that opens to that
- agent's answer, and a summary at the end. Who takes part is arranged on the
- Agents page, in a section above the cards; the order in that list is the order
- they run in, so "3 Agents" means the first three. The first seat belongs to the
- default agent and does not move, because somebody has to write the summary.
- Tasks can be run the same way from the create dialog.
-
-- **One Google account instead of the same account asked for five times.**
- Calendar, Drive, Docs and YouTube each wanted their own credential for what is
- one Google account. Sign in once under Settings, Integrations and it covers all
- four. Every one of those fields still wins whenever it holds something, so
- nothing you already set up changes; with no account connected nothing changes
- at all.
-
-- **An archived coding session can be looked at again.** Skales Code could put a
- session away and nothing anywhere could bring it back. The session list has an
- archive icon at the right of its heading now: one click shows what was
- archived, with Restore in the right-click menu, another click returns. The
- search box works in both.
-
-- **Skales can give a web page a file.** Attaching a picture to a post used to
- end in the operating system's file dialog, which Skales cannot use and which
- freezes the browser until you close it by hand: an agent would try eleven
- other ways and post without the picture. The file now goes to the page
- directly, no dialog involved, and a click that would have opened one is
- refused with the way that works instead.
-
-- **A browser session that is stuck says so and stops.** It used to keep trying
- until the budget ran out, because from the inside a blocked page and a slow
- page look the same. Now it recognises being held, ends, and tells you why.
-
-- **You can see which sites Skales is signed in to, and sign out.** Skales
- browses in one profile, so a login stayed until you deleted a hidden folder -
- and there was no list of what was in there. Settings shows the sites, signs
- out of one, or clears the profile completely. It refuses while a browsing
- session is running, because the browser writes the profile back when it
- closes and the deletion would quietly be undone.
-
-- **Replace every occurrence in a file in one go.** Changing the same thing in
- seventy places meant seventy edits, and it was easy to lose the thread halfway
- through. One call now does the file, by exact text or by pattern, and says how
- many places changed.
-
-- **A permission you can answer once for the whole session.** Flow could ask the
- same question ten times in a row with only two buttons on the card. There is
- now a third: allow this kind of action for this session. It covers the action
- AND the reason it was stopped, so agreeing to a command inside your project
- folder is not agreeing to one that reaches outside it. It lasts for the
- session, is never saved, and lifts no safety guard. Under Auto, confirming
- once is enough - which is what choosing Auto was supposed to mean.
-
-- **Your phone can watch a team work.** Starting a team task from the phone used
- to show nothing at all until it finished. The Organization view now sees who
- is working, who is waiting and what tool is running, the same live picture the
- desktop has.
-
-- **Your own transcription server has its own settings.** It shared a card with
- text-to-speech and only appeared when you set the SPEAKING provider to Custom,
- so running your own transcription meant changing an unrelated setting first.
- It sits in the Speech-to-Text section now. Both directions also take extra
- parameters for your server - temperature, language, guidance, steps, whatever
- yours has - as one field rather than a fixed list that would go stale.
-
-### Changed
-
-- **The local voice engines are marked as what they are: optional, and not
- Docker.** The card offering Kokoro and Whisper showed a `docker run` line and
- nothing else, which read as though Skales wanted Docker. It says plainly that
- Skales itself never needs it, carries a Docker-free command for each engine
- beside the Docker one, and is labelled Optional. Ask Skales to install either
- one and it now walks the whole route for your system, health check and
- uninstall included, through the normal command approval.
-
-- **Iris Orbit says it is a first version.** A quiet "Beta v1.0" in the corner of
- the window and a line in its settings block, plus the licence the engine is
- under and where to read it.
-
-- **The sidebar's Iris entry is an eye.** The dot ring stays where it means
- something: the window, the settings block and the logo.
-
-- **The user guide covers the shared Google account and Iris Orbit.** Neither
- was in it. The Google chapter walks the Cloud Console setup end to end, says
- which client type to pick, and explains why a service's own key still wins and
- why adding a service later asks for consent again. The Iris chapter covers
- opening it, speaking, typing, morphs, timers, the wake word, the right-click
- menu, languages, and where the conversations live.
-
-
-- **A plan is something you read before it exists.** "Generate Master Plan" used
- to create every task in it in the same breath, so the first time you saw a plan
- it had already been written into your board. It is a draft now: a list you
- read, untick and edit, and only "Create these" writes anything. A task the
- model got wrong no longer takes the whole plan down with it either. The usable
- ones survive, and the ones that were left out are named with the reason.
-
-- **A Quick Task is a task, not a title.** One line still starts it, but what
- gets created carries instructions the agent can act on, the project it belongs
- to and the agent best suited to it, each one a suggestion you untick if it is
- wrong. And the warning that the task will not run while Autopilot is paused is
- the switch itself now, where the warning is, instead of a sentence about a
- button elsewhere.
-
-- **Cards on the Autopilot board move by dragging.** Two moves are refused rather
- than faked: you cannot drop a card into "in progress", because that is where
- Skales puts a task when it actually picks it up, and you cannot drag one that
- is running out from under the runner. A move also clears what the new column
- would otherwise make untrue.
-
-- **The running-goals strip stops being a shelf.** Open the goal where it lives,
- carry it on, or stop it, without leaving the page.
-
-- **The composer bar never breaks into a second line.** In a narrow window it
- scrolls sideways instead, the way Code and Flow already do, with the context
- readout on the left pinned where it can always be read.
-
-- **Skales is counted while it is running, not while Discover is open.** The
- online figure came from the Discover tab and only from people with a gamertag,
- which meant it was measuring almost nobody. It now reflects running apps.
-
-- **What belongs to the conversation stands in the conversation.** The live view
- of the browser, the plan the model proposes and the offer to turn a chat into a
- goal used to hang above the text box in their own sizes, outside the
- conversation. They sit under the last message now, the width of a message, and
- scroll away with everything else that happened.
-
-- **The user guide is user help.** Five chapters of release notes had grown into
- it, the newest at the very top, so the first thing a new reader met was a list
- of things they did not know they were missing. They are gone; what changed in
- a release is in the changelog, which is where you are.
-
-- **Skales IQ: internal routing update.** Nothing to set, nothing to notice.
-
-### Fixed
-
-- **What Iris writes appears in the ring, not as a clipped line under the eye.**
- Ask her for a poem, a list or a summary and the text lands inside the frame,
- readable, with one short spoken sentence beside it. Until now the ring only
- ever opened for a tool result, so an answer that WAS the thing you asked to see
- was cut off mid-word in the subtitle line.
-- **The Iris window no longer stands still for several seconds when it opens.**
- Opening it made four requests to the local server one after another, with
- nothing on screen moving while they queued. Three of them never depended on
- each other, and the app now warms the window's own route at launch as well.
-- **Approving a file no longer throws you into the Code window.** Saying yes to
- "may I write this" is not asking to change windows. Iris says where the file
- went and stays where she is.
-- **Iris asks for permission in her own voice.** The question used to arrive as
- a chat approval card over the eye. She now says what she wants to do, the two
- answers are hers, and "yes" or "no" out loud answers it in twelve languages.
-- **Iris does not play the thing she turns into.** Asked to become a car she
- wrote a short story about being one, signed off as it, and quietly dropped the
- document that had been asked for in the same sentence. A morph is what the
- particles do; she stays herself, and the task in the sentence still gets done.
-- **Iris says who she is in one sentence.** "I am Iris, the voice of Skales",
- written out per language rather than improvised, after a weaker model answered
- "I am Iris, Skales at the wheel".
-- **She opens with something different each time.** Fifteen greetings per
- language, rotating, instead of the same sentence at every open.
-- **Iris speaks like a person about her microphone.** "Ear open", "ear shut,
- voice off" was a literal translation of an English image into all twelve
- languages. It says listening, mute and unmute now, and the German, French,
- Croatian, Turkish, Portuguese and Vietnamese lines got their accents back.
-- **Iris answers no longer ring a bell in the main window.** Every spoken answer
- produced a "message finished" toast and a chime in the window you were not
- looking at, and the Code window did the same. A companion window that is open
- is you being on that surface.
-- **"Continue with this agent" folds the cards and gives you the composer.** It
- answered the question those cards were asking and left all six answers
- standing between you and the reply you now wanted to type. A conversation
- opened again loads its finished run folded, and the cards have a ceiling and
- their own scroll instead of growing over the input bar.
-- **The working basis is a chip, not a wall of text.** Carrying on from an
- agent's answer put that whole answer into the conversation as raw text, and in
- the Code window it was glued onto the front of the next thing you typed, so
- your own "ok stop" came back with an essay in front of it. It is one line you
- can open, and it travels beside your message rather than inside it.
-
-- **Iris keeps listening.** With the ear open, the microphone still closed after
- every answer and Space had to be pressed again to say the next thing. The ear
- now reopens by itself at the end of everything she says, until "Ear shut,
- voice off" or the window closes. That entry brings her back too.
-- **A document Iris writes appears in the ring.** After the fix that stopped her
- framing the bookkeeping around a document, she framed nothing at all: the tool
- message is a receipt, and the document itself goes into this conversation's
- document panel. She now reads the receipt as what it is, a title, and puts the
- real text in the frame under its own name. The guide has a new section on what
- opens a ring at all: a picture, a document, a web search, a list.
-- **A failed transcription says which provider refused, and why.** "No
- speech-to-text provider available" with a working key configured: every leg
- had been tried, each had thrown, and the reasons were only in the server
- console. The message now carries them ("OpenRouter: 402 ...; Groq: no key"),
- and a genuinely unconfigured setup still says that instead.
-- **Iris no longer introduces herself as Skales.** On a long conversation a
- weaker model would read the voice profile, then four thousand tokens of other
- things, and open with an apology for the wrong name. The one identity line is
- repeated once at the very end of the prompt. Same profile, no second persona.
-- **The relay stops fighting itself in the log.** `socket REPLACED` once a
- second, filling the console and burying every other line. The reconnect used
- to reset its backoff every time the relay let it in, so two connections
- claiming one identity evicted each other forever at exactly one second apart.
- Being replaced now backs off and the line is printed once a minute with a
- count, and it names the process it came from, which is the fact that answers
- it. The engine that opens those connections is also single now: nothing
- stopped a second one being started over a running one.
-- **A stream that ends twice no longer throws.** `Controller is already closed`
- in the console, from closing a stream whose reader had already gone, in fifteen
- places across the update, build, plan and pull routes.
-- **Six debug lines per model response are gone from the console.** They printed
- the raw provider answer, including its content, on every step of every turn.
- One line is left, and only when a step really came back empty, saying what
- that means. A model that spends its whole reply budget on reasoning already
- ends in a proper answer that says so.
-- **Iris' voice is a list you can read.** It was a text field asking for a
- provider voice id in the provider's own spelling, which nobody knows by heart.
- It is now a dropdown of the voices your TTS provider actually has, fetched
- from the provider rather than baked in, so a voice you cloned five minutes ago
- is in it. "Provider default" leads. Where a provider has no list to fetch, it
- says so and the field stays, so an id you do know still works.
-
-- **A run with several agents survives a reload.** Asking a team of agents a
- question in a fresh conversation started the run against a conversation that
- did not exist yet: the question, every answer and the verdict were written
- nowhere, and reloading left only whatever was typed afterwards. The question
- now goes into the conversation before the run starts, the verdict is written
- into it by the server the moment it is ready, and reopening a conversation
- finds its run again and repairs a verdict that never made it in. Nothing in
- that path fails quietly any more: when something cannot be saved, the
- conversation says so at the time instead of losing it silently.
-
-- **A finished team run can be folded away.** After the verdict has been read
- and an agent picked, the cards collapse to one line saying how many agents
- answered and whether there is a verdict, and open again on a click. The state
- belongs to the run, so a phone and the desktop agree about it. The whole run,
- every answer and the verdict, is also kept as one Markdown file next to your
- other documents, so folding the cards away never loses it.
-
-- **The team cards no longer follow you into the next conversation.** Switching
- conversations, or starting a new one, left the previous run's cards sitting
- above the composer.
-
-- **"Take this result" in the Code window puts the result in the session.** It
- used to trigger a browser download, which on a phone reaching the window
- remotely meant the answer ended in a downloads folder instead of the work.
- Pressing it twice also used to add the same answer twice; it does not now.
-
-- **A team run started from the Code window's start screen is visible.** Sending
- with two or more agents before a folder was open started a real run that was
- drawn nowhere at all. The folder is asked for first, as it is for an ordinary
- message, and the run happens in the session that opens. The run is found again
- after a reload, and the agent picked to carry on is still picked.
-
-- **Speech that could not be transcribed by your own server says so.** Pointing
- Skales at your own speech endpoint and having it go down moved your voice to
- whichever cloud provider had a key, without a word. It still keeps working,
- and it now tells you it happened. Transcription through Hugging Face also
- returned something that was not text at all, so Iris said "[object Object]"
- out loud and dictation in the chat failed with an empty error message.
-
-- **The feedback form looks like a form again.** Its backdrop had a three pixel
- blur and no Safari equivalent, so through a browser it was a flat dark sheet
- over an app you could no longer see. It uses the same glass as every other
- Skales overlay now, in every browser. The same missing Safari line was fixed
- in twenty-odd other overlays across the app.
-
-- **Settings stops telling you to update an app that is already current.**
- "Desktop Control" probed a bridge that only exists in the desktop window, so
- viewing Skales in a browser always read "Update to the latest build to enable
- replay". It now says which of the two situations you are actually in.
-
-- **A voice setting that could not be saved says so.** The switch moved on
- screen and quietly went back on the next visit.
-
-- **Iris works when you reach her through a browser.** Three things in her
- window asked the desktop app for something that only exists in the desktop
- app, and did nothing at all otherwise: "Voice settings" in her right-click
- menu, Escape on the resting eye, and handing a result to Studio, the browser
- or Skales Code - which she announced out loud while nothing happened. Each of
- them now opens the place it names, in whichever shell she is running in.
-
-- **Iris can hear you.** On macOS the microphone was dead: the signed app was
- never given the right to use one, so the system never asked and Skales did not
- even appear under Privacy and Security, Microphone. It asks now, and if the
- answer is no it says so and opens the right settings pane instead of listening
- at nothing. When speech cannot be turned into text, the reason is said out
- loud and written under the eye rather than swallowed: "I could not understand
- you" with what actually went wrong, in place of silence.
-
-- **After the greeting, Iris listens by itself.** No key needed: the ear opens
- when the last word of the welcome has been spoken, the line under the eye says
- so, and it closes again when you stop talking, like any other turn. The space
- bar is still there for when you want it, and holding it down no longer cancels
- the very turn it started.
-
-- **Right-clicking in the Iris window does something.** New conversation, your
- earlier conversations, the voice settings, copy the last answer, ear shut and
- voice off, close the window. There is also a small cross at the top that
- appears while the mouse is moving and fades away when it rests, so the window
- can be closed without knowing that Cmd or Ctrl plus W works. The buttonless
- surface stays buttonless.
-
-- **Iris speaks your language.** The welcome was always English, whatever the
- app was set to, because the greeting was written before the language had
- finished loading. It now follows the app's language like everything else, and
- Iris answers in it too. There is no separate language setting for the voice:
- one language, set in one place.
-
-- **The name question takes Enter.** Answering the "what should I call you"
- prompt with the Return key did nothing at all, and only Skip worked. Typing
- and pressing Enter now answers it, and an empty Enter means "no name, carry
- on".
-
-- **Typing to Iris is findable.** The text field appears when you type and is
- invisible until then, so it was easy to miss entirely. A single quiet line
- says so once, and never again after you have used it.
-
-- **Wake-word training says what happened.** It could fail six different ways
- and reported five of them as "no microphone", including the case where the
- speech detector simply had not loaded. Each one now names itself. There is a
- Stop button that stops it, a Start over button, a live level bar so you can
- see the microphone react while you speak, and a microphone picker for machines
- with more than one input - which the whole voice loop then uses, not just the
- training.
-
-- **A team of agents answers a typed question too.** Setting the Agents button
- above one and then typing ran the message as a single ordinary turn: only
- messages with a picture attached ever reached the team. It also says up front
- when your roster is smaller than the number you asked for, instead of quietly
- running fewer, and an answer that came out of a team says how many agents made
- it rather than naming one model. Chats that ran a team are marked in History.
-
-- **Spoken conversations are marked.** They were always ordinary conversations
- in the ordinary list - which is the point - but nothing said which ones had
- been spoken. History marks them, and reopening one in the chat shows which
- turns were said out loud.
-
-- **The border around a result has no gap, and fits the answer.** The frame the
- particles form was missing its top right edge, and it was built at a fixed
- size before the panel existed, so a two-line answer got a full-page border. It
- is now measured from what is actually there, and the line of text moves with
- it. In a narrow window the eye no longer runs off both sides.
-
-- **The one question about how Skales is going comes back.** Telling Iris your
- name rewrote the identity file, and the "is this install old enough to ask?"
- check read that file's date - so answering Iris reset the clock and the
- question went quiet for another three days. The clock is now set once and
- nothing can move it.
-
-- **Running out of Skales IQ credit is a state, not an error.** A turn that hit
- an empty balance ended in a red error with a provider message in it, and a
- goal or a scheduled task that hit it was marked failed, its work thrown away.
- It now says what happened in your language and offers the way on: put your own
- key in, and carry on. A goal is paused rather than failed, keeps every step it
- had already done, and continues from there instead of starting over. Nothing
- is retried against an empty balance any more, so a background task no longer
- spends attempt after attempt on a request that cannot succeed. A user on their
- own key never meets any of this.
-
-- **Stopping a team of agents stops the one that is already answering.** Counting
- the Agents button back to one ended the run, but an agent already talking to
- its provider carried on generating an answer nobody would read, on a metered
- connection, until it was finished. It now stops mid-sentence. Its card says it
- was stopped rather than claiming it failed, and an answer that had already
- arrived stays where it is.
-
-- **A long message from us is readable on the phone.** Broadcast texts stopped at
- about sixty characters on a phone while the computer showed all of them, and a
- message with no link behind it had nothing to open. The full text is shown, and
- a long one opens out with a tap.
-
-- **A team of agents on the phone no longer gives up over a stray bracket.** If
- the agent writing the plan added a friendly closing line, the plan was read as
- unreadable and the whole run ended, over a plan that was perfectly good.
-
-- **Dragging a link no longer shows you a web address.** Nudging any menu entry
- painted the browser's own drag preview, with the page title and a localhost
- address under it. A desktop app should not show its web insides, and nobody who
- brushes a menu item while scrolling asked to see the address of anything.
-
-- **An agent you edited by hand is not overwritten by a page that was open.**
- Agent definitions are files, and a prompt shortened in the file could come back
- longer the next time the Agents page saved. Every write stamps the file now,
- the bookkeeping the app does after a run merges into the file as it is at that
- moment rather than a copy read minutes earlier, and a save that would clobber a
- newer file is refused and says so instead.
-
-- **The Studio upload to YouTube works.** It read a setting nothing in the app
- ever wrote, so it could not publish at all, whatever you had connected. It now
- asks the same place the rest of the app does.
-
-- **The YouTube tools are offered to anyone who has YouTube.** They only appeared
- for people with a separate API key, and stayed hidden from anyone whose Google
- account already covered it.
-
-- **A saved key shown back as dots can no longer be saved as dots.** Reloading
- Settings filled the Drive, Docs and YouTube fields with a masked stub, and
- pressing Save wrote the stub over the real credential.
-
-- **A decision waiting out of sight is visible.** In Skales Code the transcript
- could grow past a permission card sitting below the fold, with the session
- paused on it and nothing to say so. There is a jump-to-latest arrow now, and
- when something down there needs an answer it becomes a marked badge with a
- count.
-
-- **Skales Code names its sessions.** Every one of them was called "New Chat",
- forever, which made the history a wall you could not find anything in. A
- session is named after the work after its first exchange.
-
-- **Commands find the tools your terminal finds.** `npm` could be "command not
- found" in a Skales Code session while it worked perfectly in your own
- terminal, because an app started from the Dock does not inherit your PATH.
- It does now.
-
-- **A message you queued gets delivered.** Typed while a long task was running,
- it could be dropped when the task finished instead of being sent as the next
- turn.
-
-- **Flow shows a finished page in its own preview.** It would ask for approval
- and then open the page in your system browser, while the preview panel beside
- the chat was already showing that exact file.
-
-- **An agent reads the assignment you actually gave it.** A background task's
- instructions were welded together with the harness's own rules into one block,
- so the agent was free to summarise the lot and work from its summary - which
- is how an assignment came back broader than it was written, with examples
- nobody had supplied. Your words now arrive exactly as written and marked as
- yours.
-
-- **Publishing a subfolder no longer overwrites the homepage.** Publishing sends
- the CONTENTS of a folder, so forgetting to name a destination put a subpage on
- the site root. It happened twice. It now asks where the folder should go.
-
-- **The tool list knows about every tool.** Five that exist were missing from it,
- so anything asking "does this tool exist" got the wrong answer.
-
-- **Discover download figures respect the platform filter.** Filtering to one
- platform still summed the downloads of all of them.
-
-## v12.5.76 - Two Corrections
-
-A light window stays light when you have picked your own colours, and connecting
-a Google Calendar works on an account made after 2022.
-
-### Fixed
-
-- **Light mode with your own accent colours.** Switching to light while a custom
- accent was set left the dark background in place and turned the text dark with
- it: a page you could not read. The colours are worked out again the moment you
- flip the switch, which is the step that was missing. If you are on 12.5.75 and
- see this, "Back to the original" in Settings clears it at once. Nobody who kept
- the shipped colours was affected.
-
-- **Connecting Google Calendar.** Google stopped accepting the copy-the-code
- redirect for any OAuth client created after 2022, so a new client could not
- finish the connection at all. Skales now sends Google back to a small page of
- its own on your machine, which shows the code with a copy button. Step two in
- Settings is unchanged: paste it and you are connected. Nothing to register on
- Google's side, because a desktop client may use any local address. An older
- client keeps working exactly as it did.
-
-## v12.5.75 - Extensions
-
-Skales says out loud that you can build onto it, and the Code window stops
-tripping over its own defaults.
-
-### Added
-
-- **The Extensions pill.** The Code composer has a new toggle next to the
- attachment clip. Switch it on and the session knows, precisely, how Skales is
- extended: how a capability that worked once in chat becomes a permanent tool,
- how an app you build is registered as an API Connector Skales can call, how
- your app calls into Skales over the DevKit remote API, and how an MCP server
- plugs its tools straight in. The contracts are spelled out in the prompt, so
- a small local model can follow them as well as a frontier one. The tooltip on
- the pill says all of this before you press it.
-
-- **Skales knows it is extensible.** Asked "how do I extend you?" in any chat,
- Skales now answers with the real routes above instead of guessing at ports
- and reverse-engineering. The one honest limit is stated too: no injecting
- panels into the Skales window itself; a companion window wired over the
- connector and the remote API is the way.
-
-- **A Debian package.** Debian, Ubuntu and Mint have a proper `.deb` again,
- next to the AppImage that runs on any distro. It is the one to take on
- Ubuntu 24.04 and newer: installing it keeps the Chromium sandbox switched on,
- where the AppImage has to fall back without it.
-
-### Changed
-
-- **Home and Work reach every theme.** Obsidian, Snowfield and Neon had no way
- to switch sides, so the sidebar was stuck on whichever one another theme had
- left behind. All three carry the switch now and group their entries by it,
- under the same headings as the full sidebar.
-
-- **Codework and Swarm start switched off.** Both are a deliberate step rather
- than something to meet on the first launch: one runs commands and writes
- files in a real project folder, the other sends one task to several models at
- once. Turn either on in Add-Ons whenever you want it. An existing install
- keeps exactly what it had. Swarm also moves out of Main and down to Build,
- under WordPress, with the rest of the ways to run a job.
-
-### Fixed
-
-- **An answered question stays answered.** The occasional question Skales asks
- about how you work came back on the next reload, in full, as if you had never
- replied, sometimes days later with the old timestamp still under it. Your
- answer is kept with the message now. The same reload used to make the
- small "shape your agent" card disappear instead; it stays too.
-
-- **Your accent colour reaches the rest of the app.** Around fifteen hundred
- buttons, links, icons, headings, rings and glows were painted in the shipped
- lime by hand and never moved when you picked a colour. They follow the accent
- now, including the tinted glass behind the panels, which takes on your hue
- while keeping its own depth. Green that means something, like a success tick
- or an online dot, stays green.
-
-- **The Skales lettering is the logo again.** Under Obsidian it came out sand,
- under Snowfield blue, under Neon cyan, and elsewhere it followed whatever
- accent you had picked. The wordmark is the brand and no longer takes a colour
- from anything.
-
-- **The stripe on the active sidebar button.** The pill in the default theme
- had a hard line down its left edge in a colour from the far end of its own
- gradient. It is one smooth pill now.
-
-- **The accent belongs to Skales-X.** The other five themes are finished
- designs with an accent chosen against their own palette, and half of each
- stayed behind when a picked colour was walked through it. They keep their own
- accent; the picker says where it lives.
-
-- **A coding session without an agent runs again.** "No agent" is the Code
- window's default for a reason: no persona, no extra prompt, fewer tokens.
- A session started on that default aborted immediately with "Isolated agent
- could not be resolved", and only picking a real agent got it moving. The
- default works now, and sessions broken this way heal on their next message.
-
-- **Code stays in the Code window.** A run in the standalone Code window used
- to follow you into the app: clicking Chat landed you inside the coding
- session, its approval card and live stream included, and the coding session
- sat in the chat History as if it were a conversation. Code sessions now live
- with Code: the chat page, its recents, History and the command palette leave
- them out, and a finished coding run notifies with "Open Code" instead of
- posing as a chat reply.
-
-- **The type size buttons in the Code settings size the type.** Every size in
- the window scales with the chosen number; before, the family changed and the
- size did nothing.
-
-- **The composer's toolbar scrolls instead of wrapping.** The row under the
- input keeps one line at any width; a mouse wheel now scrolls it sideways,
- alongside the drag and trackpad swipe it already had.
-
-## v12.5.71 - Windows You Can See Through
-
-Two corrections to 12.5.7. Both are things you look at every day, and both were
-caused by a rule that was right for the windows it was written for and wrong for
-the ones it also reached.
-
-### Fixed
-
-- **Buddy is see-through again.** 12.5.7 gave each window the right to colour its
- own title bar in the current theme, so Flow and Code stop showing a dark
- caption bar in a light theme. Two windows have no title bar and no background
- at all - the desktop Buddy and the AIPointer overlay - and they were filled in
- along with the rest: a solid rectangle in the theme colour, sitting on your
- desktop. A window whose pixels come from the page is now left alone, and Flow
- and Code keep the caption bar that matches their theme.
-
-- **The Home | Work switch obeys the click, not the page you are on.** The
- sidebar follows you: land on a page only one side shows, and it turns to that
- side so the page you are on is always in the menu. That rule was also
- answering your own switch. On Discover - a Home page - pressing Work set the
- side and the rule put it straight back, so the only way from Discover to Code
- was a detour via a page both sides show. The rule now answers a move between
- pages, once. Standing still, the choice is yours.
-
-## v12.5.7 - Code Gets Its Own Window
-
-Code stops being a page inside Skales and becomes a program of its own: a
-window built for working on a repository, built for that job rather than for
-a conversation. Point it at a folder, watch it read, plan and
-ship the diff, review every change beside the transcript, and tell it - once,
-globally, or per session - how you like to work. It runs the same engine as
-everything else and none of the identity: a coding agent, and nothing else.
-
-### Added
-
-- **Hugging Face works the way it looks like it works.** The token test can
- actually fail now - it checks who you are and then makes one real call, so a
- made-up token stops passing - and a rejected token says which permission it is
- missing, with a link straight to the page that sets it. Activating a Space
- reads that Space's own API instead of guessing at it, and when a Space will not
- answer, the card says the shape was guessed rather than pretending otherwise;
- you can correct the endpoint, the inputs and the output type by hand, or have
- the API read again. Spaces built on current Gradio work, not only the old
- ones. A Space that returns a file returns a picture, a video or a sound you can
- see, saved into your gallery rather than handed to you as a link that expires.
- Speech-to-text and text-to-speech can both run through Hugging Face. Embeddings
- work. Music is no longer one single model. A dedicated endpoint is tested as
- the kind of endpoint it is instead of always as a chat. And the model catalogue
- is read properly at last: prices, context sizes, tool support and the real
- providers behind a model, instead of a list of the people who uploaded it.
-
-- **Remote MCP servers speak the current protocol.** Skales now offers the
- 2025-06-18 revision and falls back to 2024-11-05 for a server that has not
- moved yet, so servers that had begun refusing the older revision connect again.
-
-- **Skales Code has a terminal, and it is yours.** A real one, under the box you
- type in and dragged to whatever height you want: your login shell, in the
- session's folder, with your profile. Not a command field, so vim, top, an
- interactive prompt, colours, arrow keys and Ctrl-C all behave the way they do
- in a terminal. Several tabs if you want them, a switch in the title bar, and
- /terminal. It is yours in the sense that matters: the agent can neither see it
- nor type into it, so it needs none of the agent's rules, and what the agent
- runs has its own place further down.
-
-- **A command that never ends can run in the background.** A dev server, a
- watcher, a long build used to be a choice between a timeout and a blocked
- session. Now it starts, hands back a handle, and keeps going: its output
- appears under the step that started it and grows while you watch, and a strip
- above the status line lists everything that is running with what it is, how
- long it has been going, and a stop button. Nothing limits how many run and
- nothing times them out. Stopping really stops it, the whole tree, so a port
- actually comes free. A session that ends ends its own, and so does closing
- Skales. Your phone can see the same list and press the same stop.
-
-- **Call any API from Skales.** A tool for talking to a REST API or submitting a
- form: your method, your headers, your body, the answer back. A form-shaped
- body is sent as a form, which is the difference between a request that works
- and one that quietly does nothing. Reading runs, writing asks first. Your own
- network is reachable, a NAS or a second machine or a service in your tailnet;
- this computer itself is not, because that is where Skales listens, and a page
- or a document could otherwise talk it into calling the app. A switch under
- Security opens that too, and says why it is shut.
-
-- **A template for Code opens Code.** Clicking a coding template used to open Lio
- AI, and the card said so. It opens Skales Code now, with the prompt already in
- the box. Five new ones for the work a developer actually brings: find and fix
- this bug, review my current changes, explain this codebase to me, refactor this
- safely, find what is untested and test it. Each of them insists on running the
- thing rather than claiming it.
-
-- **The instructions box shows what belongs in it.** The place where you tell
- every coding session how you work was an empty field with no example. It now
- hints at the four things that earn their place there: the language to answer
- in, your own test command, your commit style, your no-gos. A hint, not a
- filled-in file: rules nobody wrote should not ride along invisibly.
-
-- **Working in Code shows up in Discover.** The biggest thing in this release had
- no place in the feed at all. A coding session, a change shipped out of the
- window and a run with several agents now do, in the same content-free shape
- the Studio entries have: never a repository, a branch, a path, a commit message
- or a file count. The feed is public and the work is private.
-
-- **Code is its own program now.** Clicking Code opens a window that looks and
- behaves like a tool for writing software, not like a chat with a folder
- attached. Its own title bar, its own type, its own light and dark, and a
- status line along the bottom that always says what is going on: idle or
- working, which folder, which branch, and how many lines it has added and
- removed. Point it at a repository and say what you want changed. You can open
- a project on this computer or clone one from an address, an ssh address
- included, and the sessions you had open before are listed underneath with the
- branch and the size of the change in each. Four ways to work: Ask reads and answers
- and changes nothing, Code makes the change, Plan writes out what it would do
- first, Auto gets on with it. Pick the model and how hard it should think right
- there in the box you type into.
-
-- **A coding session reads like a log, not like a conversation.** The Code
- window shows the whole session in one column: what you asked, what it
- answered, and every step in between as its own line. Read, Grep, Edit, Bash,
- with the file or the command beside it and how many lines it added and
- removed. An edit shows the change right there, the lines it replaced in red
- and the new ones in green, with the line numbers from the file. A command
- shows its output in a terminal block underneath. The checklist it is working
- through updates in place, with the finished items struck out. It fills in as
- it happens rather than all at once at the end, and scrolling up to read
- something earlier no longer drags you back to the bottom every second.
-
-- **Your coding sessions, down the side.** Code has a column you can open from
- the title bar: a new session at the top, a search that looks at both the name
- and the repository, and everything you have worked on grouped into today and
- earlier, with the repository and what changed under each. Right-click one to
- rename it, put it away, duplicate it, download the whole session as a
- readable file, or delete it. A duplicate brings the conversation with it, so
- you can try a second approach without losing the first, and the download
- includes the diffs, because a record of a coding session without the changes
- in it is a record of nothing. A new session is named after the first thing
- you ask it rather than after the folder, which the row already tells you.
-
-- **Review what changed, then ship it, without leaving the window.** Open the
- review panel from the title bar and the changed files sit beside the
- transcript. Pick one and you get it three ways: the change itself with line
- numbers, the file as it now reads, and, for something worth looking at like a
- markdown file or an image, a real preview. Keep it or put it back, write a
- commit line, and open a pull request: the branch is pushed and the request is
- created for you, with GitHub's own words back if it refuses. It needs the
- GitHub command-line tool signed in, and says so plainly when it is not there
- instead of just failing. Underneath, everything this session changed with a
- way to undo any single one of them, and the repository's other checkouts.
-
-- **Point at a file, drop in a screenshot, or just type a command.** In Code,
- typing @ offers the files of the repository you are in, matched on the whole
- path so a folder name finds them, and the file goes along with your message
- so it does not have to go looking. Drop files onto the box, paste a
- screenshot, or use the paperclip: video, audio, PDFs and archives all attach
- the way they do in a chat, with the chat's limits and no new ones, and they
- show as chips above what you are typing, with their size and a thumbnail for
- a picture. A slash at the
- start of the line opens the short list of things a coding session can do:
- switch how it works, open the review panel, commit, open a pull request, stop
- it, start a new session. A slash anywhere else is just a path, and stays one.
-
-- **When it stops to ask, the question is right there.** A question from Code
- now appears in the session itself, with its options numbered so you can
- answer with a keypress, or write something else entirely in the box below.
- When a tool wants permission, the same: what it wants to run, and Allow once,
- Deny, or stop asking about that tool for this session. While either is open
- the header and the bottom line both say it is waiting for you, and the box
- says so too instead of quietly doing nothing. A question you already answered
- stays readable but can no longer be clicked, so an old card cannot answer a
- question the session has long moved past.
-
-- **The status line now counts the context, and a picture you sent shows as a
- picture.** The bottom line of the Code window shows how much of the model's
- window the conversation occupies. A message that carried screenshots or files
- shows their names as small chips instead of raw text, an agent's log can be
- opened from its card, the review panel says when a very long change was cut
- short instead of ending silently, and a message typed in the last moment of a
- run is carried into a fresh run instead of quietly disappearing.
-
-- **Answers read the way they were written.** Text from Code renders properly
- now, so a filename in code marks, a bulleted list of findings or a snippet
- comes out looking like what it is, rather than showing you the asterisks.
-
-- **Watch several agents work at once, and see what they cost.** When a coding
- session splits a job across parallel agents, Code shows them down the right:
- what each was asked to do, how long it has been going, how many tools it has
- used, and its answer when it has one. Underneath, what the run has spent, split
- between the main conversation and the agents, with a total. Stop all is in the
- header while any of them is running. A figure that was never recorded says so
- rather than showing a zero that would read as free, and from now on those
- figures are recorded, including for a run that spent them and then failed.
-
-- **Code is a coding agent, and nothing else.** A session in the Code window no
- longer carries Skales' personality, its interests, or anything it remembers
- about you. It is told where it is, what it may change, which shell this
- machine runs, and how to work, and that is all. Measured on a real request:
- the instructions it sends dropped from about 13,800 characters to about
- 4,200, which is roughly seventy per cent less on every single turn. It keeps
- the whole toolset, because a coding agent with fewer tools is a worse one;
- what it loses is the handful that read or write your memory, refused both in
- what it is offered and in what it may run.
-
-- **Tell it how you like to work, once.** The gear in the Code window holds
- instructions that apply to every coding session, here and behind the /code
- command in chat, plus instructions for just this session when one job needs
- something different. Alongside them it shows what the project itself asks for
- in its own CLAUDE.md or AGENTS.md, read-only, because that file belongs to
- the repository. The more specific one wins. The gear also holds light and
- dark for this window, and follows the system if you would rather it did.
-
-- **A phone will be able to tell a coding session from a chat.** The session
- list Skales sends to a paired phone now carries, for a coding session, the
- folder it is in, how it is working, whether it is running or waiting for you,
- and how much it has changed. Nothing on the phone uses it yet; it is there so
- that when it does, it draws the same row this Mac draws instead of a second
- version of it.
-
-- **It hears you while it is working.** Type while a session is running and the
- message waits in line instead of being refused; it is picked up at the next
- step and the count of what is waiting is on screen the whole time. Everything
- the window can do it does through the same channel a phone will later use, so
- the same session can be watched and steered from somewhere else without any
- of it being built twice.
-
-- **Code sessions show up again on a machine you have been using a while.** The
- list of coding sessions was empty for anyone whose conversation list had been
- built before Code existed, because the saved list did not know what a bound
- folder was. It is rebuilt once, and they are all there.
-
-- **Code moved up, and can be put away.** Code now sits in the main part of the
- sidebar, directly above Studio, where the work you do every day lives. And
- like Studio it is an add-on: if you do not write code, switch it off under
- Add-Ons and the entry is gone. It is on to begin with.
-
-- **Code: a place to work in a folder.** A new entry in the sidebar opens Skales
- on a folder of your choosing, in its own window, so you can keep chatting and
- using everything else while it works there. It lists the folders you have
- worked in before, and you pick how it should behave: Code changes files and
- asks before anything risky, Plan only reads and shows you what it would do,
- Auto gets on with it. It is the same engine that has been behind the /code
- command in chat, with a surface of its own. Pick a session and you see the
- repository it sits in: the branch, how far ahead or behind it is, every
- changed file, and a box to commit them. A folder that is not a repository
- says so rather than showing an error, and when git refuses a commit you get
- git's own words back, not a generic failure. Click any changed file and you
- see what changed in it, with Keep and Put back: Keep stages it for the next
- commit, Put back restores the file the way it was and tells you where it took
- the old version from. A brand new file shows its contents instead of an empty
- comparison. Alongside it there is a list of every file this session wrote,
- changed or deleted, each with its own undo, and the ones you already undid
- stay in the list so the record stays complete. And you can work on two things
- at once: give a second piece of work a name and Skales checks the repository
- out again in a folder of its own, on a branch of its own, with its own
- session, so neither run can overwrite the other. Each session says whether it
- is working, waiting for an answer from you, or idle. Removing a parallel
- checkout tells you when there is uncommitted work in it rather than taking it
- with the folder, and it will never remove the repository itself.
-
-- **Your accent colours stop at the logo.** Custom accents recolour buttons,
- links, glows and selections; the Skales lettering in the sidebar keeps its own
- colours in every theme, because that is the brand, not a control.
-
-- **Your own accent colours.** Under Appearance you can now set the three
- colours the Skales gradient is made of, and everything drawn in the accent
- follows: buttons, links, the glow, the selected row, the wordmark, the command
- token in the composer. Skales adjusts the brightness where it has to so the
- result stays readable on the page you are on, tells you when it did and by how
- much, and one click puts the original colours back. The window frame of the
- Flow and Code windows now follows the theme as well, instead of staying dark.
-
-- **A team plan shows who waits for whom.** The team leader has always decided
- which subtasks can run at once and which have to wait for earlier results, and
- the run has always followed that - but the plan was printed as a flat list, so
- five agents working in parallel looked exactly like five agents working in a
- chain. The plan is now drawn in stages, each stage saying how many run
- together, and a subtask that waits names what it waits for.
-
-- **Skales says when it shortens a long conversation.** To keep a long chat
- inside the model's limit, Skales summarises the older part of it. That has
- always happened quietly, so from the outside it just looked like the assistant
- starting to forget things. It now says so in the conversation, with what was
- shortened, and makes clear that nothing was deleted.
-
-- **Sub-agents get a role instead of everything.** When Skales splits a job into
- parallel sub-agents, each of them was handed the complete tool set: a
- sub-agent researching competitors could send WhatsApp messages, write to your
- long-term memory, stop to ask you a question, and start sub-agents of its own
- without end. Three of those are now impossible for a sub-agent whatever else
- changes, and a sub-agent that needs a capability it was not given can ask for
- it mid-job and carry on. The narrower tool sets themselves are off by default
- for now and can be switched on in settings; while they are off, every request
- a sub-agent makes is recorded, so the defaults are chosen from what jobs
- actually needed rather than from a guess.
-
-- **Your own speech server, for reading aloud and for dictation.** There was a
- field for a custom text-to-speech address, and nothing in Skales ever read it:
- it was saved, carried along and backed up, and it never produced a sound.
- Settings now has a Custom voice provider with the four things such an endpoint
- actually needs, an address, a model, a voice and an optional key, and the same
- for transcription. Both are used before any cloud service, because a server
- you set up yourself is the one you meant. An address you had already entered
- in the old field starts working on its own, and if your endpoint does not
- answer, Skales says so instead of quietly reading in a different voice from
- somewhere else.
-
-- **Folders Skales must never touch, whatever else you allow.** Until now the
- only way to limit file access was to list what Skales *may* reach, which is no
- help if you run it with Full Access on purpose. There is now a list of folders
- it must stay out of, and that list is not lifted by Full Access or by
- Unrestricted mode. It covers the file tools, the shell and scripts, search,
- and directory listings, so a blocked folder is not just unreadable, it is
- invisible; it cannot be picked as a project folder either. Shortcuts pointing
- into a blocked folder are followed before the check, so they are not a way
- around it. One button blocks Skales' own program folder without you typing a
- path, which also means it keeps working after you move the app.
-
-- **Every step in a coding session opens.** A line in the Code window is no
- longer something to read past. Click a Read, a Grep, an Edit or a command and
- it unfolds: exactly what was sent, and everything that came back, whole,
- scrollable and with a copy button on each half. The line itself now also says
- how long the step took, so a slow turn shows you where the time went.
-
-- **A change decides itself where you see it.** The diff under an edit carries
- its own added and removed counts and its own Keep and Revert, so reviewing no
- longer means finding the same file again in the panel.
-
-- **Every file is a link.** A path in a tool line, a path an answer mentions in
- backticks, a file in a diff: click it and it opens in the panel beside the
- transcript. That includes the Markdown ones.
-
-- **Something to do with a turn.** Under every answer: copy it, have it read out
- loud, fork a new session from that point, or roll the whole session back to
- it, files included. Under your own turns: change what you asked and send it
- again. Your own words now sit on the right, in a block of their own.
-
-- **A page you wrote, running beside the transcript.** Click an HTML file in a
- coding session and the panel does not show you its source, it shows you the
- page: its own stylesheet and script load with it, so a widget or a small app
- works there the way it will work anywhere. Markdown is rendered, images are
- shown, and everything else is now coloured instead of one grey block.
-
-- **Point at something in a running page.** Turn on the picker, hover the page,
- click what you mean, and the element goes onto the message you are writing:
- what to call it in CSS, its own markup, and a picture of just that piece. So
- "change this one" is a thing you point at rather than describe.
-
-- **Chat or Split, from the title bar.** The panel used to be reachable only
- when git had noticed a change, which has nothing to do with wanting to look at
- a file. Now it is a switch, and it is there whenever a session is.
-
-- **Hand a coding session over to Code.** A chat that is bound to a folder and
- running in a code mode gets one button in its header: it stops what is
- running, cleanly, and opens that same session in the Code window. Not a link
- and not a second copy. The same session, in the window built for it, with
- your chat left exactly where it was.
-
-- **The landing says something different every time.** Twenty-two greetings, in
- every language, in rotation rather than at random, and the headline flickers
- for a moment every few seconds. It holds still for anyone who has asked their
- computer for less motion.
-
-- **Dictate into the box.** A microphone in the composer, running the same
- speech-to-text everything else in Skales runs. What it hears joins what you
- already typed instead of replacing it.
-
-- **Which agent runs the session, next to the model.** By default: none. An
- agent carries a system prompt and a persona, and a coding session should have
- neither unless you deliberately picked one. If you do pick one, it is that
- agent's prompt from the next turn.
-
-- **The Skills button says what else is connected.** Which MCP servers are on
- and how many tools each one brings, and which connectors are set up, instead
- of leaving you to wonder whether any of it reached this window. A step that
- went through an MCP server now says which one.
-
-- **A message you send from your phone while it is working joins the queue**
- instead of starting a second turn on the same session.
-
-- **Two files for your standing rules, not one.** SKALES.md is still how you
- want to be answered, everywhere. AGENTS.md is new and is about code only: the
- style, the checks to run, the folders never to touch. One editor in Settings,
- a switch above it, and both are ordinary files you can open in a text editor.
- A project's own CLAUDE.md still wins for that project.
-
-- **Look at what you published.** Two new tools alongside the FTP upload that
- has been there for years: list a directory on your server, and download a file
- from it. Same saved profiles, all three protocols, and an agent working on its
- own website can finally read its own files back.
-
-- **Every coding session gets a scratch folder of its own.** Somewhere to put a
- captured log or a half-finished file that is not part of your repository and
- should never show up in your diff. It is emptied when you clear the session and
- deleted when you delete it.
-
-- **`/clear` in Code.** Empties the session and its scratch folder and keeps the
- repository, because starting again almost never means starting somewhere else.
-
-- **The sidebar has two sides now: Home and Work.** A switch under the provider
- status, and the menu regroups: Home leads with Chat, History, Discover, Memory
- and Lio AI and keeps Browser, Group Chat, Spaces and Wrapped together; Work
- leads with Code, Organization, Autopilot, Teams and Swarm and puts Codework,
- Browser, Playbooks, Workflow and WordPress under Build. Workspace, your own
- widgets and System stand in both. Your choice is remembered, and collapsed the
- switch becomes one button showing the side you are on.
-
-- **Skales asks once how you want to use it.** A single question after the
- privacy notice, with two equal answers, and all it decides is which half of
- the menu you see first. Skippable, and changeable any time.
-
-- **Code pulses once, quietly, if you have not opened it yet.** It stops the
- first time you click it and does not come back in the next release.
-
-- **Hand a design to a coding session.** The paperclip in Code now offers a
- Studio or Flow project as well as files: pick one and the session gets a
- snapshot of it, the markup, a list of the files it is made of, and a picture
- of how it looks. Read-only, so the session can match it and never write into
- it. So the thing you designed and the thing you are building can finally see
- each other without an export in between.
-
-- **No emoji in Code, at all.** Where a step or an answer arrives carrying a
- tick, a cross or a warning sign, the window draws its own icon and keeps the
- meaning. Everything else is removed, and Code is told not to put emoji in the
- files it writes either.
-
-- **Code follows Skales, not your Mac.** Light stays light and dark stays dark
- because of what Skales is set to, and it changes the moment you change it,
- even with the Code window open beside the main one.
-
-- **Skales Code has a browser, and the project's files beside it.** A page the
- agent just wrote, a dev server it just started, a file you want to look at:
- there is a panel for that now, next to the transcript rather than instead of
- it, with an address bar, back and forward, reload and open-in-your-real-
- browser. The file column lists what is in the folder and opens it with a
- click. Its own panel, because "what changed" and "what does it look like" are
- two questions and the review panel only answers the first.
-
-- **A page printed in the answer can be run where it stands.** When Code writes
- a whole HTML page and shows it to you, the block carries two tabs: the source,
- which is still what opens, and the page itself, running. No copying it out to
- a file to find out whether it looks right.
-
-- **The panels on the right can be dragged wider.** Both of them, from their
- left edge, with the mouse or with the arrow keys, and each window remembers
- what you set.
-
-- **New coding sessions start on what you choose.** The gear in Code now holds
- the defaults: which model a new session opens on, which agent runs it, and
- which working mode it starts in. Nothing there touches a session that is
- already open. Beside them, how the window reads - the typeface and the size -
- which stays on this machine.
-
-- **Code says what it is allowed to do.** The settings panel now states, in
- words, that a coding session runs on the working mode you pick in it and not
- on the safety mode in Settings, what each mode asks about, and the five things
- no mode lifts: the folder is never left, Skales' own files are never
- writable, commands that cannot be undone are refused, and a coding session can
- neither read nor write your memory.
-
-- **Skales itself can run a coding session.** The agent list in Code offered
- every agent on the machine except the one the chat window runs on. It is
- there now, second in the list, for when you want Skales with its memory on a
- repository rather than the lean coding identity.
-
-### Changed
-
-- **All eighteen AI providers are on the providers page, not four.** OpenAI,
- Anthropic, Google, Groq, Mistral, DeepSeek, xAI, Together, Minimax, Moonshot,
- GLM, Qwen, Cloudflare and Nvidia were folded into a collapsed "More LLMs"
- section, so the one page that answers what Skales can talk to showed you an
- aggregator, a local runner, a text field and a router. They are all listed
- now, in a sensible order, and the "show only active" checkbox above the list
- is there when you want it short. The cards themselves used to be built twice,
- in two places that had drifted apart three times in past releases; they are
- built once now, so a button can no longer exist for half the providers.
-
-- **Lio AI now has an address that matches its name.** The sidebar entry read
- "Lio AI" while the page it opened lived at /code, so links, bookmarks and the
- assistant's own answers about where things are all pointed somewhere that
- said something else. Lio AI is at /lio from now on. An old /code link still
- takes you there, so nothing you saved stops working.
-
-- **Code asks you, not your safety mode.** The leash is the mode you picked when
- the session started: Ask and Plan read, Code asks before anything risky, Accept
- edits applies file changes and still asks for the shell, Auto runs. Set the
- rest of Skales to Unrestricted and Code still asks, because that is what
- choosing Code meant. Nothing that is not a question changes: the blocked
- folders, staying inside the bound folder, the command block and the shell rules
- were never approvals and are never lifted.
-
-### Fixed
-
-- **A mailbox that refuses your password by design now says which code to get.**
- QQ never accepts an account password over IMAP, it issues a separate
- authorisation code, and Skales answered "verify credentials", which sends you
- to check the one thing that was already right. QQ now gets the sentence that
- helps, in the words its own settings use, and so do 163 and 126, GMX, Web.de,
- iCloud, Zoho and Yandex: each names what to get and where. A provider that is
- not on the list at least learns that this is a thing that happens. And a
- failure that arrived as one word, "AggregateError", now says what actually
- went wrong on each address it tried.
-
-- **The mail test button works before you save, and says what it does.** It was
- dead until the account was stored, which is backwards: testing is what you do
- before committing settings you are unsure about. It now checks what is in the
- form, and it says out loud that it only checks the login, sends nothing and
- changes nothing in your mailbox.
-
-- **Searching a project no longer stops everything else.** Skales runs on one
- loop and every chat, poll and button shares it. A search across a source tree
- held it for a quarter of a second in one block and listing a project held it
- on every session open, so a running conversation stuttered for reasons that
- had nothing to do with it. Search, listing, reading and the project walk now
- give the loop back while they wait.
-
-- **You can watch the browser work instead of seeing the pictures afterwards.**
- A screenshot only travelled with the finished tool result, so a browsing job
- of a dozen steps left the chat empty for a minute and then dropped every still
- in at once. The current picture now appears above the composer while the job
- runs and updates with each step, and it disappears when the run ends rather
- than sitting there looking live. The screenshots in the conversation are
- unchanged.
-
-- **Skales can offer to turn a long answer into a goal, and it asks first.**
- The automatic goal start could only judge a message before it ran, so a short
- request that turned into fourteen open steps and thirty tool calls was never
- recognised as the bigger job it became. When a finished turn ends with several
- open steps, a lot of tool work, or an unfinished checklist, a line above the
- composer offers to carry it on as a goal. It never starts by itself: one click
- runs it, one click makes it go away, and doing nothing leaves the chat exactly
- as it was. When you have switched on the automatic start, a promoted message
- now says so in the conversation instead of only in a log, and neither half of
- this runs inside an isolated agent.
-
-- **A team task that stops moving ends, and tells you where it stopped.** If a
- step never came back, the task stayed on "running" for as long as the app was
- open, and the log only listed which tools had been called, never what came of
- them. A step that produces nothing for six minutes is now given up on and
- marked as stopped, which is its own outcome and reads differently from a step
- that failed. The log lines carry the result of each tool, and a failed one
- says what went wrong right there in the timeline. The clock restarts on every
- answer, so a long task that keeps working is not interrupted.
-
-- **The updater cannot install a version that is no longer the current one.**
- It checked for updates every four hours and then treated that answer as
- settled: the download took whatever the last check had named, and the install
- checked nothing at all beyond the file being there. On release day that meant
- two machines installed the previous version half an hour after the new one had
- gone out. Skales now re-reads the release feed before it downloads and again
- before it installs, and if the download has been overtaken it says so and
- offers the current one instead of installing the old one. When the feed cannot
- be reached, an already-downloaded and verified update still installs, and the
- log says it could not be re-checked rather than pretending it was.
-
-- **A team task keeps its files together instead of on your Desktop.** The
- agents in an Execution Team were never given a place to work, and their
- instructions literally told them to save to the Desktop or the home folder,
- which in unrestricted mode is where the files ended up. Each run now has its
- own folder under the workspace, the agents are told to write there, and the
- run says which folder it is before it starts, so you can find the result.
-
-- **Flow keeps its Render button after the first render, and you can start one
- from the chat.** Once a Motion project had been rendered, the button was gone:
- the preview had switched to the finished video, and the button only appeared
- while an HTML composition was on screen. It belongs to the project now, not to
- whatever file you happen to be looking at, and clicking it puts the
- composition back in the preview so you see what is being rendered. When it
- cannot run, it stays visible and greyed out with the reason on it, including
- on a machine where FFmpeg is not installed, instead of quietly not being
- there. The same action is now on the file card in the conversation, where the
- round that produced the composition is.
-
-- **The Execution Team room is visible in a light theme.** The scene was drawn
- in white on the assumption of a dark room, so in Snowfield, or in any theme
- set to light, the desks, the people, the monitors and the names were white on
- white and the panel looked empty while a team was working in it. It is drawn
- in your theme's colours now and follows a theme switch immediately. The same
- thing hid your own dot on the Discover network map; it shows up too.
-
-- **"Fetch Available Models" works on providers where it never could.** The
- button asked the provider directly from the window you were looking at, and a
- browser refuses that unless the provider explicitly allows it. Most do not, so
- the request never left your machine and the card reported a failure on a key
- that was fine. Nvidia NIM was where it was noticed; it was true for a dozen
- providers. The request now goes out from Skales itself, and when a provider
- does refuse, the card shows you what the provider said instead of a generic
- error. The model picker in chat was quietly affected by the same thing: it
- promised your actual models and fell back to a fixed list without saying so.
-
-- **The Debug Panel works again on all five of its tabs.** Memory, Sessions,
- Tools and Status all answered with an authorization error, because the panel
- called the developer API with a token it had guessed rather than the one in
- your own DevKit config. Only the Adaptive tab worked, and only because it was
- the one tab that read the app directly. All five read the app directly now, so
- the panel shows you your machine whether or not the developer API is switched
- on. The API Playground stops guessing too: it fills the token field from your
- config, and says so plainly when there is no token to fill it with.
-
-- **Settings no longer shows you the name of a text instead of the text.**
- Four provider cards printed something like settings.providers.nvidia_nim.desc
- where their description belongs, and all six theme cards in Appearance showed
- the word "Appearance" as their description. A few other places did the same
- thing: the standing instructions panel, the day picker in Planner, the refresh
- button on the Autopilot board, and four of the twelve voice languages in
- Studio. All of them now say what they mean, in all twelve languages, and a
- check now fails the build if a text is referenced that no language actually
- has, so this cannot come back unnoticed.
-
-- **The desktop buddy now wears the theme you picked, and its bubble can no
- longer be cut off.** Four things were wrong with that little window at once.
- It ignored the theme system completely: it carried its own fixed set of
- colours and reduced every theme you can choose to a single light-or-dark
- guess, refreshed at most twice a minute, so it never matched the app and
- lagged behind a theme switch. It now reads the same colours as everything
- else and follows a change the moment you make it. In a light theme the
- "Open Chat" link was white text on a white bubble - invisible until you
- hovered it, which made it whiter; the speech tail stayed dark under a light
- bubble too. Both follow the theme now. The action button on a notification
- ("View Tasks", "Open calendar") did nothing at all when clicked; it now opens
- that page in the main window, and a notification without a page opens the
- chat rather than nothing. And the window itself was a fixed height that could
- not grow, with the page hiding anything that did not fit - a long answer was
- not scrolled, it was cut. The window now grows to fit what it has to say,
- upward from wherever you parked it, never past the edge of your screen, and
- goes back to its normal size when the bubble is gone. If a message is longer
- than the screen can hold, the bubble scrolls instead of hiding the rest.
-
-- **A chat that sends out sub-agents now shows them working while they work.**
- Handing a job to parallel sub-agents puts a live status card under the message
- that started it, one chip per agent. That card only turned up after leaving the
- chat and coming back, or after a reload, so a job that was very much running
- looked like nothing had happened. The launch signal never reached the open
- chat: it travels on the same internal channel as the proactive dashboard
- notes, that channel is read-once, and a single unreadable item made the read
- throw away the entire batch. The signal was gone before anything could act on
- it. The channel now hands over exactly what it promises, and one bad message
- can no longer take the rest down with it, so the card shows up within seconds
- and its chips move as the agents finish. Two more things that ride the same
- channel come back with it: the "job finished" notice with its per-agent
- report, and proactive dashboard messages appearing in chat. And a job sent off
- in the very first turn of an empty chat, where there is no message yet to hang
- the card on, now shows it above the input instead of nowhere at all.
-
-- **A click in the built-in browser no longer gives up because the page redrew
- itself.** Reading a page marks every element so the agent can act on it
- exactly, but a site that rebuilds itself while it works - most modern ones do,
- constantly - throws those marks away, and the click then failed on everything,
- including a site's own logo, while typing carried on working. Skales now also
- remembers what each element WAS, its role and its label, which survive a
- redraw, and falls back to finding it that way. If even that misses, the
- remembered label is handed to the two strategies that were already there and
- were never reached before, instead of the attempt ending.
-
-- **The browser window that opens by itself now says whose it is.** A visible
- browsing run opens a real browser window next to Skales, and then it starts
- clicking and typing on its own. Nothing said where it came from, so closing it
- was the obvious reaction - which ends the run. Skales now says, once, that the
- window is its own doing, that it will move by itself, and that it should be
- left alone until it finishes. The sign-in flow says the same, and now also
- points at the chat for telling Skales you are done, rather than leaving people
- hunting for a button in Settings.
-
-- **"Flow finished" opens Flow instead of taking over the window you are in.**
- Flow runs in a window of its own, but the notification's button navigated the
- current window, so the main Skales window turned into the Flow workspace.
- It now opens or brings forward the Flow window, the way the Studio tab
- already did.
-
-- **A rendered video no longer reports a check that never ran.** Validating an
- MP4 needs ffprobe, which was looked up on the system path - and an app started
- from the Dock does not inherit the path from your terminal, so it was
- regularly not found. The check then fell back to "the file is not empty" and
- still showed a green Validated, next to a resolution of 0x0, a length of 0.0s
- and 0 fps. Skales now looks for ffprobe where it looks for ffmpeg, including
- the places a manual install puts it, and when it genuinely cannot be found the
- card says the file was only measured by size instead of claiming a verdict.
-
-- **The settings panel in Code stopped loading and started answering.** It sat
- on "Loading..." forever with nothing loading, because the instructions could
- not be read at all. They can be read now.
-
-- **Tooltips in Code are no longer cut off.** The ones on the title bar opened
- upwards, out of the window, and the ones by the box you type in were painted
- over by the bar beneath them. They are drawn above everything now and nudged
- back on screen at the edges.
-
-- **Code is quiet.** Every reply used to raise a notification and play the
- chime, in the coding window, about the coding window. Worse, the button on it
- navigated that window away to the chat and took the session you were watching
- with it. Notifications belong to the main window; Code does its work without
- interrupting you.
-
-- **The agent list in Code shows four and scrolls.** It opened as a column of
- every agent you have, over the transcript you were choosing it for.
-
-- **A model picked in the Code window is the model that runs.** The Code model
- in Settings quietly overrode it, so the picker showed one name while the run
- used another. The setting still applies where it should: to /code in chat, and
- to a session that has no model of its own.
-
-- **A Discover mention obeys the box that says it should not arrive.** With
- notifications muted and both Discover types switched off, mentions and replies
- still filled the dashboard. They were arriving under a name nothing here
- recognised, so every Discover setting let them past. Three places read that
- list and all three go through one check now.
-
-- **An emoji in Skales is Skales' emoji.** The agent faces in Code were drawn by
- the operating system, so the gecko looked like Apple's gecko. And a lookup
- fault meant most hearts, hands and weather fell back to the system everywhere
- in the app, not only there. Both are fixed, and the coding agent wears a
- laptop, which is a face Skales can actually draw.
-
-- **The sidebar breathes evenly.** The gap under the provider box now matches
- the one above it, and with the sidebar collapsed the Home button is exactly as
- wide as the box above it.
-
-- **Chat and History are in Work as well as Home**, in the same place, and the
- sidebar always shows the page you are on: opening Discover from a notification
- no longer leaves the menu with nothing marked.
-
-## v12.5.6 - An Answer You Can Actually Give
-
-Being asked a question is no use if the form has nowhere to put your answer.
-Skales could ask you four things and accept none of them, and the setup screen
-for browsing the web reported itself ready on the same page that said it was
-not. Both were the same kind of fault: something reported a state instead of
-checking it.
-
-### Added
-
-- **Skales can speak on your phone in a voice from this Mac.** This computer has
- dozens of voices installed, the premium ones included, and they cost nothing
- and never leave your own two devices. The paired app (2.5.6 and newer) offers
- them under Voice: pick one from the list this machine reports and it speaks
- for the phone over the pairing connection. Nothing is played on this
- computer's speakers - the phone asked for the audio, not for the room to hear
- it - and if this machine goes away mid-sentence the phone falls back to its
- own voice rather than going quiet. The same is available locally at
- /api/tts/say. A computer that is not a Mac says so instead of reporting an
- empty list of voices.
-
-- **The Organization scene shows you who is stuck.** The office under a running
- team was decoration: wood-grained desks, a steaming mug, a bouncing ball you
- could kick. With six agents working, none of it told you which one had stalled
- without reading six labels. It has been redrawn around that one job. Colour
- now means state and nothing else - the furniture is near-monochrome and a pool
- of light on the floor under each desk carries the agent's status, readable
- across the whole panel before a single word is. The monitor is the status lamp,
- breathing while there is work and dim while there is not; the ring around an
- agent grows when the state changes instead of blinking, so a glance mid-change
- still reads correctly; and the chip under each desk shows the tool actually
- running rather than the word "Working". The mug, the bookshelf, the clock and
- the kickable ball are gone with it.
-
-- **Instructions you write once, followed everywhere.** Telling Skales how you
- want it to work - answer in German, keep it short, never open with a
- compliment - lasted exactly one conversation, and the next chat started
- ignorant again. Code mode could read a project's own CLAUDE.md, but only with
- a folder bound, and that file is about a project rather than about you.
- Settings, under Chat & Code, now has a box for standing instructions. What you
- put in it travels with every conversation, with tasks and schedules that run
- while you are away, and with custom agents you built yourself. Two places
- leave it out deliberately: an isolated agent, which carries nothing of yours
- by design, and the internal calls Skales makes to itself, whose answers are
- read by the app rather than by you. A project's own CLAUDE.md is the more
- specific instruction and still wins for that project.
- The box is not the only way in. It is backed by a plain SKALES.md file in your
- data dir, the path is printed under the box, and editing it in a text editor
- is the same edit: the box re-reads the file whenever you come back to the
- window. If it changed on disk while you had unsaved text here, it says so and
- refuses to save over it, so the two ways of editing cannot quietly eat each
- other's work.
-
-- **Right-click a key field and paste into it.** A password field in Settings
- gets its own menu - copy the value while it stays masked, reveal it, remove it
- - and the app suppresses the ordinary Cut/Paste one so the two do not draw on
- top of each other. Between the two, nothing offered to put a key *in*: the
- menu opened, showed three things you could do to a key you already had, and
- none of them helped when the key was sitting on your clipboard. Paste is in
- that menu now. It respects a selection, so pasting over a highlighted key
- replaces it, and the keyboard shortcut keeps working as it always did.
-
-### Fixed
-
-- **A page you visit cannot reach the Skales API on your own machine.** The
- local server is reachable from any page open in your browser, and that was
- being closed one route at a time: the same-origin check sat on 2 of 225
- routes. There is one shared check now, in front of every /api request, and it
- refuses exactly one thing - a caller that CLAIMS an origin which is neither
- this server nor your own machine, which is the shape a page on another site
- makes when it fetches your local port. Everything that is not a browser is
- untouched: the messaging bots, the relay, the CLI, the phone over your LAN and
- the app talking to itself all keep working, and each of them has a test. The
- image context menu had the same hole in miniature and no longer opens a menu
- for a message no frame on the page sent.
-- **Discover stops passing its own template off as your model's work.**
- Composing a post gave the model 45 seconds, then quietly substituted a local
- phrase bank and returned it with no mark on it. With auto-post on, that
- template was published under your name as if your agent had written it. The
- ceiling is 180 seconds now - measured: at 45 seconds ten live models were
- being called dead, and at 180 seconds seven of them answered - the composer no
- longer gives up before the server does, a fallback draft says that it is one,
- and auto-post holds for it. Cloud versus local was also decided by a list of
- five provider names, so every provider added since counted as running on your
- own computer; the endpoint decides now.
-- **The in-app guide is the guide the assistant reads.** The served copy of the
- guide, which is what fetch_skales_docs answers from, was five releases behind
- the source one while both showed the current version in their header. They are
- copied on every version stamp now, and the guide covers 12.5.1 through 12.5.6.
-- **Muting notifications is now actually silent.** The switch promises no pings
- on any channel while everything still lands on the Notifications page, and
- those two halves were the leak. A notification is written to that page first
- and only then silenced, and the toast that pops up in the corner was being
- built from the page's own records - so muting stopped the browser
- notification and the messengers, and a toast and a sound came through anyway.
- In practice that meant a Briefing nudge and a Discover mention still
- interrupted you. The toast is a channel like the others now and falls silent
- with them. One exception stays, the same one the other channels make: a
- request that is waiting on your answer still reaches you, because silencing
- that would leave work stalled with nothing to show for it. What was silenced
- is not saved up either - unmuting does not empty the quiet period onto your
- screen; it is all on the Notifications page, where it always was.
-- **A mention in the Discover feed stopped outranking your own settings.** It
- was filed at the urgency reserved for things that block work until you
- answer, which let it through Mute, through quiet hours, and past the limit
- you can set on how often Discover may ping - while still using up that
- limit's daily slot, so a mention could silence the normal feed ping. It is
- filed as ordinary now and obeys all three.
-
-- **A question with choices no longer stops at the choices.** When Skales asked
- you something and offered buttons to answer it, the buttons were all there
- was. A question like "What is your name and what do you do?" came with four
- roles to pick from and no way to type a name, so the one answer being asked
- for could not be given - the card sat there, nothing moved, and the run looked
- like it had died. Every set of choices now ends with one you can write into
- yourself. The rest of the card caught up too: the short label a question
- carries is shown, and the line of explanation that comes with each choice -
- the part that tells you what picking it actually means - is shown under it
- instead of being dropped on the way to the screen. Questions asked before this
- release still open and still work.
-- **Browser Control says whether it can browse, not whether it once could.**
- The setup screen read a note written the last time an install had gone well,
- and nothing ever revised it. So the panel could sit there marked ready while
- the same panel, a few lines further down, refused to open a window and said
- the browser was missing - and it was: the app had moved on to a newer browser
- build that had never been fetched. The screen now asks the question it was
- claiming to answer, every time you open it: would opening a page work right
- now, with the browser you have chosen. If it would not, it says which part is
- missing and offers the button that fixes it, instead of hiding that button
- because it believed everything was fine.
-- **A failure to launch is no longer reported as a failure to install.** Any
- problem that so much as mentioned the browser engine was rewritten into "not
- installed", including the engine's own complaint that it could not find the
- browser it wanted. The advice attached to it sent you to the screen you were
- already looking at, to press a button that screen was hiding. Each of those is
- now named for what it is, and the advice names something you can do - fetch
- the browser, or switch to the one already on your machine.
-- **Install Chromium can install Chromium.** The button had two ways to do
- nothing. With "Use full Chromium" turned on it took a shortcut meant for
- people using their own Chrome, reported success and downloaded nothing. And in
- the packaged app the downloader itself had not been shipped: the build keeps
- the parts of a library it can see being used, and the piece that fetches a
- browser is started in a way that keeps it out of sight, so the one build that
- needs the button most was the one where it failed. Both are fixed, and the
- button now confirms the result against a real launch rather than against its
- own exit code.
-- **A privacy note stopped introducing itself twice.** The desktop screenshot
- notice began "Privacy: Privacy:", because the screen put a heading in front of
- a sentence that already carried its own - in whichever language you read it
- in. The heading the translation uses is now the only one shown.
-
-- **The Nvidia model list is the models that answer.** Every entry was called
- once and checked: of the 118 the catalogue lists, 44 reply to an actual
- message. Three of the six that shipped had been retired - including the one
- labelled "(Best)", which is the first thing a new user reaches for, so their
- first message failed and read like the provider was down. The list is now the
- verified ones, fastest first, and two things it taught are carried with it: a
- model that appears in a catalogue is not the same as a model you can call, and
- a slow first token is not a dead model.
-- **A retired model is reported as a model problem, not a setup problem.** Nvidia
- answers a request for a withdrawn model with the words "404 page not found" -
- no code, no JSON - and that was surfaced as though the connection or the key
- were wrong, sending you to check settings that were fine.
-- **The curated model picks are the ones you actually see.** The list was filed
- under one name for the provider and looked up under another, so the app
- decided there was no curation and showed the first few hardcoded entries
- instead. Refreshing against 118 live models changed nothing on screen.
-- **An empty "Allowed tools" field means every tool, not none.** Leaving the box
- blank on a custom agent saved an empty list, and an empty list is a complete
- answer: this agent may use nothing. So the one thing you do when you don't
- know the tool names gave the agent no tools at all, and it failed by politely
- explaining it had no capabilities. Blank now means what a person means by it.
- The field also stopped requiring you to know the answer in advance: it filters
- all 184 tools by name and description, shows what you picked as chips, and
- says so when a name matches nothing.
-- **An isolated agent cannot reach the two tools that were a shell in disguise.**
- Running commands is denied to an isolated agent on purpose. Two other tools
- each took an optional command and passed it straight through without a list of
- what is allowed, which made either of them a full shell by another name. They
- are denied now and gone from that agent's catalogue. No isolated agent could
- reach them in practice, because both need a bound code folder - an accident of
- plumbing rather than a decision, and it would have stopped being true the
- moment one got a folder.
-- **Video rendering works on Linux.** Every Linux build shipped a copy of the
- video encoder built for macOS, and it was the first one the app found, so
- rendering a Studio video or a Flow motion died at the moment it started the
- encoder. The app now refuses a binary that cannot run on the machine it is
- on - on any platform, not just macOS - and looks in the place the correct
- Linux build actually ships, falling back to one installed on the system and,
- failing that, saying plainly that it found none.
-
-## v12.5.5 - Nothing You Made Is Gone
-
-A turn used to be a one-way door. Regenerating an answer deleted the one you
-had; a Flow turn rewrote the deck you liked and the previous version was gone
-from the disk. Both keep their history now, and both let you walk back into it.
-Around that: Stop actually stops, an uncapped task runs uncapped, an isolated
-agent stops writing into your memory, and GLM and Qwen become providers you can
-simply pick.
-
-### Added
-
-- **A crash screen now tells you what crashed.** When a page stopped drawing,
- all you got was a line like "Minified React error #482" and a link to look it
- up. The detail that actually identifies the fault was being collected and
- written straight to the browser console, which is no help at all on the phone
- where these crashes tend to happen. Both crash screens now show that detail on
- the page: a ready-made report with the error, the page, and where in the app
- it broke, in a box you can select, plus a copy button. If copying is blocked -
- which it is on a plain connection to your desktop - the button says to select
- the text instead of pretending it worked.
-- **Flow projects can be put away or thrown out.** The project grid only ever
- grew: a run that went the wrong way, a project started by accident and a
- throwaway test all stayed in it for good, because nothing in Flow could remove
- one. Each card now has two actions. Archiving is reversible and changes
- nothing else - the project keeps its files, its versions and its chat, and
- only leaves the grid, with a switch in the toolbar to look through what you
- put away. Deleting is the one that removes work, so it does not simply erase:
- the project folder moves to a trash folder next to your projects and can be
- recovered by hand. What deleting really ends is the conversation, and the
- confirmation says so instead of asking whether you are sure.
-- **Motion can be made in a technique, not just in a palette.** Asking for
- papercut, pixel art or particles used to get you a normal animation with those
- words in the copy: the composition could only move boxes and text around, so
- that is what came back. Motion now has a Style setting of its own next to the
- model and the template, with fourteen techniques - papercut, pixel, particles,
- whiteboard, kinetic type, motion graphics, sketch, ASCII, isometric,
- blueprint, neon, comic, origami and silhouette - and three effects that stack
- on any of them: stop motion, glitch and film grain. Each technique brings the
- drawing it needs rather than imitating it, so pixel art is really drawn on a
- pixel grid and particles really are a particle system. It is a separate choice
- from a style pack, which decides the colours and the type: pick both and you
- get your palette, made that way.
-- **A rendered video is checked before you are told it is ready.** Flow could
- already look over a finished design, but for Motion it looked at the
- composition, never at the video that came out of it. So the faults that only
- exist in the file went out unseen: a clip that freezes on one frame halfway
- through, a stretch that comes out black, a video that never moves at all
- because the animation runs on the clock instead of on the frame being drawn.
- Every finished render is now read back frame by frame, and anything found goes
- straight back into the project as a correction, with the exact seconds it
- happens at. The checks that need no vision model run on every machine; where a
- vision model is set up, the frames that actually differ from one another are
- looked at as well, so a broken transition or clipped text is caught in the
- file rather than in the preview.
-- **A Flow project keeps every version of its artifact.** Each turn's result is
- archived before the next one can overwrite it, up to the last twelve, and the
- history opens from the preview toolbar as real renders of each version rather
- than a list of dates. From a version you can go back to it, download it as a
- ZIP without disturbing what you have now, or branch it into a project of its
- own and try the other direction. Coming back is not a one-way door either: the
- state a rollback replaces is kept as a version too, and the chat is told what
- happened so the next turn works from the files that are actually there.
-- **Flow can just do what you said.** Every mode hands the run a starting file,
- a catalogue of layouts and a checklist. That is what makes a one-line brief
- come out finished, and it is in the way when you already know exactly what you
- want. A "Free" switch next to the brief turns all of it off: your brief is the
- whole specification, and nothing else is added. What you chose yourself stays
- on - your files, your template, your Brand Kit, your style pack. Motion keeps
- its starting file, because that one is the format the video is rendered from,
- not a suggestion.
-- **A video can be as long as you asked for.** The length was a row of fixed
- buttons, one of them always picked, and that pick beat anything the brief
- said - so asking for 45 seconds quietly got you 5. "Follow instructions" is
- now one of the choices, and with it the length comes from your brief alone.
-- **A Flow project can be handed to someone else.** Every export so far turns
- the artifact into something else and loses whatever that format cannot carry.
- The export menu now also offers the project itself: every file as it stands,
- in one ZIP, in any mode. It downloads rather than landing in the project, and
- it carries your work and nothing else - not the version history, not the
- backups, not the deployment settings, and none of the app's own bookkeeping.
-- **Auto: the brief picks the output type.** A pitch deck for a coffee brand
- already says deck, and already says which structure, so translating that into
- a chip and a template before you have described anything is the same work
- twice. The Auto chip reads the brief and picks both, shows what it chose so
- you can correct it, and starts the run normally from there. If the helper
- model is unavailable it falls back to a keyword guess rather than to a Start
- that does nothing.
-- **Ask first: a few questions before anything is built.** Flow could already
- ask mid-run; this asks while it still costs nothing, as a card of clickable
- choices, and folds your answers into the brief instead of turning them into a
- second turn. Anything you leave on "decide for me" is left out entirely, and a
- brief that is already specific gets no questions at all.
-- **Undo and redo in the Flow editor.** The file editor is a plain text box, so
- the only undo it had was the browser's, which breaks the moment the text is
- set by anything but a keystroke. It has its own now, reachable from the
- toolbar and from the usual shortcuts: typing groups into words rather than
- characters, a line break always ends a step, and the cursor comes back to
- where the edit was instead of jumping to the end of the file. Taking a version
- the app put there over yours is undoable too.
-- **Asking again keeps the answer you already had.** Regenerate an answer, or
- reword the question, and the previous turn is kept as a version instead of
- being deleted. Arrows and a counter at the answer step through them; switching
- restores the whole turn it belonged to, including its tool results, and
- survives a reload.
-- **A deck leaves Flow as PowerPoint, a document as Word, a course as SCORM.**
- All three had exactly one way out until now: PDF. PowerPoint is built from the
- laid-out slides, so the text boxes are real ones the client can retype, with
- the pictures, the speaker notes and the slide colours; an artifact with nothing
- extractable in it exports as exact slide images instead, and says so. Word
- keeps headings, lists, tables and pictures editable. A course packages the
- project for Moodle and other LMS, with a runtime that reports the lesson
- complete inside the LMS and does nothing at all outside it, so the same file
- still previews here.
-- **GLM and Qwen are providers you can pick.** DeepSeek was the only
- China-frontier provider you could just choose; these two were reachable through
- nothing but OpenRouter or a custom endpoint you had to wire yourself. Both are
- registered the whole way through: the provider card and its model list, the
- context and output limits, the live model refresh, the group-chat and
- custom-skill pickers, onboarding, and what Skales says about itself.
- Descriptions in all twelve languages.
-- **Gemini 3.6 Flash and Flash Lite** are in the model lists, and are the
- default.
-- **GitHub as an MCP server without Node.** It can now be added over its official
- remote endpoint with a sign-in, so it needs no npx and no access token at all.
- The local server stays, and both entries say which is which.
-- **You hear about it when an agent schedules itself.** An isolated agent can
- give itself standing work that from then on wakes on its own, unattended. The
- only way to find out was to open the Schedule page. It is a notification type
- of its own now, so it reaches WhatsApp, Telegram and the phone with your quiet
- hours and channel choices applied, and it can be switched off like any other.
- You scheduling something in your own chat is not news to you, so that stays
- silent.
-- **An isolated agent gets a memory of its own.** Each agent now has its own
- store, and its lessons go to its own file. Normal chat, custom agents and
- scheduled runs keep the owner's memory unchanged.
-- **Look at an image by URL.** Image analysis takes an http(s) address now,
- instead of the download-then-look detour that put every picture the agent only
- wanted to see on the disk.
-- **The no-time-limit setting is reachable.** A checkbox next to the slider, and
- the message at the end says what actually stopped the run.
-
-### Added
-- **Right-click a chat message for what you can do with it.** Chat bubbles had
- a small row of buttons that only showed on hover; a right-click did nothing.
- Now it opens a menu: Copy, Quote in reply (drops the text into the composer as
- a quote, ready to answer), Save to document, Read aloud, New session from
- here, Delete - plus Regenerate on an answer and Edit and resend on your own
- message. Deleting asks once before it happens, and right-clicking a passage
- you have selected leaves your normal copy alone. The menu it uses is one
- shared component, so the same behaviour is coming to the other lists.
-- **Right-click a chat in your history for everything you can do with it.** The
- same menu now opens on a saved chat, in the History page and in the session
- list in the chat header. It offers Open, Pin, Rename, Archive, Download and
- Delete. The header list used to offer only a delete button on hover; now it
- has the full set, renames a chat in place, and deleting takes two clicks
- instead of a dialog. In all twelve languages.
-- **Right-click a task or a schedule for its actions.** A task card now opens a
- menu with what fits its state - Run while it is waiting, Stop while it runs,
- Copy the title, Copy the result once it is done, and Delete. A schedule row
- opens Run Now, View Logs, Copy, Edit, Pause or Activate, and Delete. Both use
- the same menu as the rest of the app, delete takes two clicks instead of a
- dialog, and a system entry keeps only the actions it is allowed. In all twelve
- languages.
-- **The image menu now works when you open Skales in a browser too.** Right-
- clicking a generated image, a Flow image or an image in a preview used to give
- the app's menu only in the desktop window; from the phone companion or a
- browser tab you got the browser's plain menu with no "Open in Studio". Now the
- same Save image, Copy image and Open in Studio menu appears in a browser as
- well, including for images inside a live preview. In all twelve languages.
-- **Right-click an image or a link and get the menu you expect.** Images and
- links had no menu of their own anywhere in the app - not in chat, not in Flow,
- not in a preview, not in the built-in browser. Right-clicking an image now
- offers Save Image As, Copy Image and Open in Studio (which drops it straight
- into the image generator as a reference); right-clicking a link offers Copy
- Link and Open in Browser. Saving keeps the original file when it can, so a
- JPEG stays a JPEG, and falls back to a copy of what is on screen when it
- cannot, telling you by the name it suggests. It works the same in every window
- and in the in-app browser, in all twelve languages.
-- **Right-click a link in a chat answer for more than copy.** A link in a
- reply now opens a menu with Copy link, Open in browser, Open in the built-in
- Browser, and Summarize - which starts a fresh chat that fetches the page and
- sums it up for you. The full menu shows in the desktop app and just as well
- when you open Skales from your phone or the web, in all twelve languages.
-- **Right-click an API key field for what you want to do with it.** A key or
- secret in Settings stays a row of dots, and the only way to remove one was a
- small icon that appeared inside the field. Right-clicking a key field now
- offers Copy value - the real key goes to the clipboard while the field stays
- masked - Reveal or Hide, and Remove, which asks once before it clears. It
- works on every key in Settings, from the provider keys to the image, video and
- search service keys, in the desktop app and over the web, in all twelve
- languages.
-
-### Fixed
-- **A stuck reasoning model gets the bigger reply budget when a bigger one
- exists.** When a model spends its whole reply budget thinking and produces
- nothing, Skales retries it with more room. That retry was only offered when the
- turn had been running at the model's own limit, so a turn that had been capped
- lower - a background run, a Playground call with a set limit - was denied the
- extra room and re-ran on the same starving budget, then gave up. The retry now
- measures against what the turn actually used, so it gets the larger budget
- whenever there is one to give.
-- **Things that are supposed to exist once now exist once.** Inside the app, a
- page you open and an action you trigger were each running their own copy of
- the same machinery, and the parts that only work as a single copy quietly
- came in pairs. The effects were spread across the app and all had the same
- root: the autopilot's hourly call limit could be counted twice, so it allowed
- roughly double the number of calls you set, and raising the limit in Settings
- only reached one of the two counters; switching autopilot off could leave a
- second heartbeat still ticking; a connected tool server could be started
- twice and stopping it left one running; the browser Skales controls could be
- reported as closed while it was open; the desktop could announce itself twice
- on the local network; a log file could be rotated by both copies at once and
- lose the lines the other had just written; a calendar settings change only
- took effect on one side; and two files being saved in the same instant could
- share one temporary name, which is a way to lose what was just written. All of
- it now lives in one place per app, as it was always meant to.
-- **A project's deploy password stopped travelling with the project.** The
- settings a project keeps for publishing to your own server - host, user and
- password - are stored in a file inside the project folder. That file was left
- out of the project download, and nowhere else: it was listed among your
- project files where it could be opened and edited, it came back over the
- preview address to anyone who could reach your desktop, every saved version
- kept a copy of it, and it went out again inside a course package and a version
- download you hand to someone else. Publishing even uploaded it to the very
- server it describes. Now one list decides everywhere, and the files Skales
- keeps for itself - the project record, the deploy settings, the saved versions
- and the caches - stay out of every list, every download and every page you can
- open. Your work is untouched, and so are the versions you already have: they
- simply stop carrying the password with them.
-- **Turning the effort dial down actually turns it down.** On a model without
- its own reasoning setting, picking Low or Medium for a chat still sent the
- full deep-reasoning instructions whenever the global toggle was on, so the
- answer was as slow and as expensive as before. The dial is the newer, more
- specific instruction and now decides: High and Extra high add the deep pass,
- Low and Medium leave it out. With the dial untouched, nothing changes.
-- **A third of every request stopped being sent twice.** The instructions that
- frame a turn (the Flow brief, Code mode, a goal's plan, the live working
- state) were being placed at the top of the request AND appended at the
- bottom, so the model read the same pages twice and you paid for both. On a
- measured Flow deck turn the request fell from 86.3 KB to 56.6 KB, a third
- less, on that turn and every one after it. Nothing was removed: the second
- copy was the only thing dropped.
-- **Your own instructions to Skales apply again.** The duplicate had taken the
- place where the personality and the system prompt you write in Settings
- belong, so on any turn that carried a frame, which is every Flow turn, every
- Code turn, and every step of a chat after the first, Skales ran without them.
- They are back where they belong, in front of everything else.
-- **A picture sent over WhatsApp is answered, even when the picture does not
- arrive.** WhatsApp hands a photo over in two steps, and the reply was being
- written before the second one finished, so on a slower computer the file was
- never there. Sent on its own, the message vanished without a word; sent with a
- caption, only the caption came through and nothing said a picture had been
- missed. The file is now waited for and asked for again a few times, and if it
- truly does not arrive you get an answer saying so instead of silence, with the
- reason in the WhatsApp log. The same goes for anything this channel does not
- read yet, like a video: it is named rather than quietly ignored. Telegram had
- the same quiet corner and now says it too.
-- **A note about an image no longer disappears behind its caption.** When the
- picture arrived but could not be read (no vision provider, a provider error,
- or image reading switched off for WhatsApp), the explanation replaced your
- message instead of joining it, so a photo with a caption looked like an
- ordinary text message and the reply never mentioned the picture. Both are kept
- now.
-- **WhatsApp keeps working after an app update.** The record of which WhatsApp
- Web build your session runs against was stored next to the app itself, where
- macOS does not allow writing at all and where a Windows update deletes it. It
- now lives with your data, and an existing one is carried over, so a connection
- that works today keeps working after the next update.
-- **A slide is centred on any screen, not only on a 1920x1080 one.** A deck is
- laid out at a fixed size and scaled down to whatever window it is shown in,
- but the scaled slide was being pushed towards the bottom right instead of
- sitting in the middle, so part of every slide was cut off in the preview and
- in any window smaller than full HD. It is anchored to the centre now, at any
- size. Exports were never affected and are unchanged.
-- **The media settings you picked are actually used.** Aspect ratio, quality and
- a preferred model reached the model only as a sentence asking it to honour
- them, so a project set to 9:16 kept coming back square with nothing saying the
- setting had been dropped. They are now written into the generation call
- itself, and they win over what the model typed. Only values the tool can
- actually take are written: the rest is named in the result, so the reply says
- the picture is 4K only when it is. A model you name mid-project still wins over
- the preference you set on the landing.
-- **A message can no longer be lost when two parts of the app write the same
- chat at once.** The lock that serializes chat writes lived per copy of a
- module, and the app loads that module more than once, so two writers could
- each hold their own lock, write the same temporary file, and leave the loser
- with nothing to rename and its messages gone. Seen live: a tool result
- vanished from a Flow turn, leaving the model to answer about work it could no
- longer see. The lock is now shared, and each write uses a temporary file of its
- own, so even a writer the lock cannot reach costs at worst an overwrite instead
- of a lost message.
-- **A turn no longer loses the file you have open in the editor.** Nothing
- re-read that file after a turn wrote it, so the editor kept showing the
- version from before and the next Save quietly put it back over the new work.
- The editor now follows the file: with nothing unsaved it simply shows the new
- content, and with unsaved changes it says the file was rewritten and lets you
- keep yours or take the new one, instead of one of the two disappearing.
-- **Leaving a file in the editor and coming back keeps your unsaved changes.**
- Switching to the preview and opening the file again re-read it from disk and
- threw the edits away without a word.
-- **The version history is reachable while you are editing a file.** The editor
- takes the place of the preview, and with it the only way into the history.
-- **Present a deck, not just show it.** A deck opens in a Presenter view of its
- own now: the current slide, the one coming next, the speaker notes for what
- you are on, an elapsed timer and the clock. The audience gets its own window
- with nothing but the deck, and the two move together, from the buttons or the
- arrow keys. Decks are also written with speaker notes from now on - what you
- say over a slide, not a repeat of what is on it - and those same notes travel
- into PowerPoint and into a new PDF handout that puts each slide next to its
- notes. A deck that has no notes says so instead of handing you empty pages.
-- **Build a Brand Kit from a website.** Paste an address into the Brand Kit
- settings or into the New kit dialog in Flow, and the kit fills itself from the
- page: the palette, the heading and body typefaces, the logo, the name, the
- tagline and a short description, with the site kept as a reference link.
- Colors and fonts are read out of the page and its stylesheets, so what lands
- in the kit is what the site actually uses, not a close-looking guess - the
- page says which color is its surface and which is its body text, and the roles
- follow that rather than whichever color happens to appear most. Nothing is
- saved by reading a site: the fields are filled and you still press save, so a
- bad read costs a glance.
-- **Using the logo can be decided mid-project.** Which Brand Kit a Flow project
- designs with switches from the toolbar at any time, but the "use logo and
- brand assets" half of that same choice could only be answered on the landing,
- before the project existed. A project started without the logo could never be
- given it, and one that was meant to stop demanding a logo file had to be
- started over. The toggle now sits under the kit in the toolbar menu, applies
- from the next turn, and stays where you left it when you switch kits or turn
- the kit off and on again.
-- **The wireframe and phone modes are reviewed as themselves.** Both start from
- the same file as the web prototype, and were reviewing themselves against its
- checklist too. So a phone mockup was told to make sure its layout collapses to
- one column on a narrow window, which means nothing for a strip of fixed phone
- screens and invites the run to break the strip, and neither mode was ever
- asked whether it had done the one thing that mode exists for. Each has its own
- review now: the phone mode confirms it built phone screens rather than a
- website, with buttons in thumb reach and the labels the product would ship;
- the wireframe confirms it stayed grey, kept its image boxes crossed out and
- annotated, and used the real navigation. Both also say where a picture comes
- from, which neither did before: the phone mode may find and place real images
- like every other mode, and the wireframe deliberately fetches nothing and
- keeps labelled boxes, instead of each run deciding that for itself.
-
-- **Stop now ends the session.** Stop reported success whether or not it stopped
- anything, so with a stale run the panel said stopped while the run kept
- generating, and reopening found it still live. A stop now answers honestly and
- sweeps the session: the run you are looking at first, then every other live run
- on it. Chat does the same sweep.
-- **A stopped run stops looking like a failed provider.** The fallback chain read
- a stop as a dead provider: it walked the rest of the chain and put the primary
- on cooldown, so pressing Stop quietly moved the next turn to another provider
- and the run ended in "all providers failed". A stopped run now ends the chain
- and says it was stopped.
-- **No time limit now means no time limit.** The task timeout has documented 0 as
- run-to-completion for a while, but the code took the zero literally: the
- deadline fired on the next tick and the task died before its first step. An
- uncapped run is now judged on whether it is still doing something, and ends
- only after fifteen quiet minutes.
-- **A resumed workday remembers what it already tried.** The task loop had no
- repeat guard, so a stuck model reissued the same call with the same arguments
- until the clock ran out, and the checkpoint carried that stuck state into the
- next resume while every counter restarted at zero. The guard now rides along in
- the checkpoint, so "already tried" means across the whole chain. A chain that
- adds nothing new ends itself; one that is actually advancing resumes for as
- long as it advances.
-- **A chain of runs stops saying it is done seven times.** Each round announced
- itself as finished. A run that queued its own continuation now holds its
- notification and its report until the chain really ends.
-- **Continuing a task by hand keeps the agent it belongs to.** Typing "continue
- the task" carried the work on in your chat as the default agent, no matter
- which agent the task was bound to, with that agent's identity and data replaced
- by yours. It is now queued as a continuation under that agent, and the chat
- says so. An unbound task continues in chat exactly as before.
-- **Saving a chat stops blocking on every message.** Each message read the whole
- session file back, stringified it and wrote it, all synchronously, on paths
- that are mostly remote-facing. Creating a session also wrote its own file
- outside the shared writer, so a new chat skipped the auto title, the base64
- stripping and the message cap.
-- **Skales asks its questions again.** It could ask you a question through a form
- card since v11.3, and mostly did not: only one way of writing the call was
- accepted, so a model that phrased it slightly differently produced nothing, and
- a failed tool goes to the model, not to you. It retried until the turn ran out
- and the session ended empty. All those shapes are read now, and a question that
- still cannot be shown says so on screen instead of vanishing.
-- **Test Connection tests what you typed.** It only ever tested the saved
- provider, so pasting a key and pressing Test failed while the long way round
- worked.
-- **Fetch models and the key remove button exist on every provider card.** Both
- were built into a block only five providers render, so Fetch existed on
- OpenRouter and nowhere else. The remove button also moved into the key field as
- an icon, so the card stops jumping when a key is saved.
-- **MCP servers can find the tools you installed.** A double-clicked app is
- started without a login shell, so nothing installed with Homebrew, nvm, fnm or
- into a local bin folder was visible, and every npx-based MCP server died with
- "spawn npx ENOENT" while the same command worked in Terminal. The usual install
- locations are now on the path a server is started with, anything you set
- yourself still wins, and a command that cannot be found is reported by name
- with the two ways out.
-- **A skill loads when you name it, not when a word contains its name.** A skill
- was pulled in on a substring, so one called "teach" matched "can you teach me
- about anything at all" and pushed ten kilobytes of unrelated instructions into
- the prompt. Naming a skill still loads it, including the slash form; a name
- buried inside another word does not.
-- **A stranger cannot stop your turn in a Telegram group.** Everyone writes to
- the same chat, and the stop machinery only knew the chat: someone else typing
- "stop" killed your running turn, and someone else's chatter counted toward the
- follow-up threshold. Both now follow the sender. One-to-one chats and WhatsApp
- were never affected.
-- **Importing a skill brings its files.** The GitHub import asked whether a skill
- had scripts and references, recorded the answer, and then downloaded neither,
- so a skill claimed to ship with its scripts while any instruction pointing at
- its own files failed. Both folders are fetched now, and the flags describe what
- actually landed on disk.
-- **Long-term memory asks where a fact came from.** Memory harvesting mined
- assistant turns without asking about origin, so a summary of a stranger's email
- or a fetched page was a candidate like any other, and a sentence planted inside
- that content could walk into memory through the summary. A turn that pulled in
- content written outside the conversation is no longer mined. Your own files,
- your own machine and your own messages are unaffected. The cost is real and
- stated: a fact you mention while discussing fetched research in that same turn
- is lost with it.
-- **An error from a phone-initiated turn opens the right conversation.** It
- carried no session, so tapping the notification opened the chat landing instead
- of the conversation it belonged to.
-- **A finished answer is not argued with.** Any sentence anywhere that looked
- like an announced next step re-prompted a completed answer up to three times,
- including recaps of work already done and offers handed back to you. The check
- now reads the closing stretch of the answer and no longer mistakes an offer for
- a promise.
-- **The activity log stops losing lines.** Its rotation state was kept per copy of
- a module rather than shared, so two parts of the app could rotate at once and
- the loser's lines were lost.
-- **A scheduled job keeps its shape and its agent.** The cron API wrote its own
- envelope into the job file, so an update persisted the word "update" as a
- field. A job can also be bound to an agent when it is created, rather than
- needing a second call, and a run started over the API is filed as the manual
- run it is.
-- **Good files stop being called broken.** The download validator was rejecting
- real files: a favicon served as PNG bytes, a one-pixel GIF under the size
- floor, an SVG with a licence comment in front of it, a ZIP with a stub before
- its first entry, and a QuickTime file that opens on a different box. The
- bot-wall case it was built for is still caught.
-- **Reveal in the file manager stays a reveal, not a launch.** The action that
- shows a folder or file in Finder or Explorer handed the path to the system
- opener as-is, and the opener treats some paths as "run me" rather than "show
- me": an app bundle, a script, a desktop launcher would start instead of being
- revealed. A file is now handed to the opener in reveal mode, so it is always
- shown in the file manager and its own handler is never invoked; a folder still
- opens as before. The same endpoint also turned away a request coming from
- another site in your browser - a page you happen to be visiting could reach
- the local app and ask it to open something. It now accepts the request only
- from Skales itself.
-- **Uploading a skill no longer runs it.** To show a skill's name and
- description in the list, the app used to execute the uploaded file, with full
- access to the machine, the moment it arrived - before you had looked at it,
- and whether or not you ever pressed anything. A file that did its damage and
- then failed was simply filed as broken, so there was nothing to notice. Those
- few lines of information are now read out of the file as text, without running
- any of it, and every skill in an uploaded archive is treated the same way. A
- skill runs when you run it, and not before. Uploading also stopped accepting
- requests from other sites in your browser.
-
-## v12.5.4 - Pictures and Providers
-
-Sending a picture is a normal chat turn now. It was not: it ran on a path of its own, outside everything the app does to keep an answer alive, and the previous release patched that path instead of removing it. The path is gone, and the same class of loss is fixed everywhere else it hid - in Buddy, in the AI Pointer overlay, and in the model recognition that quietly shrank a new model's context window to an eighth of its real size. Moving that turn onto the normal path also broke a promise nobody could see from the outside: an incognito chat with a picture in it was being written to disk. An incognito chat now runs on a conversation that has no file at all. Moonshot AI (Kimi) joins as a provider.
-
-### Fixed
-
-- **Incognito keeps its promise on picture turns.** Since image turns became normal chat turns, the answer to a picture was written to the chat file by the part of the app that generates it - which never knew about incognito. An incognito chat with a picture in it was saved, and saved half-blind: your own message suppressed, the reply kept. An incognito chat now runs on a conversation that exists only in memory: nothing about it is written anywhere, not the messages, not the live reply, not an entry in History, and it is gone when you leave incognito or close the app.
-- **The app's own nudges stop being mistaken for your instructions.** When a turn needs a correction, the app writes one into the conversation the same way you would, because that is the only way it reaches the model. Everything that then looked back for "what did the user ask?" found the correction instead: a parked task could adopt a nudge as its objective, and it could be written into memory as something you had said. A checklist created during such a turn also went invisible, so a task that had a plan stopped continuing on its own. The corrections still reach the model, unchanged; they are no longer read back as yours.
-- **Stop reaches an image turn during the handoff too.** Stopping while the picture was being prepared cleared the screen, and the turn then started anyway a moment later: the answer arrived after you had stopped it. It now ends the run it just started.
-- **Buddy still sees the screenshot after you approve an action.** Approving a tool continues the same turn in a second request, and that continuation went back to the stored conversation - where the picture is a placeholder by design. The stored picture is read back for the continuation, so the rest of the turn is answered about the screenshot too.
-- **A picture no longer pushes the conversation out of the request.** An attached screenshot was sent at full resolution - several megabytes - and the app measured that literally when deciding how much history still fit, so one picture could take the entire budget and the model answered it with no conversation behind it at all. Pictures are now sent at the size every provider scales them to anyway (the stored copy stays full resolution), and they are counted for what they cost instead of by their raw size. Long chats with images also got noticeably faster.
-- **Moonshot's region switch works everywhere it should.** Refreshing the model list ignored the picked region and asked the international host with a China key (two separate accounts - the key is invalid there), so Refresh reported no models on a correctly set up provider. The selected region also did not look selected, because it was painted with colours that only exist in the AI Pointer window. Both fixed, and a stored address in a slightly different but equivalent form is now recognised as its region.
-- **Buddy sees the screenshot you attach.** Attaching a picture in Buddy showed it in Buddy's own bubble and then answered past it, because the image never reached the model. Session files deliberately do not store raw image data, and Buddy read its history back from that stored copy - so what arrived at the model was the placeholder the storage layer had left behind. Opening the chat afterwards showed that same placeholder as a broken image. The picture is now stored as a file the chat can render, and the live one is handed to Buddy directly, routed to a model that can actually see.
-- **Pictures from Buddy and the AI Pointer show up in the chat.** Those turns store the picture as a file and no preview copy, and the chat bubble only ever looked for the preview - so opening the chat afterwards showed the message with no image at all. The stored picture is now shown directly.
-- **A screenshot sent from the AI Pointer overlay to the chat keeps its picture.** The overlay sent the image along with the question, and the chat import ignored it, so the turn arrived as text with the picture missing.
-- **New model generations work the day they land.** Model recognition was written as per-version lists, so each one broke the moment a vendor shipped the next number. A Gemini generation newer than the app ran through OpenRouter at an eighth of its context window and an eighth of its output limit, which made long chats compact themselves far too early and cut long answers short. Kimi's vision-capable generations were a literal two-entry list that the next release would not have matched. Both are rules now: a new Gemini or Kimi generation is sized and detected correctly without waiting for a Skales update.
-- **Extra custom endpoints get their real context window.** Every endpoint added under Additional Custom Providers fell back to the smallest possible window rather than the one custom endpoints are meant to get, so a 128K model was run at half of it and conversations started compacting after a handful of turns. The single built-in custom slot was never affected.
-- **The switches on a custom endpoint do something.** On endpoints added under Additional Custom Providers, turning tool calling off did nothing - the tools were sent anyway, and an endpoint that cannot call them answered with garbage. The Vision checkbox was worse: nothing in the app ever read it, on any custom endpoint, so images were stripped out before the request left regardless of what you ticked. Both are honoured now, and both still default to what they did before.
-- **An image turn no longer reports itself dead while it is working.** The composer watches for a locked input with no live turn behind it and frees it after 25 seconds. An image turn never registered as a live turn, so on every picture you sent - since the release that added that guard - the app unlocked the composer and said the run had stopped responding, while your Vision Provider was in fact still reading. Reading one image with a local model takes longer than that on its own, several images take minutes. Image turns are real turns now, and the guard leaves them alone.
-- **You can watch an image answer arrive.** The reply now streams token by token like any other, instead of showing a spinner until the whole answer is finished. On a long answer from a local model that is the difference between a working turn and a turn that looks frozen.
-- **Stop stops an image turn.** The button could not reach the call that was running; it cleared the screen while the work continued in the background and the answer landed afterwards anyway. It now ends the actual turn.
-- **Your custom agent answers your images.** With an agent selected, an image turn used its model but not its instructions, so the picture was answered by the default assistant wearing the agent's name. The agent's own prompt and its assigned skills now apply to an image turn. A picture handed straight to the chat model is still answered without tools, so that the model reads the image instead of reaching for something else; with a Vision Provider set up, the turn keeps the full toolset.
-- **A reply cut off mid-sentence finishes itself.** When a model hit its output limit, the answer simply stopped where it stopped. The rest is now requested and joined onto what came before, so one complete message lands. This applies to every chat turn, and to the ones nobody is watching: a task, a Flow run or a scheduled run used to accept a half sentence as its finished result and close on it.
-- **A picture that could not be answered says so, permanently.** If starting the answer failed outright, the error showed once and was gone: after a reload the picture sat there with no reply at all, and the model later saw a conversation ending on an unanswered question. The failure is now part of the conversation. The helpful guides for the two most common causes - the Ollama vision model is not installed, or the Vision Provider has no key - are also back, in all twelve languages; they had been lost in the rebuild and every cause showed the same generic text.
-- **Telegram and WhatsApp answers carry the same protection as the desktop chat.** The way those two channels handed their instructions to the model switched off the assembly step that, among other things, tells the model to treat fetched web pages, files and messages as data rather than as instructions. Both channels now go through the same assembly as everything else, and material coming back from tools is marked as untrusted content there too.
-- **A reply cut off mid-sentence finishes itself on every channel.** The repair that stitches a truncated answer back together existed only in the desktop chat. On Telegram and WhatsApp a long reply simply ended mid-sentence; from the phone it was worse - the beginning of the answer was dropped and the message started mid-sentence. Buddy never checked at all. All four now request the missing remainder and deliver one complete message.
-- **A model that thinks itself empty gets rescued on every channel.** When reasoning consumes the entire reply budget, the desktop chat raises the budget, then turns reasoning off, then says honestly what happened. Telegram, WhatsApp, Buddy and the phone either retried the identical doomed request or invented a cause: Buddy said a bare "No response.", the phone blamed your provider settings. All of them now climb the same ladder and, if it truly ends empty, tell you what actually happened - in your language.
-- **A tool call torn by the reply limit no longer runs broken on Telegram and WhatsApp.** A file write cut off mid-content executed anyway with half its content or failed on unreadable arguments. Like the desktop chat, both channels now hold the torn call back and ask the model to redo it in smaller pieces.
-- **A picture sent from the phone reaches a model that can see.** An image turn from the phone ran on whatever model was active; if that model cannot process images they were quietly removed before the request went out, and the answer ignored the picture. The turn is now routed the same way a desktop image turn is: to a model that can actually look at it.
-- **The recovery steps a stuck turn is supposed to take now actually happen.** When a model spends its whole budget on internal reasoning, repeats a failing tool call, or announces a step and then does nothing, the app sends it a correction and tries again. Those corrections were being dropped before they left the app, so the retry re-sent the identical request and the model repeated itself. They reach the model now. Instructions from your own hooks were dropped the same way and now arrive too.
-- **An incognito chat stays incognito, and stops being lost.** Leaving the chat page and coming back reopened the incognito conversation as an ordinary chat: no purple marking, no "nothing is saved" notice, and because it kept running on a conversation that has no file, everything written in it from then on was saved nowhere and gone at the next start. Coming back to an off-record chat now finds it marked as one.
-- **The picture in an incognito chat is not written to disk either.** Attaching an image saved the full-resolution original into the workspace uploads folder before the chat even existed, and nothing removed it afterwards, so it stayed readable by every file tool long after the conversation was gone. Off the record now covers the picture too.
-- **The Buddy carries the same protection as everything else.** The way the Buddy handed its instructions to the model switched off the step that tells the model to treat fetched pages, files and messages as data rather than as instructions, and its tool results were not marked as untrusted. It is the assistant with the widest reach and the one that runs a step without asking, so it needed that protection most. It now goes through the same assembly as the desktop chat and the two messaging channels.
-- **A cut-off answer no longer disappears when finishing it fails.** A reply that hit the output limit is continued and joined into one message, but if that second request failed or you pressed Stop while it was running, the half you had already watched arrive was thrown away and only an error remained. What was produced is kept now, whichever way the turn ends.
-- **A picture is no longer sent to a model that cannot see it.** On providers where the app cannot swap in a seeing model by itself, an image turn from your phone or from the Buddy went out to whatever model was active, and a text-only model answered with the provider's raw rejection. The turn now says plainly that the model cannot read images, and how to set up one that can.
-- **Kimi's own default model reads pictures again.** Recognition was written around version numbers, and the rolling model name Skales itself ships as the Moonshot default carries no version number, so the app treated its own default as unable to see: through OpenRouter it quietly moved the picture to a different vendor's model, and on Moonshot it reported the model as blind. That name and the 2.5 generation are recognised now.
-- **Gemini's image models get their real context window.** They were sized by the family rule at a million tokens, thirty times their actual window, so the conversation was never compacted, the context meter read far too low, and the request failed once the chat passed the real limit. They now carry their own sizes, and a picture-model released later is sized cautiously instead of far too large.
-- **"Test connection" honours the Moonshot region.** The button always tested against the international host, so a correctly set up China account reported an invalid key, and a passing test saved a configuration still pointing at the other service. The region you picked is now used for the test and stored with it.
-
-### Added
-
-- **Moonshot AI (Kimi) is a provider you can just pick.** Paste a key, choose International or China, and it works the way every other provider works: in chat, in Lio AI, in Codework, in Flow, in group chats, in custom skills, in the onboarding, and in the model picker everywhere models are chosen. The two Moonshot services are separate accounts with separate keys, so the region is a switch rather than something to type. Kimi's context window, its tool-call handling and its tuning profile were already in the app; what was missing was the provider itself.
-
-## v12.5.3 - Flow Hotfix
-
-The 12.5.2 reasoning watchdog cut healthy long-thinking runs: models that built a motion graphic fine one day earlier spent half an hour in cut-and-retry ladders the next. The guard is now behavioural, the stall card actually does what it offers, custom widgets are usable again in every theme, and custom agents see your connected MCP servers again.
-
-### Fixed
-
-- **A custom agent sees your MCP servers again.** The chat agents you build yourself run on their own prompt, which left out everything about the environment they work in: the connected MCP servers were never mentioned to them, and the tools were waiting behind a load step the agent was never told about. Asked about a connected server, such an agent answered that none were set up. Custom agents now get the same truthful MCP status as the default assistant - which servers are connected, and the exact step to pull their tools in - plus the index of every other tool group they can load.
-- **An image with text gets a real answer.** Sending a picture along with a question could end in a silently empty reply: a reasoning model that spent the whole turn thinking produced no visible text, and the image turn had none of the safeguards normal chat turns have, so nothing appeared and no error was shown. In older versions the same gap showed up as a reply cut off mid-sentence. An image turn now retries once asking for the answer directly, stitches a cut-off reply back together with one continuation, and when a model truly never answers it says so honestly instead of staying silent.
-- **New vision models are recognized without waiting for an update.** Skales asks your Ollama daemon what each model can actually do instead of matching names against a built-in list, so a model that just came out is treated as vision-capable the moment it appears. Kimi K2.6 and K2.7 stop being flagged as text-only, and the name list stays only as a fallback for older daemons.
-- **The image-description row fits its content.** The collapsed "Vision" row under an attached image stretched across the full bubble width, which next to a narrow image looked like a broken layout. It now hugs its label and only widens when opened.
-- **Custom widgets scroll again.** A widget taller than the window could not be scrolled at all: it grew past its pane while the pane itself refused to scroll, so everything below the fold was simply unreachable. Clicking or typing also made the page slide down under the pointer, because the widget's own click handling was being mistaken for new content.
-- **Custom widgets follow your theme.** They were painted in fixed dark colours, so on a light theme the text was unreadable. They now inherit the theme's own colours and switch along with it, without losing what you had entered.
-
-- **Long-thinking models can think again.** The previous release stopped a reasoning stream once it crossed a fixed share of the reply budget, which killed exactly the deep runs it was meant to protect, and then retried the doomed step with the very same budget. The stream is now stopped only when the model is provably stuck repeating its own trace, or exceeds the physical context window; whether a budget is exhausted stays the provider's call. The pointless same-budget retry is gone, and the request no longer asks for a reasoning bound the failing models were measured to ignore.
-- **The stall card's buttons work.** "Keep going" was re-raised by the next status poll within a second, and Stop had to win a race against that same poll, which could take a dozen clicks. Dismissing the card now sticks for the rest of the run, and Stop takes effect immediately while the server abort settles in the background.
-
-- **Motion scenes no longer bleed through each other.** Scene transitions layer the incoming scene over the outgoing one, and a transparent scene showed both at once (old text under new text). Scenes are opaque by default now, and Flow renders frames from inside the transitions after every motion turn and checks them - plus a code-level check for exactly this class of conflict - firing a correction automatically when something is off.
-- **Your model choice in Flow is the one that runs.** If a model was set under Settings > Chat & Code > Code model, it silently overrode the model picked in Flow: the project kept showing your choice while the run used another model, and switching models to escape a stalled run changed nothing. Flow's own pick now wins in Flow.
-- **Options that promised something they did not do.** The reasoning-effort dial is hidden on models that have no such control instead of being offered and ignored. Switching mode clears a pinned image or video provider that does not exist in the new mode. A pinned skill that you switch off in settings stops riding along. Pinned-skill and style-pack chips show real names instead of raw slugs.
-- **A short motion brief stays short.** Asking for 3 or 5 seconds used to arrive at the model alongside two other duration ranges from the built-in guidance, and the model spent the turn reconciling them. Your duration now overrides everything else and shapes the scene plan.
-- **Image and video projects see your attachments.** Those modes skipped part of the briefing: an attached reference image was never mentioned to the model, and they had neither the fact-checking rules nor the scratchpad the other modes get.
-- **Flow's option rows scroll instead of wrapping.** The mode chips, the composer options and the project toolbar stay on one line and slide sideways at every window size. Dropdown menus can no longer leave the window: they clamp to the frame, flip upward when there is no room below, and the reasoning-effort row in the model menus is always visible instead of hiding behind the model list's scroll.
-
-### Added
-
-- **Changing the model mid-run resumes the work.** Picking a different model while a run is working closes that run, moves the project to the new model and continues the work in place, without restating the brief. This is what the stall card's "Change model" now does, and the project's model menu behaves the same way.
-- **PDF export for decks and documents.** Both modes have advertised it since Flow shipped without an actual export anywhere in the app. The preview bar now has it: a deck exports one full-bleed landscape page per slide, a document exports through its own print layout.
-- **A scratchpad in every Flow project.** The agent keeps a small working-memory file - plan, decisions, open items - current while it works, and a panel in the project shows it live. It carries over into new chapters, so long projects keep their thread.
-
-## v12.5.2 - Flow and Security
-
-Flow gets real controls for long projects and long-thinking models, a security hole on the local command interface is closed, and the app finally reports its own version correctly.
-
-### Fixed
-
-- **A closed door that was standing open.** The local command interface accepted requests without a valid access token, so anything on the same network could reach it. It now requires the token, the way it always should have.
-- **Reasoning models that thought forever now get stopped.** Some models spend their entire reply budget on internal reasoning and never write an answer. A live watchdog now ends a run that has clearly thought past its whole budget, instead of leaving you staring at a spinner for twenty minutes.
-- **Flow's Stop button actually stops.** After a page reload it used to lose track of the run and only pretend to stop while the work kept going. Stop now reaches the real run every time, and reopening a project reconnects to it.
-- **A crashing preview no longer takes the whole project down.** A render error is caught in place and recovers, instead of blanking the page and orphaning the run.
-- **The app reports the right version.** System properties and the About box were frozen at an old number regardless of the installed release; they now match.
-- **DevKit can be enabled on an installed app.** The developer command interface only worked from a source checkout; it now reads its config from the data directory (`~/.skales-data/devkit/`), so the CLI works on a normal install. Requires the matching DevKit CLI v0.3.0.
-
-### Added
-
-- **New chapter in Flow.** After many rounds of changes a project can continue in a fresh, lighter session that keeps the files and a short summary but drops the long transcript, so it stays fast and costs less. Offered on its own once a project gets heavy; your brand kit, template and model choices come along.
-- **Change the model and reasoning effort mid-project.** Both were only settable when starting a project; now they sit in the project's model menu and apply from the next step.
-- **A way out when a model stalls.** If a model finishes twice with nothing but reasoning and no answer, the panel offers to switch to a different model instead of leaving you stuck.
-- **Option chips scroll on narrow screens.** The mode and parameter rows slide sideways instead of stacking into several lines.
-
-## v12.5.1 - Hotfix
-
-Connected MCP servers are visible to the AI again, background tasks stop running in circles, slow models get the time they need, and the Tasks page becomes something you can actually keep tidy.
-
-### Fixed
-
-- **The AI can see your connected MCP servers again.** Their tools are fetched on demand to keep messages lean, but the AI was being told they were already active, so it concluded it could only see the built-in tools. It now knows how to pull in each server's tools and does so on its own in chat, Flow and scheduled runs.
-- **Slow models no longer lose a whole task partway through.** A single model reply was cut off after a fixed short limit, and because that counts as an error the task failed with no resume - painful for local and reasoning models that legitimately take longer to answer. The reply now has the full task time to finish.
-- **Background and scheduled tasks now see the full results of their file reads and searches.** Until now they received only a bare count or metadata, so a task that needed to look something up could repeat the same search endlessly, burning through its time limit run after run without ever finishing. The same blind spot affected agent teams and the Buddy after a tool approval. All of them now see exactly what chat sees.
-- **Deleting an old reminder on the Tasks page works now.** Rows that came from the Planner were silently ignored by the delete button, so one-time reminders could never be removed from the list.
-- **Flow shows what the model is actually doing.** A reasoning model can think for minutes before the first visible word, and Flow showed only a bare "deciding" spinner the whole time. The status row now shows that the model is reasoning, a live glimpse of its train of thought, and the elapsed time. The Stop button remains available throughout.
-- **An answer no longer comes back empty after a long reasoning phase.** Models with extended reasoning could spend the entire reply budget on thinking and finish with no visible answer. That case is now caught and retried once with a much larger reply budget and a nudge to answer directly.
-- **Flow menus stay inside the frame.** The model and template dropdowns could open partly outside the visible area near a screen edge. On narrow screens they now open as a bottom sheet that is always fully visible.
-
-### Added
-
-- **The Tasks page can be searched, filtered and cleaned.** A search box, status filters with counts (All, Active, Done, Failed, Stopped), newest-first ordering, and a one-click "Clear finished" that removes every completed, failed and stopped task - running and pending ones always stay.
-
-### Changed
+- **The picture reader's key field no longer fills itself with a key you never typed there.** Letting the reader borrow the key you already gave a provider also put that key into the Browser Control card, so the field showed a secret nobody had entered there, and pressing Save laid a second copy of it down on disk and froze it. Rotating the key in AI Providers then left the reader quietly using the old one until every picture came back as a failed request. The card shows what is stored again, an empty field stays empty, and the borrowing happens at the moment a picture is read - so a new key is in effect on the next image, with nothing to re-save.
-- **The task time limit can now be set up to 60 minutes** in Settings, for long-running local models that need room to work.
+- **Clicking a control by name now lands on that control, on a Retina screen too.** A button addressed by its name is found through the accessibility tree, which answers with a point of the desktop - but it was then read a second time against the last screenshot, so on a scaled display the click landed at half the distance from the top left corner, and if another app had come to the front in between it was refused with a sentence about a screenshot that was never involved. A point that came from the tree is now used as it is; a coordinate read off an image is still checked against that image as before.
-## v12.5.0 - Skills
+- **Looking at the screen stays allowed in Plan and Ask.** Four ways of reading the desktop - listing the displays, listing the open windows, reading what is on screen and checking whether something is true there - were refused in the read-only modes, while taking a screenshot went through. They only observe and are now treated that way, so a plan can be worked out by looking before anything is changed.
-Skales learns skills: a built-in library it can reach for on its own, style packs that give Studio a committed aesthetic, ready-made agent teams - and an agent that keeps working instead of stopping halfway.
+- **A skill you installed before there were permissions keeps the doors it always used.** Giving skills a bounded space of their own closed four of them at the same time: a skill that fetches from the internet could not reach it even when its own permissions named the site; the data folder and the workspace were handed over and then refused on every path inside them; the settings it used to read came back empty; and a skills folder reached through a shortcut was not recognised as belonging to you. All four are open again and none of them wider than before - the internet only to the sites the skill's permissions name, the two folders only because they were handed over, and the settings with anything that looks like a secret removed, exactly as a command run from a shell gets them. A request carrying something other than text is refused by name instead of being sent empty, and what a site answers with is readable again.
-### Added
+- **The same skill folder now gets the same package fingerprint here and on the phone.** The fingerprint exists so two copies can be compared without trusting where either came from, but the two platforms listed the files in different orders - and the order is part of the fingerprint, so any folder holding a SKILL.md or a README.md never matched itself across devices. Both sides list in one order now.
-- **A built-in skill library, ready on install.** Skales now 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 by the AI on its own when the task fits; the others run when you name them. Every skill shows its author and license right on the card, the library is searchable with category filters, built-ins cannot be deleted (only switched off), and any skill you import under the same name simply takes over.
-- **Skills in chat, on demand.** Mention a skill with @ or just ask for it by name - the AI loads the full instructions itself before acting.
-- **Style packs for Studio.** 74 aesthetic references - each clearly labelled as inspired by a well-known product or brand look - can be pinned to a Flow project. A pack shapes typography, color and layout mood; your brief always wins on content, and no real logos or trademarks are ever reproduced.
-- **Flow knows its skills.** The Flow landing gets two new controls: pin up to three skills that ride along on every step of the project, and pick a style pack next to the Brand Kit. Your choices show as chips on the first message, so a project always 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; everything stays editable, and building a team from scratch works as before.
-- **Images carry their origin.** A picture that really came from image generation wears a small origin badge - text that merely claims a picture does not. The generation prompt now sits in a collapsible "Show prompt" section under the image instead of flooding the chat.
+- **Building a release no longer leaves the terminal and the global hotkey broken behind it.** Both are pieces that have to be rebuilt for the machine they will run on, and that rebuild was being skipped at the end of a release - so afterwards each failed with an unreadable error that looked like a fault in Skales itself. It happens again now, on every kind of release build, and there is only one place it can be forgotten from.
### 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 now queues its own continuation and picks up from the checkpoint - overnight jobs no longer stall until someone pushes them.
-- **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 - an error page pretending to be an image fails loudly instead of leaving a broken file behind, and re-fetch attempts identify the app honestly.
-- **Websites that block the built-in browser now say so.** Instead of failing quietly, the run explains the block and points to the visible-browser login handoff.
-- **Knowledge-graph import shows real progress** instead of sitting at zero for minutes.
-- **Every provider key can be removed** - including the ones that previously had no remove button.
-- **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.
-- **A cancelled goal says so once.** The cancel notice appears exactly one time and no longer returns on every reload.
-
-### 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.
-
-## v12.4.5 - Checkpoint
-
-The release that locks the doors before adding more rooms: remote access becomes one honest switch, pairing asks you first, videos work from the first message, and your phone can hand the desktop videos, PDFs and its whole photo batch.
+- **The licence question is asked before the update is delivered, not after it.** Automatic downloads are on by default, so a version that comes under a new agreement used to arrive downloaded and installed, and the licence screen came afterwards - the decision landed after the delivery. An update that crosses the licence boundary now holds: the notice names the change, links the text, and the Download button is the decision. An installation that already runs on this agreement updates exactly as before, and somebody who presses Download themselves is never held up. The one group this cannot reach is the one already on the old licence, because the holding happens in the build that does the holding - so the release feed itself now carries the notice, which is the only thing an older Skales shows you that we write at release time.
-### Added
+- **Two lines stand between the licence and the buttons.** Private use stays free of charge, as before; commercial use needs a licence, and that was true under the previous licence too. They sit below the text rather than above it, because they are the last thing read before the press. There is nothing to click that classifies you as private or commercial: a press cannot make that true, and a promise that "nothing changes for you" would be a promise to a group the app cannot identify.
-- **Remote access is one switch, default off.** Off means the server listens only on this machine - in a hotel or cafe network the port simply is not there. On means reachable over LAN/Tailscale AND token-protected, as one inseparable setting; the in-between that used to be the default (reachable, no token) no longer exists. The access token now survives restarts, and Settings > Security shows it at the very top with a copy button, a QR code and a regenerate button. Swarm users are unaffected: an enabled Swarm keeps the port open for its peers.
-- **Pairing asks you first.** A phone (or any device) that knocks with your pairing code no longer connects by itself: the Mobile page shows who is asking - device name plus the tail of its key - and you allow or refuse. Your Allow counts only for the device you were shown, refusing is "not now" rather than a ban, and unanswered requests expire on their own.
-- **Pair without a camera.** The desktop now shows a copyable pairing string under its QR code - for remote onboarding, headless devices, or whenever scanning is impractical.
-- **Video from the very first message.** The chat landing page accepts videos (picker and drag & drop) up to 100 MB, shows a thumbnail chip, and hands the analysis to the same frame-sampling flow the in-session composer uses.
-- **Videos, PDFs and full photo batches from your phone.** The bridge speaks a new protocol: the paired app (2.4.5) streams videos and PDFs to the desktop in chunks, every attached image arrives instead of only the first, and the phone can display the desktop's knowledge graph. Older app versions keep working exactly as before.
-- **Attach audio, and videos get their soundtrack read.** Drop an audio file into either composer and it is transcribed through your configured speech-to-text provider, then reads like any attached document - with an honest message if no provider is set up. Video analysis now also listens: the soundtrack is transcribed alongside the sampled frames, so "what is said in this clip" finally has an answer. Both are best-effort and never block the run.
+- **Declining says what you keep, and stops handing out an old build.** The refusal screen shows the data folder, untouched, and now says the thing that is actually true and costs nothing: an installation that was already running is unaffected and stays on the licence it was supplied under. It names that licence - the Business Source Licence 1.1 - and quotes the part of it that has not changed: its Additional Use Grant covered personal use, learning and non-commercial internal testing, so commercial use needed a licence there too. What it no longer does is offer a download of a superseded release. The same screen is shown to somebody installing for the first time, a superseded build gets no security fixes, and an installed copy can never change the address it shipped with - so the one link goes to the published licence and nowhere else.
-### Changed
+- **The app carries no price, and it never will again.** The Business page held twelve euros a seat in twelve languages, in a block where the numeral and the unit were two type roles - and a website can be corrected every day while an installed copy keeps what it shipped with until somebody updates. In its place stands what is true today: an evaluation licence, free, the full scope, no seat limit, confirmed in writing with an end date, which is a written commercial licence at no charge rather than a stretched trial. The commercial terms come down the same live wire as the figures, so setting a price later is a change on the server and not a release, and a machine with no connection shows the honest statement instead of a number that has aged.
-- **Server actions joined the token gate.** With remote access on, the token now guards the entire local surface - not only the /api routes but the internal action calls pages are built on. Verified against the packaged server: without a token, nothing executes and nothing is written.
-- **Frames fold into an accordion.** A video message shows its sampled frames as a compact expandable strip under a thumbnail instead of ten full-size images, and the bubble reads like your question again instead of an attachment note.
-- **The knowledge graph stays yours.** Machine-written sessions (briefings) and Skales' own product terms no longer enter the graph - the extractor filters them at every entry point, and Settings offers a one-time cleanup that removes previously collected ones without touching your real entities.
+- **The Business page stops saying things it is not allowed to say.** "Free, with your own key" now names which half is free where it is said: your own key decides who pays the model, the licence decides whether you may use Skales commercially. The line about nothing being locked in no longer says "stop paying and the free agent keeps working with your own key" - a sentence about the end of a payment, with the licence unmentioned, reads on a business page as a free commercial alternative. And the waiting list says at the point of consent that Skales may write about commercial licensing, with what the box asked for stored beside the privacy notice version it was agreed against.
-### Fixed
+- **What the offer carries instead of a badge: your logo on our page, and your name to your own clients.** The fifth line of the evaluation licence used to promise a founding badge on an account and thirty minutes on the phone. From v13 it is a company logo under "Trusted by" on skales.app, linked back to your own site, and your name listed there as a qualified contact for setting up your clients - a link somebody can follow and a mention somebody can act on, instead of a mark inside an account only you can see.
-- **Friend Mode no longer double-texts.** Two overlapping schedulers could each generate their own check-in inside one cooldown window - two different messages in the same minute. One in-flight guard now spans both, including across the app's two server bundles.
-- **The @ and / menus close when you click away** - and a slow, deliberate click on an entry no longer loses the selection.
-- **Stop works while frames are being sampled.** The red stop button now aborts a video analysis during the "sampling video frames" phase instead of only after it.
-- **The pairing QR told phones the wrong desktop version** - hardcoded at 10.1.0 since that release; it now always carries the real version.
+- **The v13 section says what is coming, not what is missing.** The band over it read "not built yet", and the paragraph under it listed what the download you can get today does NOT contain. That it comes with v13 already says it is not here now. Nobody reading a business page should have to learn our release plan to understand an offer.
-## v12.4.0 - Piranha
+- **FFmpeg is in the third-party notices, with its licence and where to get its source.** Skales uses it to merge audio and render video, and it ships in the package - but only the small JavaScript wrapper around it was ever attributed. The build that ships carries the GPL, version 3 or later; FFmpeg runs as its own program and is never built into Skales, so those terms cover FFmpeg and not the application around it. Its full licence text and FFmpeg's own notice now travel with every download, and the notice says where the matching source can be had.
-Generated media shows up where you asked for it, your phone sees what the desktop is doing and can stop it, goals know when they are done, and every agent gets its own name, hooks and task list.
+- **Everything in the download is attributed, and a new piece cannot slip in unattributed.** The notices used to be assembled from a list that could not see the programs added to a package last. Whatever ends up in a download now has to be named and described before the build will finish.
-### Added
+- **The third-party notices read as the public document they are.** The file ships with every download and is published beside the licence, and it was carrying a block of internal notes and a handful of internal file paths out with it. Both are gone, and it stays that way.
-- **Your phone sees a running desktop chat - and can stop it.** When the paired app has a session open that Skales is working on, the phone shows a live banner with a Stop button, streams each tool step as it happens, and receives generated images right in the bubble. The desktop also gained a "send to phone" action that forwards a workspace file or note to the paired app, with a push if the app is closed. Work done from the phone now shows up as proper tool cards when you open that session on the PC.
-- **User hooks.** Run your own text snippet or shell command at chat lifecycle points: session start, after a tool result (optionally one specific tool), or when a goal finishes. Command hooks go through the exact same safety gates as the model's own commands, and isolated agents never fire them.
-- **Tick the task list yourself.** The checklist Skales keeps in a normal chat is now clickable when the chat is idle: mark items done or undone by hand, and your ticks survive reload. A fresh plan from the model still wins over stale hand edits.
-- **Publish over SFTP.** FTP publishing now speaks SFTP as well as FTP and FTPS: pick the protocol per profile, and on the first successful test Skales shows the server's SSH host-key fingerprint and pins it to the profile. If that key ever changes, the next upload is refused, so a swapped-out server can't quietly receive your files.
-- **Goals can declare themselves done.** A long-running goal with broad criteria used to be nudged to keep going forever; the agent can now state completion explicitly once the work is verifiably done, and the checker respects it.
-- **Isolated agents got a real household.** An isolated agent's generated and edited images stay in its own workspace instead of your gallery, downloads are confined to its folder with a hard size cap enforced mid-download, its dates are labelled UTC instead of your timezone, planner tasks run as the agent that owns them, and each agent can have its own email account without seeing yours.
+- **A picture a tool took is shown in the coding window, under the call that took it.** Until now a screenshot taken during a run reached Skales Code as a message from YOU: pictures travel to a model in a user-role slot, because that is the only slot every provider wire has for an image, and the window read that transport detail as authorship. So a live check with two screenshots in it produced a turn on the right-hand side, in your name, carrying the fence written for the model - with an edit button and a send-again button on it. The chat has told the two apart for months; the window now asks the same reader, and goes one better than the chat can: it has a line per step, so each picture hangs under the call that produced it rather than under the round. A tool that merely shows a picture again does not draw it a second time - it says it showed one from further up - while two captures of a screen that did not move stay two looks at the screen, however identical their bytes.
-### Changed
+- **The pictures you attach to a coding turn are in the transcript.** They were counted ("2 pictures") while the session was live and vanished entirely once it was reopened, because saving replaces the parts with plain text and keeps the files beside the session. The same thumbnails the composer showed you before you sent are in the log afterwards, and after a restart, and any of them opens full size.
-- **Multi-agent runs report in the conversation.** The per-agent progress trace now lives under the message that dispatched the job, and the finished report keeps a compact version of it, instead of a banner over the whole chat.
-- **The Knowledge Graph got a proper physics engine.** The map lays out with d3-force, settles quickly at up to 300 nodes (up from 150), and stays smooth to drag, zoom and expand.
-- **Media cards are forgery-proof.** Image and video cards in chat render only from real tool results, never from text that merely looks like one, so a model cannot fake a "generated file" card. Video bubbles from before this release show their file path as plain text instead of a card - the file itself is untouched.
-- **Images open in the lightbox.** Clicking an image card in chat opens the full-size viewer; the dead "Open in Studio" link on media cards is gone.
-- **One-tap deploy knows Cloudflare.** Deploying a Code-mode project now recognises a Cloudflare Wrangler setup alongside Firebase, Vercel, Netlify and an npm deploy script, and points you at your own CLI for anything else.
-- **Skales IQ asks for an update when it needs one.** If a build is too old for the Skales IQ service, chat now shows a clear "install the latest update" message instead of a raw error.
-- **Custom agents answer as themselves.** Ask a custom agent who it is and it answers under its own name (still honest about the model it runs on) instead of introducing itself as Skales.
-- **WhatsApp knows who "me" is.** "Send it to me on WhatsApp" reaches the owner instead of the bot's own number, contacts saved in national format (0676...) match, and a failed media send explains the file path problem before anything leaves the machine. The isolated-agent chat badge became a lock icon.
+- **A block of code in an answer is finally treated like code.** It arrives tinted by the same scanner the file viewer uses, so a fence and the file it was cut from are not two different colour schemes; it carries the language it claims to be; and it has a copy button - the same button the open tool step has, not a second one that looks like it. It also moved off the dark terminal plane onto a card, where those colours were chosen to be read; the terminal plane now belongs to actual command output, which is what it was for.
-### Fixed
+- **You can see what the model thought.** The window has offered a reasoning-effort dial for months and then showed a blank column while the effort was being spent. The thinking is folded under the answer after the turn lands, and it is open WHILE the turn is still only thinking - it closes itself once, the moment the answer starts, and after that the fold is yours.
-- **Generated images render in chat and Flow again.** An image produced by the built-in generator shows up as a picture in the conversation and lands in the Flow project folder, on every backend - previously the turn ended early with a broken link.
-- **Video generation works in the installed app on every platform.** The Windows build was missing the Google video engine and ffmpeg entirely, and every platform's package now carries the ffmpeg binary for its own operating system instead of the build machine's. Veo model names are consistent everywhere, so the tier you pick is the tier that runs.
-- **Finished answers stop re-typing themselves.** At the end of a streamed reply the text no longer wipes and types out a second time.
-- **Google Calendar tells you what is actually wrong.** The dashboard widget and calendar tools now name the real cause when the connection fails - an expired authorization (with the 7-day testing-mode hint), missing OAuth fields, or a network error - instead of a generic "reconfigure" line. And an empty week on a connected calendar links to the planner instead of suggesting you set up a calendar you already set up.
-- **Web search fails forward.** A failed search names the fallback that will work (fetch the page, extract text), Windows gets shell hints that match its own tools, and the keyless DuckDuckGo path retries with a hard 25-second budget instead of hanging.
-- **Three tools were invisible to the model.** A manifest generator gap hid connector building, connector requests and web search from the tool catalog in some setups; all 182 tools are listed again and a guard keeps future descriptions from silently dropping out.
-- **A broken video clip cannot stall the chat.** Analyzing an uploaded video now stops sampling after a minute and answers from the frames it already has, instead of hanging on a slow or partly corrupt file.
-- **The chat input can no longer stay locked.** In rare cases the composer stayed disabled after a reply finished (Enter and Send dead until a hard reload); a watchdog now releases it whenever no run is actually in flight.
-- **A dispatched multi-agent job is always visible.** When the dispatching reply carried no visible tool card, the live worker chips had nowhere to attach and the job looked like nothing had started; they now attach to the latest reply instead.
-- **Video renders stop making promises.** After starting a background video render the model no longer invents wait times or tries to poll a progress URL; it states that the render is running and where the finished file will appear.
-- **Chat exports skip empty blocks.** A reply that only ran tools exported as an empty "Skales" section; exports now show a compact tools line instead, in both the /export command and the History download.
-- **Background runs stop paying for the whole tool catalog.** Multi-agent workers, planner tasks and scheduled catch-up runs used to receive every tool definition (tens of thousands of tokens) on every single step; they now start from the same lean core set as chat and pull extra tool groups on demand. On a metered model this was the single largest hidden cost per day. GLM 5.x models get on-demand tool loading too, instead of always paying for the full catalog.
+- **A file a step wrote is a card, not a path in a sentence.** Name, type, size, download, open the folder - the card the chat has had, with the check that the file is still there, drawn in the coding window's own colours.
-## v12.3.5 - Flying Gecko
+- **A turn that looked something up brings its cards with it.** Repositories, articles, places and their map render in a coding session the way they do in a chat.
-Your knowledge graph comes alive, multi-agent jobs show every worker and finish on their own, and finished work stops being sent twice.
+- **One card, drawn twice, is one card.** The three shared cards above are the chat's own components mounted inside the coding window through a token bridge: the window defines the names those components read out of its own palette, so they inherit Skales Code's colours in light and dark without a chat value ever reaching them. The alternative was seven hundred lines of card rebuilt a second time, and two copies that disagree with each other within a release.
-### Added
+- **The + that starts a new chat is the one filled button in the chat header.** It sat in the same grey outline as the six tool buttons beside it, so the way OUT of a conversation looked exactly like the controls that act on the one you are in - seven identical outlines, and the one that leaves is not the same kind of thing as the six that act. It is filled in the accent of whatever theme and accent tone you run, and the lettering colour is COMPUTED against that accent rather than assumed dark, which is what keeps it readable under Skales X, Classic and Flat, in light and in dark, and under an accent you picked yourself. It is the only filled control in the row on purpose: a second one would take the distinction away again. Nothing about its size, its spacing, its radius or its icon changed, so the row still collapses the same way on a narrow window, and it is the same button in the compact window.
-- **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 can 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. Light and dark both fit, and Reduce Motion drops straight to the settled layout.
-- **Watch every agent in a team job.** When Skales fans a task out to several workers, the chat now shows one live chip per worker moving from queued to working to done or failed, with a running count and elapsed time, instead of a single "running" line. Click it to open the Tasks view.
-- **A private scratchpad for each conversation.** Skales can now keep its own working notes for a chat and gets a private work folder per session, so intermediate files from one conversation never collide with another. Its notes come back to it at the start of every turn, so a long task keeps its train of thought.
-- **Drop a video into chat and ask about it.** You can now attach a video (or drag one in) and have Skales actually watch it: it samples frames across the clip and reads them together, so you can ask what happens, check why a short is working, or figure out how a clip was made, its hook, pacing, cuts and on-screen text. Works with any capable model; camera clips just need FFmpeg, which installs from Settings.
-- **Look through a camera.** Skales can now take a single frame from a camera on demand and tell you what it sees: an IP or WiFi camera address, or a camera already set up in Home Assistant (which is how a Ring doorbell is reached). Ask "is anyone at the door" or "what's in the garden", and pair it with a schedule to check a camera every few minutes and only ping you when something changes. It asks before looking, and needs a Vision Provider.
-- **Isolated agents with a life of their own.** A custom agent can now be marked Isolated: it runs with its own memory, its own workspace folder and only the tools you hand it, and it never sees your identity, facts or saved memory - not even in the background passes that build your knowledge over time. Pair it with a pinned model, a bound FTP profile and a scheduled goal and it can run a project of its own, end to end. If its definition ever can't be loaded, the run stops rather than falling back to your data.
-- **Publish a folder over FTP.** A new publish tool uploads a workspace folder through a saved FTP profile, and profiles can be bound to a single agent so only that agent can use them. FTPS (explicit TLS) is supported and 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 now carry a "Let your agent explore this space" button: one tap prepares a visit for your own agent - it reads the site's welcome file, looks around and leaves a short guestbook comment, with your usual confirmations. You review the request before it's sent, and each post can be run once.
+- **Wrapped stops being a switch you can forget you turned off.** It never had a menu row - its one door is the button in the Discover header beside Spark - so a switch of its own could do nothing but take that button away and leave the page unreachable from the navigation. The switch, its card on the Add-Ons page and both readings of it are gone, the button is drawn for everybody, and whoever does not want Wrapped does not open Discover.
-### Changed
+- **The Playbooks switch is gone, because it decided nothing.** Recording and replaying live in the chat and inside the Browser surface, and the two tools were always handed out with the browser tools - so the switch on the Add-Ons page changed nothing except what Skales SAID about playbooks, and it could say "switched off" while playbook_run sat in the very same turn. The page, the palette entry and the tools are untouched; what is gone is a control that never controlled anything, and the assistant now reads the Browser Control switch when it answers for them.
-- **Notifications carry their real button.** Announcements now show exactly the action label they were written with instead of a generic "Learn More", update notices get their own look with a one-tap jump to the in-app update page, and messages can be aimed at desktop or mobile users specifically.
-- **Signing in by hand works everywhere.** The visible sign-in window for website logins no longer announces itself as an automated browser, so Google and other providers accept the login instead of blocking it as insecure.
-- **Toast messages stay put.** Stacked notifications in the corner could leave an invisible gap behind that pushed each new toast lower until the stack drifted toward the middle of the screen. Old toasts are now always cleared, and any that a background window left stranded are swept up on the next one, so the stack stays anchored top-right.
+- **Organization is opened from the chat, and from nowhere else.** `/organization` is its door on this machine exactly as it is on the phone: no sidebar row, no palette hit, and a card that names the command instead of a menu entry. The switch keeps everything it gates - the team tools and the team block the assistant reads.
-### Fixed
+- **The sidebar stops asking every minute about a badge it cannot draw.** It fetched the weekly summary once a minute, for the whole time a window was open, to light a dot on the Wrapped row - and that row has not been drawn since the entry was taken out of the navigation. The request, its timer and the dot are gone; the Notifications dot beside it is untouched.
-- **A team job finishes on its own.** When Skales split a goal across several workers, it could stall once they finished and wait for you to nudge it. It now picks the goal back up automatically 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 that had already happened, so a file or message could go out two or three times. Once a step has genuinely succeeded, its result now counts as proof and is never repeated to "re-verify" it.
-- **Ask for an image, get an image.** When you ask for a picture, poster or graphic, Skales now creates it instead of handing back a text description of what it would draw. If no image provider is set up, it says so plainly and points you to Settings rather than quietly returning a prompt. And right after a big task wraps up, a follow-up like "now write the launch post" reuses what was just made instead of starting the search over.
+- **The Add-Ons page and the first-run picker stop promising doors that were closed.** Two entries still said Organization and Playbooks had their menu rows back and one said Wrapped's row follows its switch, months after the rows were taken out; the check that would have caught it existed only on the phone. It exists here now: an add-on offered during setup whose whole promise is a page has to name a door - a row, or a written reason why it has none - and the same reading found one more card that could never have shown the mark. What Skales tells you about Templates, Organization, Wrapped and Playbooks matches what the windows actually do again.
-## v12.3.0 - Flying Gecko
+- **The Plugins page says what a plugin is for, instead of how it is stored.** It opened on "a plugin has its own menu entry, its own page and its own storage" and an empty text field, and the one route that actually works - describing what you need in a chat - was named only in the empty state, which is to say only until the first plugin existed. The page now opens on what you get, and under it a section that explains the whole thing once: what a plugin is, that anything which would send, publish or spend waits for one click from you even at three in the morning, and the three ways one gets here as three cards rather than one button and two sentences. It closes with an X in the corner and does not come back - the flag sits in the settings file next to every other machine-level preference, so it survives a restart, a cleared cache and a restored backup rather than living in browser storage like the four other one-time notes in this app.
-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.
+- **The shelf and the catalogue are one card language now, not two.** The same object was drawn as a full-width row in one place and a two-column grid card in the other, with different padding, different glyph sizes, and permissions rendered once as a sentence with middots and once as capsules. Both are the same card, the same glyph tile and the same pills, from one shared file - and what a plugin may do reads as separate facts instead of a run-on line.
-### Added
+- **Which tools a plugin may use is a question you can answer.** The picker was a flat list of up to two hundred bare names; the category each tool belongs to was being loaded and never drawn. It is grouped now, with a count of what is on per group, the house switch instead of a raw browser checkbox, and a line naming what is never on offer at all.
-- **Sign in to a website, once.** When a task needs you logged in, like posting on X or checking an order, Skales opens a visible browser window so you can sign in by hand. Your login is saved to the Skales browser profile and reused the next time, so it can work inside your accounts without ever seeing your password. Open a login window any time from Settings, Browser Control ("Log in to a website").
-- **Watch the browser work.** A new "Show browser window" switch under Settings, Browser Control opens a real, visible browser window instead of running hidden, so you can see each step as it happens.
-- **Multi-line scripts run without shell traps.** Skales can now write a script to a file and run it directly with the right interpreter (Python, Node, Bash, PowerShell) instead of squeezing code through a shell one-liner. On Windows especially, quotes, braces and special characters used to break perfectly good code before it even ran; that whole class of failure is gone.
-- **Every file change has an undo point.** Before Skales overwrites, edits, appends to or deletes a file, it quietly keeps a copy of the previous version. If an edit turns out wrong, the earlier version is right there to restore, 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 now remembers when it last ran successfully. A tool that has not proven itself recently is marked as such, so Skales tests it before relying on it and reads it before changing it.
-- **Flow checks its own work before handing it over.** After Flow builds or edits a deck, prototype, page, or visual, it now takes a picture of the finished result and looks at it the way you would, checking for text that runs off the edge, unreadable contrast, pieces cut off outside the frame, or a layout that has collapsed. If it spots a clear rendering problem it quietly fixes that one thing and shows you the corrected version. It makes at most one extra pass, never holds up a finished draft, and only steps in when a Vision Provider is set up. You can switch it off under Settings, Vision Provider.
-- **Fill your memory graph from what Skales already knows.** Once you turn on learning in Settings, Memory, the Knowledge Graph can now be seeded in one pass from your saved long-term facts and a slice of your recent chats, instead of only filling up reply by reply going forward. Choose how much history to pull in, watch it count through in the background, and stop it whenever you like. Running it again never creates duplicates.
-- **Your desktop makes Studio visuals for a connected phone.** When your phone is paired in remote mode but has no provider key of its own, it can now ask your desktop to build a Studio visual for it. The desktop generates it on the side and sends back the finished design, without ever adding anything to your chat history.
-- **Friend Mode can reach your phone.** Turn on "Mobile App" in the Friend Mode channels and your desktop's proactive check-ins are sent to your paired phone too, where they land in your Buddy thread. It uses the same cooldowns and quiet hours as every other channel, and the message stays encrypted end to end.
+- **The page holds together on a narrow window.** The row's controls used to sit in a column that never wrapped, so as the window got smaller the description was squeezed into a strip one word wide. Everything below the glyph now shares one column that wraps, the tool search takes the full width when there is no room beside it, and the sentence explaining a tool is the first thing dropped rather than the thing that crushes the layout.
-### Changed
+- **A plugin page can have icons now.** It never could: a plugin runs in a sandbox the app's own icon components do not reach, so all ten bundled plugins shipped with not one glyph in them, which is most of the reason they read as documents rather than as tools. Sixty-two lucide glyphs are available to a page by name, handed in the same way the house fonts are - only the ones a page actually asks for, so a page without icons costs nothing. A name this build does not know still draws the plugin square, because a `use` pointing at nothing renders nothing and a row that loses its glyph shifts for a reason nobody can see. The geometry is generated from lucide rather than traced, and the build fails if the two ever fall out of step.
-- **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.
-- **The Proactive operator says what it is waiting for.** An empty operator pane on the Memory page used to show only "No initiatives yet", which read like a broken feature. It now explains what Skales is watching (upcoming meetings, failed scheduled tasks, blocked work, unread email) and shows when the last background check ran, so quiet genuinely reads as good news.
-- **Skales reads before it edits.** Changing a file it has not actually looked at in the current session is now refused once, with the instruction to read it first. What is on disk can differ from what the model remembers, and blind patches on working scripts were the main way tools got destroyed.
-- **Failing approaches now end in an honest question, not a seventh attempt.** When the same tool keeps failing, Skales first gets told exactly what failed and is pushed to step back and test one hypothesis at a time; if the failures continue, it stops trying on its own, summarizes what it attempted, names its best guess at the cause, and asks you for the one thing it needs. No more endless rabbit holes.
-- **Unrestricted mode keeps its word.** A few command rules used to stay blocked even with Safety Mode set to Unrestricted. In true Unrestricted they now run instead of being refused, and the result card in the chat carries a clearly visible amber warning strip naming what was flagged, pinned to the exact command it belongs to and kept in the conversation history. Safe and Advanced keep their protections unchanged.
-- **The working glow is easier to notice.** The ambient background that breathes while Skales works now drifts further, on a slightly quicker rhythm, and sits a touch more present at its peak, so you can actually see that something is happening. It stays well within readable contrast, idle stays still, and Reduce Motion still switches it off entirely.
+- **A plugin now follows the theme in shape as well as in colour, and it keeps following it.** The frame handed a plugin page seven values, and all seven were colours - so a plugin looked exactly the same under Flat, which is square and unshadowed, as under Skales X, which is neither. The radius, the shadow and the easing travel now too, with Flat's own values for them arriving for free. And switching theme or accent while a plugin is open used to leave that page on the colours it opened with until you navigated away and came back: the new values are pushed into the running page instead, without reloading it out from under whatever it was doing.
-### Fixed
+- **Every plugin page now stands on the same floor.** There were three: the ten bundled plugins carried a real little design system, a plugin started by hand got twenty lines with its own hardcoded colours and a fixed English sentence, and anything the chat built got eight lines - box-sizing, no margin, a colour and a font. The one that was good was reachable from ten files and nowhere else. It now lives in one place, is laid under EVERY plugin page by the frame, and has grown the parts that were missing: a head band with a glyph, a title, a sentence and one primary action; sections and toolbars; rows that lift under the cursor and give way when pressed; a figure with its label for the counts; a skeleton in the shape of what is loading instead of a spinner over a blank rectangle; a failure band that is not the same thing as a remark; a focus ring; and a staggered entrance that is off entirely for anyone whose system asks for less movement. It is a floor and never a ceiling: it sits in the head, so a page that styles something itself still wins, and it contains no `!important` anywhere. A page that never thought about its edge gets one, and takes it back with one line of its own.
-- **Long conversations no longer lose track of the task.** In a long session, especially one with many working steps, Skales quietly fed the model only the most recent stretch of the conversation, so the original request could fall out of view while the context meter still showed plenty of room, and a later remark could be mistaken for the task itself. The model now always works from the full conversation, sensibly condensed, and the original request stays anchored no matter how long the session gets.
-- **Work already done is not redone.** When a long task was interrupted and you asked Skales to continue, it could lose track of the files it had already written, claim a script was gone, and write it again from scratch. Skales now keeps a live list of everything the task has written, checks that each file is really on disk, and picks up from there instead of starting over.
-- **Skales changes course instead of giving up.** When a step kept failing the same way, Skales used to stop with a generic "I got stuck" message, throwing away an otherwise healthy task. It now tells the model exactly which call keeps failing and why, blocks that one call, and pushes it to try a different approach; only if it still cannot move on does it stop, and the stop message now names the step that jammed instead of leaving you guessing.
-- **The calendar tells you why it is empty.** If none of your configured Google calendars could be read, most often because a calendar's display name was entered where its ID belongs, Skales showed an empty calendar as if you simply had no events. It now says which entries failed and how to fix them, and the settings field explains where to copy the real Calendar ID from. One broken entry among working ones still does not disturb the rest.
-- **The Memory page tells the truth about Knowledge Graph learning.** The graph card claimed new facts were extracted automatically after every reply, while that learning is actually a switch that ships off, hidden under a name nobody connected to the personal graph. 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.** Asking which mode is active used to get you a guess or a request to check Settings yourself. Skales is now told its active mode and what that mode means, and answers directly.
-- **Legitimate commands are no longer refused as dangerous.** A handful of everyday commands, such as scheduling a task on Windows or listing output with a format flag, were wrongly treated as system-destroying and refused regardless of your Safety Mode. Command safety is now decided in one place, honours your chosen mode everywhere, and only the genuinely destructive commands stay blocked.
-- **Windows: scripts no longer crash on special characters.** A Python script that printed an emoji or other non-Latin text could die with an encoding error on Windows. Script output now uses UTF-8 there by default.
+- **A plugin you start by hand opens on the house look, in your language.** The blank page it creates carries the same floor every bundled plugin does and says what it is in the language the app is set to, rather than in English regardless.
-## v12.2.5 - Past Freeze
+- **The Business page now shows what Skales already does instead of claiming what it might.** The page opens on one night's work - receipts read out of the mailbox into a month sheet, an invoice chased with the reminder already written, a post published with its image and its SEO, a summary sent to the phone - with the tool behind every line named. It says out loud that the reminder was drafted and not sent, because a scheduled run has nobody at the screen and sending waits for a click. The old opener promised an agent that "finds clients", and there is no lead, no source and no duplicate check anywhere in the tree to make that true.
-Your phone can now start, watch and stop the work your desktop does. Agents know who you are and what day it is. Vision lands on the right provider, the composer never stays stuck, and Flow stays out of your chats.
+- **Six things Skales already stands for, before anything about a price.** It runs on your machine; your key or no key at all; Skales IQ carries zero data retention on the request itself rather than as an account setting; it asks before it sends, publishes, deletes or spends; nothing locks you in, since a plugin exports as one file and the free agent keeps working with your own key; and erasure is a branch that runs, with a 36-month sweep behind it. Underneath them, the boundary in one line: this is not a certificate, and what a procurement reviewer gets is a document naming the providers a request can reach.
-### Added
+- **Skales Business is on the page, and marked as not built.** Workspaces, cost control with a usage journal and budgets that reserve before a run starts, infrastructure profiles for a company gateway, department packs, joining by invitation, SSO and SCIM, and an admin portal that runs without us. It is drawn as a rail with hollow marks rather than a row of cards with ticks, because a tick is a claim that something is there.
-- **The chat background breathes while Skales works.** The accent glow slowly grows and drifts while your question is being thought through, stays a touch more present while the answer arrives, and settles back when it is done instead of snapping off. A running goal gets its own livelier, greener state, so you can tell at a glance that something is still working for you. The thinking dots and the status bar move on the same tempo. Nothing moves while Skales is idle, the animation stops while the window is in the background, and it is off entirely when your system asks for reduced motion.
-- **Skales Mobile is on the App Store.** The Skales Mobile page now links the App Store next to Google Play, and the iOS button opens the store in your browser like the Android one. The in-app guide and roadmap note both stores are live.
-- **Start a goal from your phone.** A request the phone marks as a goal now runs on the desktop as a proper background goal, on its own session with its own plan, instead of one long blocking reply. It keeps working after the phone acknowledges it, so you can close the app and come back to the result. Goals honour the Settings goal switch and need a real provider (your own key or a local model), like the scheduled goals.
-- **Control a goal from your phone.** The Skales mobile app can now list the goals running on your desktop and stop, pause, or continue any of them. Stop and pause halt the live work at once, the same as the desktop.
-- **Get notified on your phone when a goal finishes.** When a goal started from your phone completes, pauses for input, or stops on an error, the desktop notifies that phone, even if the app is closed. The notification is routed only to the phone that started the goal, and it arrives as its own message so it never interrupts whatever else you are doing in chat.
-- **Your phone sees the live plan.** When the desktop works through a multi-step task from a phone request, it sends the same step-by-step checklist to the Skales mobile app, which shows it as a strip above the chat, updates it the moment a step changes, and brings it back when you reopen the conversation.
-- **Goal and origin badges reach your phone.** The desktop now shares each session's goal progress and whether it started on a phone, so the phone shows the same goal pill and Mobile badge you see on the desktop.
-- **Browse and fetch desktop files from your phone.** The phone's Workspace shows a Desktop folder while connected: browse your desktop workspace and pull any file (up to 2.5 MB) to the phone over the encrypted link. Read-only and strictly limited to the workspace folder.
-- **Finish a connector's setup from your phone.** The Skales mobile app can now list your desktop's API connectors and complete the parts that used to require the desktop: confirm the connector's domain and set its key (and the client id and token address for OAuth2). The key is sent once, stored encrypted, and never read back; confirming the domain and setting the key stay separate steps.
-- **Discover shows when someone forks a skill or template.** Forking a skill or a template now posts a short note to the feed, the same way sharing does, and only when feed sharing is turned on.
-- **Safety Mode now has three levels: Safe, Advanced and Unrestricted.** Safe blocks dangerous shell commands and asks before critical actions. Advanced runs without asking and allows dangerous commands, while a residual guard still refuses the handful that would destroy the machine or its data outright. Unrestricted removes those last blocks too and, with file access also set to unrestricted, lets Skales reach system locations, install software and talk to hardware. If you were on the old Unrestricted, it is now called Advanced with identical behaviour, and a one-time note on first start points you to the new, genuinely unrestricted mode. You change it under Settings, Security, or during onboarding.
-- **Name an image or video model in plain words.** You can ask for a model by loose name, for example "flux schnell", "gemini flash" or "veo 3.1 flash", and Skales matches it to a model your providers can actually run, picking the right backend for it. If the name is ambiguous it asks which one you meant instead of quietly falling back to a default.
-- **Skales can set up an API connector from documentation, mid-conversation.** In Advanced or Unrestricted mode, after reading a service's API docs Skales can scaffold one of your REST connectors during the chat. The connector is saved switched off and without a key; you confirm the address and enter the key yourself under Settings, API Connector before it can be used, so the key never passes through the model.
-- **API connectors speak OAuth2 (client credentials).** A cloud API that exchanges a client id and secret for an access token at a token endpoint, like Husqvarna or Gardena smart systems, now works as a connector: pick "OAuth2 (client credentials)" as the auth scheme, enter the token URL and client id, and put the client secret in the key field. Skales fetches and refreshes the token on its own, entirely on your machine. Static extra headers such as an application key are supported too.
-- **Tools Skales builds for itself survive the session.** When Skales writes itself a working script, say a Bluetooth scanner or a TV caster, it keeps it. Every future conversation knows about those tools, so Skales reuses what it already built instead of starting over.
+- **The free half is named before the paid half.** Ten plugins, every tool, triggers on the clock, a folder and the mailbox, and any provider key stand in their own column beside what the seat adds. A buyer does that subtraction anyway.
-### Fixed
+- **The offer is an object rather than a sentence in gold.** What it is stands at the display step with its scope beneath, and the five conditions run down the same object as a hairline column instead of five full-width cards that outweighed the offer itself. The geometry does not care whether the left cell holds a name or a figure, which is what lets the terms arrive over the wire. The partner terms and the v13 sunset are two statements on two lines, not one that trails off.
-- **Agents now know your saved facts and today's date.** A custom agent, or an agent from the Agents page such as the Research Analyst, ran with its persona only: none of your Knowledge & Facts applied, and it did not know what day it is, so a web research report could carry the publication date of its sources as the report date. Every agent now gets the real current date, is told that a date found in a source belongs to that source and is never today, and receives the facts you saved. Agents that read the web are also held to the same rules about untrusted content as the rest of Skales.
-- **Your saved facts survive minimal compression.** With compression set to minimal, or chosen automatically for a small local model, what Skales knows about you was shortened so aggressively that every entry under Memory, Knowledge & Facts fell away, so a fact you explicitly saved was simply absent when the model answered. Your saved facts are now always kept.
-- **Stopping a running goal from the header now actually stops it.** The Stop control on the goal activity pill only marked the goal stopped without halting the work, so it kept running and reappeared on the next refresh. Stop and Pause now halt the live run at once. A stopped or crashed goal also no longer silently resumes itself on your next message: an interrupted goal is parked so it only continues when you choose Continue, and a goal that hit an error is marked failed and leaves the active list instead of looking like it is still running.
-- **A long task no longer times out on your phone while the desktop is still working on it.** In remote mode a request that ran long (a goal, a slow build, one long model step) was given up on and its real answer dropped, and a retry re-ran the whole thing. The desktop now keeps the phone informed while it works, so the phone waits and the answer arrives.
-- **OpenRouter works as your Vision Provider.** Switching the Vision Provider to OpenRouter kept sending image reads to whatever address a previous provider had left behind (a local Ollama, for example), so every read failed with a "not found" error even with a valid image model, while the exact same endpoint entered as a custom provider worked. OpenRouter now always uses its own address, a leftover address from a previous provider is cleared the moment you switch providers, and a "not found" error now names the model and how to correct its id.
-- **The image description sits under the picture, inside your message.** The collapsed eye view with the Vision Provider's description floated above the message bubble instead of belonging to it. It now sits inside the bubble directly under the attached image, styled to read clearly on the bubble's background, both live and after reloading the conversation.
-- **The composer no longer stays stuck after a reply stops responding.** If a reply stalled partway (a hung tool call, an orphaned run, a dropped connection), the input stayed locked with no way to send until you reloaded the page. It now releases on its own after a run goes quiet for too long, with a short note that you can send again, and without stopping whatever the server is still doing. Typing in very long conversations is lighter too.
-- **Skales Visuals no longer times out while it is still working.** Creating an image or visual in Flow could fail with "the operation was aborted due to timeout" even while the model was actively writing, because a full single-page visual routinely takes a while. Flow now allows the whole generation and only gives up on real silence, and when a timeout does happen the message explains what stopped and suggests trying again or picking a faster model.
-- **Forked and created templates appear in your library.** Forking a template from Discover, or creating one, showed a success message but the template never appeared under Templates, because the library never read your saved custom templates. They now show up with a Custom or Forked chip and a delete button, so a mistaken fork is not permanent, and a forked template carries its full prompt instead of a short preview.
-- **Flow projects stay out of your chats.** A Flow (Studio) project could surface as a recent chat on the dashboard, in history, in the command palette and on your mobile device, and finishing a Flow turn popped a "finished your chat reply" note that opened the Flow conversation inside the chat view. Flow projects are now kept out of every chat list, a finished Flow turn shows a Flow note that reopens the workspace, and opening a Flow project link sends you to Flow instead of rendering it as a chat.
-- **Long answers from Ollama Cloud models no longer cut off mid-sentence.** A strong model served through Ollama Cloud was limited to a small answer length regardless of what it could actually write, so longer replies stopped partway. Each family now uses its real answer length, and you can still raise it under Override Model Limits.
-- **Ollama Cloud models use their tools instead of just talking about them.** A frontier model running on Ollama Cloud was treated as if it were small hardware on your own desk, so it described what it would do instead of doing it, and a slow first response could wrongly mark it as not supporting tools at all. These hosted models are now recognised for what they are.
-- **Reasoning no longer leaks into the chat as "[thinking]" text.** Some models write their private reasoning inline using square brackets, which showed up as literal "[thinking]" text in the reply. That reasoning now goes into the collapsed Reasoning section like every other model, and the visible answer stays clean.
-- **Image generation works when your only provider is fal.ai or Replicate.** Asking for an image in a chat or a Flow could return a written description instead of a picture when the only image key you had configured was fal.ai, or a Replicate key stored the newer way, because those two were overlooked when deciding whether to offer image generation. Every image provider you have configured is now recognised, so the tool is offered whenever any of them can produce an image.
-- **A PowerShell display command is no longer mistaken for disk formatting.** Formatting command output for display (for example listing Bluetooth devices with Format-Table, Format-List or Format-Wide) was refused as if it were a drive-format command. Only real drive formatting stays blocked now; the display formatters run.
-- **Pairing a phone opens the right panel and confirms right away.** After scanning the pairing QR code, the panel jumped to Settings instead of Connected Devices, and a scan showed no confirmation until a background refresh caught up. It now opens Connected Devices immediately with a success note.
-- **Removing a paired phone really cuts it off.** Removing a device only closed the current connection, so the phone could quietly reconnect on its own. Removing a phone now tells it to forget this computer and refuses to re-pair it until you deliberately show a fresh QR code, so a removed device stays removed.
+- **Enterprise and white label read as an invitation:** "Tell us what you need and we will talk about it. Starting in 2027."
-### Changed
+- **Your memory has a ceiling now, and the Memory page says what it is.** The three record folders behind the counters grew by one file per turn and nothing ever trimmed them: on this machine 519 recent notes, 556 long-term records and 598 journal entries in eight weeks, of which 465 were written overnight by the dreaming pass rather than by anybody. That is not a disk problem, it is a reading one - the list is read from start to finish every time Skales builds a prompt, and at thirty-six thousand files that is a sixty-five millisecond pause on every single turn. Each folder now keeps its newest thousand records. Nothing you can still reach is lost by that: the prompt reads the newest five, the Memory page shows thirty, and the search reaches four hundred, so long-term had quietly passed the point of being findable months ago. What was filed on purpose is never rotated out, however old it is: a fact you asked Skales to remember, a todo or calendar note you gave the pointer, a pinned record, and the closed records that let a fact show its history. In the long-term store only the overnight pass is ever trimmed. One line under the three counters says all of this on the page itself, and Skales knows it too, so asking whether it keeps everything for ever gets an honest answer.
-- **The interactive Playground is taking a break.** The mini-app builder is temporarily unavailable while it is reworked and will return in a later release. For now its sidebar entry, its Settings tab and its Add-Ons card are hidden, including for anyone who had it switched on.
-- **Skales IQ is described as free included usage, not a trial.** The Skales IQ copy now makes clear it is free included usage rather than a trial of a paid plan, with nothing to buy: when the included usage runs out you connect your own provider and everything keeps working. Added an explicit "this is not a subscription" line to the Skales IQ settings box.
-- **Flow places your attached images instead of redrawing them.** In a design or build project, an attached image is now described so the model knows what each file shows and embeds the actual file where it fits, instead of trying to recreate the picture in code. In an image or video project the attachment still passes straight to the generator untouched, so an edit works on your real picture.
+- **Building a prompt stopped re-reading the memory folder every time.** The list of recent notes was read fresh off the disk on every turn, and twice on a turn where the profile is withheld - once for the turn, once for the rebuild. It now reads through the same short-lived cache its sibling has used all along.
-## v12.2.0 - Freeze
+- **The sidebar opens with four entries you cannot lose, and five you can.** Dashboard, Chat, Code and Studio are the floor of the column now. Teams and Iris Orbit used to stand there too and could not be taken away, which is the wrong rank for a messenger a lot of people never open and a voice surface behind its own switch - both are ordinary rows now, and both can be sent into More like anything else. Nothing moves on a machine that is already set up: whatever your column looked like, it looks the same after the update, in the same order. A fresh install opens with Voice, Memory, Discover, Business and Skales Mobile under the four, and takes any of them away with one press. Custom Skills and Plugins are no longer handed out at the start - both are shelves for things you have built, both are empty on a new machine, and a row that opens onto nothing is how you learn that the rows are noise. They are one tap away under More, and whoever already had them keeps them.
-The first release after the freeze: images answer with the model you chose, the ChatGPT subscription works everywhere, and profiles keep their full guidance.
+- **Templates moved out of the column and into the System button.** The library of starting points and the Guide answer one question between them - what can I start from, and what can this machine do - and they stood a whole column apart. Templates is now a row in the same popup, directly above the Guide, and it still follows its own switch on the Add-Ons page.
-### Fixed
+- **Pointing Skales Code at a repository no longer throws you into a session.** Picking a folder with nothing typed opened a session on the spot, which took you off the landing before you had chosen a model, attached a file or picked an agent - off the one page whose whole job is making those choices. Picking now binds the folder and leaves you where you are: the chip names it, the branch appears beside it, and the @-menu lists the repository's files, all before a session exists. A press on the x lets the folder go again. The way in is the send button, or the "Open the session" button that appears beside the folder for the times you want the file tree and the terminal without typing anything first. Opening a folder from outside - `skales .`, a folder dropped on the dock, the folder's own context menu - still opens a session straight away, because that is a different request.
-- **Synced model profiles keep their full guidance.** An imported tuning profile's guidance text was cut at 600 characters on import, which silently truncated every profile in the shared library mid-sentence and dropped the tail of the instructions. The bound is now generous enough for a real multi-rule hint, and a profile that still exceeds it is logged instead of being trimmed in silence.
-- **Images keep your chosen chat model.** When a dedicated Vision Provider read an attached image, the answer was silently handed to a different model than the one you picked (a local vision model, or a small cloud fallback), so the reply looked like it came from the wrong place and, on paid providers, billed the wrong model. The picture now goes to the vision model only when a raw image is actually sent; once the Vision Provider has described it, your chosen chat model answers. Vision-capability detection now reads what the local model actually reports instead of guessing from its name, a Vision URL that already points at the chat endpoint is no longer mangled into a broken address, an explicitly chosen local vision model is attempted instead of pre-blocked, and the image-related tips and error messages now point at the Vision Provider settings. Profiles can also carry a vision-capability override so a new model is recognised without an app update. A picture and its description are now saved with the conversation, so the very next question ("what did you see?") is answered from what was actually read instead of "I do not remember", and a failed image read no longer vanishes from the history. The saved description shows under the image as a small collapsed eye accordion (like the reasoning view), never as raw text in the bubble.
-- **The ChatGPT (Codex) subscription is selectable everywhere.** A signed-in ChatGPT account never appeared in the model pickers (new-chat landing, in-session switcher, picker search, agent and Codework provider dropdowns) because those lists only recognised API-key credentials; the subscription sign-in now counts as configured and the Codex models show up like any other provider's. Flow already offered it.
-- **Every surface reads image capability from one place.** Telegram kept its own list of which models can see, with the multimodal gemma3 and gemma4 families wrongly marked text-only, and several surfaces disagreed about what counts as a local vision endpoint. Chat, Telegram, WhatsApp and the desktop buddy now decide this the same way, a pasted address is tidied up first, and the buddy no longer insists on a vision model id where every other surface accepts the built-in default.
-- **The ChatGPT (Codex) subscription is on one tested path.** The plain-chat and tool-using code for a signed-in ChatGPT account had drifted into two hand-copied copies, which is how the subscription kept breaking one release at a time; both now share a single request builder and stream parser covered by tests, so a fix can no longer land in one and miss the other. The current model list gained GPT-5.5 (an up-to-date id was falling through to the manual entry field), and a short-lived access token is now refreshed automatically before a turn instead of leaving a still "signed in" account failing every message until a manual sign-out and sign-in.
+- **A session started from the Skales Code landing keeps what you set there.** The model you picked, the effort you dialled, the mode, the agent, the extensions, the skills you attached and the files you dropped were all read from a stale copy the moment the session opened: the model was never pinned, so the composer went on naming it while the run used the settings default; the effort quietly fell back to medium; and anything staged without a message was cleared. All of it is now written onto the session before it opens, and read back from there - so the phone opening the same session finds the same composer. If the model could not be pinned for any reason, it is no longer shown as pinned either.
-## v12.1.1 - Freeze
+- **Every question Skales asks before doing something now looks and reads the same.** The permission prompt, the offer to install a language checker, the offer to import another editor's MCP servers, the chat's approval card and Flow's all asked the same four things - allow this once, allow it for this session, allow it always, refuse - in five different shapes, and in three different orders: the refusal was last in the Code window, on the right in the chat and on the LEFT on the phone. They are now one card with one order everywhere: one answer per row, full width, the narrowest yes first and the refusal last and set apart, and each answer carries a plain sentence inside it saying how far it reaches and for how long. The language-checker card, which used to offer four bare words, says it too.
-Long-running tools, portable backups, and Flow polish.
+- **The Skales Code window's title bar no longer breaks apart on a narrow window.** A session title runs to sixty characters and the bar is a fixed 44 pixels that does not wrap, so on a phone-width or portrait-tablet window the title spilled out of the chrome. It is now cut with an ellipsis at every width; below 720 pixels the repository and branch step aside, and below 560 the title does too - both are still on screen, in the status bar at the foot and in the history column.
-### Added
+- **Wrapped is opened from Discover, not from the sidebar.** The weekly card draws its week from the feed and its Sparks, and a row of its own in the Main list was a door standing away from the thing it belongs to. The row is gone; a Wrapped button stands in the Discover header beside Spark, follows the add-on switch like the row did, and the palette still finds the page.
-- **Send several images in one message.** Pasting, dragging or picking multiple pictures used to keep only the first one; now up to five images ride along per message, on the new-chat landing as well as in a running session. Each queued image shows its own preview with its own remove button, a shared window capture joins the same queue, the model receives all pictures in one turn (a dedicated Vision Provider reads each one and hands over numbered descriptions), all of them are saved with the conversation and shown in the bubble, and editing a message that carried several pictures resends all of them.
-- **Discover marks mobile posts.** A post made from the Skales mobile app shows a small phone icon next to its timestamp, matching the marker in the mobile feed. Under the hood the device bridge gained session control and a tool catalog for the next mobile release; everything is additive, a current mobile app is unaffected.
-- **Wrapped share cards get more looks.** The weekly recap card now offers nine themes (Skales, Obsidian, Snowfield, Neon, Sunset, Rose, Aurora, Grape and Teal), each in a light and a dark variant, chosen from a swatch picker right above the card. The choice is remembered for next week, and the download, copy and share-to-Discover buttons all export the themed card.
+- **The chat sidebar's More block holds History and the Cockpit, and both can be pinned.** Group Chat, Browser, Playbooks and Add-Ons stood there a second time, opening as windows over the app, while the same four are rows of the Home column that open as pages; two doors with two behaviours for one place. The block now holds the two surfaces you reach for between conversations rather than during one - History searches what was said, the Cockpit shows the work that came out of it - and either can be pinned into the column, where it stands under Agents. The four others are reached where they live. Group Chat carries speech bubbles instead of a second pair of shoulders, so it is told apart from Teams at a glance.
-### Changed
+- **A More row with nothing behind it is not drawn.** Pin every row out of the block, in either column, and More stayed - a chevron that opened an empty space and said nothing about anything. Both sidebars now leave the row out while it hides nothing, and bring it back the moment something is unpinned or a plugin, a widget or the DevKit gives it something to hold again.
-- **Flow designs with a stance now.** Every authored mode carries a design essence on top of its structural seed: commit to one aesthetic direction instead of a safe generic middle, name the system (layout family, fonts, background rhythm) before building, let typography carry the hierarchy, give each artifact one hero moment, treat whitespace as material, and make requested variants genuinely divergent. Each mode adds its own craft posture - slides think in sequence rhythm, motion thinks in beats with anticipation and one reveal, mockups respect platform conventions and touch targets, wireframes stay deliberately negotiable, documents read as typography with structure - each with concrete taste anchors. Real photography from Unsplash and Pexels joins the asset sourcing (downloaded into the project, never hotlinked), and a design that wants more typographic character may swap the seed's font tokens for a characterful pairing; motion compositions stay fully offline.
+- **Playbooks has no menu entry any more.** A playbook is recorded and replayed where the browsing happens - from the chat and inside the Browser surface - so a row in the Tools list was a door to a workbench nobody navigates to. The row is gone. Nothing else is: the page answers at /playbooks, the palette finds it, the add-on switch is where it was, and run_playbook and list_playbooks work from any chat.
-### Fixed
+- **The sidebar row that opens Iris Orbit is called Voice.** The row read "Iris Orbit" in every language, a product name where every other row says what it does; a person looking for the place to talk to Skales did not find the word. The row is now Voice, in all twelve languages, and the window it opens keeps its name.
-- **Windows commands run in the right PowerShell.** Agent commands on Windows were hardwired to `powershell.exe`, which is always Windows PowerShell 5.1 - a shell that rejects `&&`/`||` chaining, so every chained command a model wrote failed and had to be retried in single steps, burning time and tokens. Skales now runs agent commands in PowerShell 7 (`pwsh`) whenever it is installed and only falls back to 5.1 without it; on top of that, the agent is told exactly which shell it is in (in chat, Codework and the Telegram bot alike), so even on 5.1 it writes `;`-separated PowerShell commands correctly on the first try instead of learning it from errors. Linux setups also get their own platform hint instead of borrowing the macOS one, and the batch-folder shortcut in the prompt now shows Windows syntax on Windows.
-- **Discover links that contain a `#` are clickable again.** A shared or auto-posted link whose URL had a fragment (the `#section` part, e.g. a Wikipedia section or an app deep link) had that fragment mistaken for a Discover hashtag: it was pulled out and highlighted as a tag, leaving the rest of the URL as dead, unlinked text. Feed post text now recognises URLs first, so the whole link stays one clickable link (opening in your browser) and only real standalone `#hashtags` are highlighted.
-- **Backups no longer lose FTP passwords or stored secrets across machines.** The portable-secrets export covered the settings and connector keys but missed the two stores that encrypt in a different format (the internal secret store and saved FTP profile passwords); those travelled as the source machine's ciphertext and, with the encryption key deliberately kept local, silently vanished on restore to another machine. They now travel in the same portable form and are re-encrypted with the receiving machine's key; anything genuinely unrecoverable is reported instead of dropped.
-- **The 1:1 Wrapped card is square again.** The share card labelled 1:1 was rendering as a cropped portrait: sized as a flex item, it was squeezed narrower than intended while its height stayed fixed, so the content overflowed and the corners were cut off, and the exported image inherited the same wrong shape. The card now holds its exact intrinsic size, the on-screen preview is scaled separately to fit the window, and the content is measured and scaled to fit the box, so 1:1 exports a true 1080x1080 square and 9:16 a true 1080x1920, with nothing clipped in any language.
-- **Custom skills can be shared to Discover again.** The Skill AI generator added `require('fs')` and `require('path')` at the top of every skill even when the skill never used them, and the share screen rejects any skill that touches the file system. A share now strips those unused requires first (a require that is actually used still stays and is still screened), so a pure-compute skill is no longer blocked by dead boilerplate, and newly generated skills omit the unused imports.
-- **Detect Ollama knows about Ollama Cloud.** With Ollama Cloud switched on, the Detect button and the failure messages used to tell you to start a local daemon (`ollama serve`), which is nonsense for the hosted service. They now point at your Ollama Cloud key and connection instead, and the provider is simply labelled "Ollama" (not "Ollama (Local)") since it can be either.
-- **The Telegram bot names the real problem.** A network drop while polling (DNS failure, reset, timeout) was logged as "bot may have an invalid token", sending you to check a token that was fine. Network errors are now labelled as network errors (and recover on their own), and only a genuine rejection from Telegram is reported as a token problem.
-- **API connectors that need two query credentials work now.** Some APIs authenticate with a public application id next to the secret key, both in the query string (Adzuna's `app_id` + `app_key` is the common one). The connector could only carry a single credential, so these always failed. A connector now has an "Extra query params" field for the public companion (sent in the clear, never encrypted, kept out of the model context), the secret query parameter's name is editable instead of a fixed `api_key`, and the docs importer recognises the two-parameter pattern automatically. On top of that, saving a connector with an auth scheme the request layer does not understand is now rejected with a clear message instead of silently sending an unauthenticated request that reads back as the API rejecting the key.
-- **MCP tools get the time they need.** Every MCP request was cut off after a blanket 30 seconds, which aborted legitimately long operations (video analysis, renders, large scrapes) in the middle of their work. A tool call now has five minutes by default, each server can raise or lower that on its own up to 30 minutes, and the quick control requests such as connecting and listing tools keep their short timeout so a hung server is still detected fast.
-- **Ollama reads images reliably, including Ollama Cloud.** Pictures sent to Ollama went through the OpenAI-compatible endpoint, which the hosted vision models answer with a server error; the same models describe the same image fine on Ollama's native chat endpoint. Skales now routes image turns over the native endpoint automatically - no setting, no URL to change - and text conversations keep exactly the path they had.
-- **Claude over OpenRouter stopped re-paying its own prompt.** The direct Anthropic connection has long cached the static prompt prefix (tool schemas plus system prompt), but the same Claude model routed through OpenRouter sent no cache marker and re-billed that whole prefix on every turn. Claude models through OpenRouter now reuse their cached prompt the same way the direct connection does, so long conversations cost what they should; the volatile clock line stays outside the cached span so the cache survives minute boundaries, and every other vendor's request is untouched.
-- **The OpenRouter speech model is actually spoken through now.** The provider card has offered a speech (TTS) model slot, but no read-aloud surface read it: chat read-aloud fell back to the system voice and voice notes only ever used the dedicated voice providers. With an OpenRouter key and a speech model configured, read-aloud in chat, WhatsApp and Telegram voice notes, Studio voice and the AIPointer overlay speak through OpenRouter first; an explicitly chosen voice provider (ElevenLabs, Azure, OpenAI, local) still wins.
-- **Settings backups travel between machines.** Exports now carry secrets in a portable form while the local encryption key never leaves the machine; imports re-encrypt everything with the receiving machine's key instead of overwriting it, older backups migrate automatically, and anything unrecoverable is reported instead of silently dropped.
-- **Thinking stays out of the chat.** Reasoning models (the DeepSeek, Qwen, GLM and MiniMax families among them) that spent a turn thinking could get their entire thinking transcript posted as the visible answer, in Flow a wall of inner monologue instead of the design file. The trace now stays on its own collapsed channel: a turn that burned its budget mid-think recovers with a re-request instead of dumping the transcript, an answer cut off inside its thinking block is treated as thinking, and the Flow live preview hides inline thinking while the model works. Plain helper calls on local reasoning-only models keep working as before.
-- **The Flow window lost its redundant close button.** Flow runs in its own window, so the corner X only duplicated the window's own close. The project-files button moved into the preview and editor toolbars as the first icon of one right-aligned row next to the other actions.
-- **The Flow composer got wider.** The attach button now sits above the send button in one column, so the message field takes the full remaining width.
-- **The Start button looks like a button again.** Inside the Flow window it rendered without its styling; it is now the same accent icon button as the chat send button, and the scoping-questions Continue and Brand-Kit create buttons got the same treatment.
-- **Editing a message keeps its picture.** Editing a question that carried an image resent only the new text: the image was silently dropped and the model answered blind. The edit now re-attaches the original picture from its saved copy (with the in-bubble thumbnail as the fallback), so a reworded question about an image still sees the image.
-- **The new-chat page remembers your agent.** The agent picker on the chat start page reset to Skales on every visit, so anyone who mostly chats with one custom agent had to reselect it each time. The picker now remembers the last choice for the next visit. Skales stays the default character: nothing is stored until you pick someone else, picking Skales again clears the memory, and a deleted agent falls back to Skales.
-- **Approving a phone-driven action replies to the right message.** When a chat run from the Skales mobile app pauses on a tool that needs confirmation (sending an email, pushing to git, running a command), the desktop asks the phone to approve or reject and, once you decide, sends the result back. That reply now carries the conversation's id through the whole approve-reject round-trip - the initial request, any follow-up approval, the final answer and the error path - so the phone pairs it with the exact message that was waiting instead of guessing by order, and a reply missed while the phone was briefly offline still settles cleanly. Local and non-mobile use are untouched.
-
-## v12.1.0 - Flow
-
-Design by conversation. Studio gets a new front door.
+- **Organization is opened from the chat, not from a menu.** A team of your own agents working one job is something you put to work from the conversation, and its row in the Tools list and in the chat column's More block was two doors to a place the chat already reaches. Both rows are gone; typing /organization opens it, the command stands in the slash menu with the others, and the palette still finds the page. An add-on that is switched off is not opened blind: the command names the switch instead. The Add-Ons card says where the door is now rather than marking the surface as parked.
-### Added
+- **The line under the wordmark at start changes.** The splash screen carried one fixed line. It now draws one of twelve at every start, from a list the phone and the computer share byte for byte, and the fixed line stays as the fallback when the list cannot be read.
-- **Flow, the conversational design workspace.** Open Studio and describe what you want; the Skales agent designs it as real files, with a live preview, the files and the code sitting next to the conversation. Eight modes cover the ground: slide decks, interactive prototypes, wireframes, mobile app mockups, print documents, generated images, generated videos, and motion graphics that render to a real MP4. Every mode carries its own design discipline, so the first result already looks deliberate instead of improvised. Flow is a beta.
-- **Flow opens in its own window.** In the desktop app, Flow gets its own window so you keep working in chat and the other tools while a design generates. In the browser it stays an in-app overlay.
-- **A composer that carries real controls.** Attach up to ten files (documents like PDFs become content sources the agent reads, not decoration), reference an earlier Flow project, pick a Brand Kit, pick a template, and choose the model and reasoning effort per project. The model picker searches the same live catalogs as the chat page, favourites and recents included.
-- **Brand Kits drive the design, including what to avoid.** The Brand Kit gained fields for what the brand is, reference links the agent studies before designing, free design notes, and explicit bans: fonts, design directions and anything else that must never appear. Activating a kit in Flow makes the palette, the typography and the bans binding for every artifact, and a toggle makes the logo and uploaded brand assets a requirement, not a suggestion. New kits can be created right inside Flow; the settings page remains the full editor for the default kit.
-- **Templates that shape the output, not just the prompt.** Picking a template (pitch deck, invoice, dashboard, logo reveal, ...) binds the right mode and injects a structural directive into every turn of the project, so a pitch deck follows a pitch deck's arc even five edits later.
-- **Media generation with a choice of engine.** Image and video modes list every backend that is actually usable right now: connected MCP media servers, the configured cloud providers, local engines, and Skales Visuals, the built-in engine that designs images and videos itself, at the exact format you asked for, no external image model needed. Pin one, set aspect, resolution, quality and duration, and optionally name the model you want.
-- **Generated images can be edited by talking.** Ask for a change after an image lands ("make the keyboard white") and the agent treats it as an edit of the existing file, through the image-edit tool or the media server's image-to-image, saving each revision as a new file.
-- **Skills and connectors answer to "@" in Flow.** Typing "@" in a Flow composer lists your imported agent skills and connected MCP servers; a mention activates the skill or steers the turn to that server, the same way the chat composer does it.
-- **Flow artifacts land in the shared gallery, and images can go to Discover.** Generated media is mirrored into the Studio gallery automatically, every turn shows a card of the files it produced, and any image artifact can be submitted to the Discover feed, compressed and reviewed before it appears.
-- **Discover can now carry video posts.** The feed renders link-based videos with a muted inline player that respects the clip's format, moderated the same way as images.
-- **OpenRouter carries every modality now.** One OpenRouter key no longer means text only: the provider card gained per-capability model pickers, fetched live from the catalog and filtered to what each capability actually accepts, for image generation, video generation, music, speech (TTS), transcription (STT) and embeddings. Image models join Studio and Flow as a first-class engine, video models generate through the async job API with honest progress and errors, voice input can transcribe over OpenRouter, read-aloud can speak through a dedicated speech model with a voice of your choice, and the semantic document index can embed through OpenRouter as well.
-- **Skales IQ inherits the same breadth, tier by tier.** Skales IQ now understands every modality (chat, vision, image, video, music, speech, transcription, embeddings), and what your plan can do is managed per tier and can grow without an app update. A capability that is off or not in your tier answers with a clear message in chat instead of a cryptic failure, the same way an offline Skales IQ already does.
-- **Drag and drop lands in ongoing chats.** Dropping a file onto the composer of a running conversation now attaches it, PDFs through text extraction, images to the vision slot, text inline and archives to the Workspace, exactly like the paperclip. The starting page accepts PDFs and archives too, and a PDF that cannot be parsed still attaches with a Workspace copy instead of being skipped.
-- **Flow asks before it builds.** On the first turn of a project, when the brief leaves essential decisions open, the agent poses a handful of tailored scoping questions rendered as a clickable form in the preview pane: option chips, multi-select where several picks make sense, a free-text field and "Decide for me" per question. Continue sends the picks back and the artifact is built to them; a clear brief skips the questions entirely.
-- **A template gallery on the Flow home.** All twelve templates fan out as cards under the composer; hovering one previews its brief right in the prompt box, clicking binds the matching mode and structure, and a blank-project link sits below.
-- **Change course mid-project.** The Flow workspace header carries a small toolbar: switch the model (with live catalog search) or switch and deactivate the Brand Kit for the next turn, without leaving the project. The workspace composer gained an attach button too, so fonts, images and archives can join a running project, and a live token counter shows what a run actually costs.
-- **Skales unpacks zip archives.** A new extract tool unzips archives natively (no shell), guarded against path escapes, available to the agent in chat and everywhere else. A zip attached in Flow is unpacked automatically into its own assets folder.
-- **Ollama Cloud, one switch away.** The Ollama card gained a Cloud toggle in its own row: flip it, paste your ollama.com account key, and Fetch Models fills the regular model list with your hosted models, no local install involved. Detection doubles as a cloud connection check, the local-install sections (setup, marketplace, the local tool cap) step aside in Cloud mode, and the hosted frontier-size models are treated as the cloud models they are, with none of the local limits.
-- **Your OpenRouter choice leads.** When OpenRouter is the active provider or one of its modality models is pinned, image generation, voice transcription and read-aloud speech go through OpenRouter first; the other configured providers remain fallbacks.
+- **The licence names its licensor at the end, not in its first lines.** The agreement opened with a full postal address, telephone number and e-mail address in its fourth line, before it had said a single word about the software. The address, the telephone number and the e-mail address now stand together in a section of their own at the end, section 12, where a reader who needs to write to us finds them; the first lines say who the licensor is and point there. The English and the German text changed in the same way, and nothing else in the agreement moved.
-### Changed
+- **The licence screen says "du" in every language it speaks.** Skales addresses one reader, informally, and English and German did - but the licence screen changed to the formal voice the moment it was read in Spanish, French, Croatian, Turkish, Russian or Chinese, so the one screen everybody has to read was the one screen that spoke like a form letter. All twelve languages now use the same voice there: pronouns, verbs and imperatives together, not half a sentence at a time. The computer and the phone also say each of those sentences with the same words again - the line explaining that the binding text is in English had drifted into two versions, formal on one machine and informal on the other. The agreement itself is untouched and keeps its legal register, which is a different thing from how a screen talks to you.
-- **The classic Studio surfaces moved behind one "Studio v1" group.** Design, Media, Audio, Type, Scenes and Gallery keep working exactly as before, folded into a collapsible rail entry below Flow while Flow is the front door.
-- **The Studio landing for Flow is honest about readiness.** One column: the checks on top with a wide open button, the explanation below. Configured MCP servers count as ready (they connect on first use), and the media row mirrors, key for key, what the image and video pipelines actually accept.
+- **A chat that is working keeps saying so, all the way to the last step.** The conversation decided whether anything had changed by looking at a handful of facts about the last stored message - how many there were, whose it was, how long it was, when it arrived. Everything a running turn knows that is not in that sample moved without the screen noticing: which step it is on, which tool is open, what its helpers are doing, what the browser is looking at. Measured over one ordinary turn with tools, seven polls out of thirteen carried a change nobody looked at - which is how an answer from an earlier turn could end up standing over a picture you had just re-sent, with no thinking panel under it. The run now carries its own version number, and the chat redraws when that number moves and at no other time; it also notices when something else has replaced what is on screen and puts its own view back. The working indicator hangs on the run as well: it used to leave at the first token and never return, so a turn with eleven more steps to go looked finished for all of them. It now steps aside while an answer is being typed out and comes back for every further step, and a thinking panel you opened stays open across the whole turn. The Skales Code window asks the same question of the same field, so the two windows cannot disagree about the same piece of work.
-### Fixed
+- **Skales Desktop now runs on a newer generation of its web engine.** The
+ ground the app stands on was a whole generation behind, which meant every
+ release carried weaknesses that are no longer repaired at the source. Moving
+ up puts the app back on a foundation that is still maintained and still gets
+ security fixes, and it changes how pages and requests are handed around
+ inside the app. What you see and what you click stays exactly where it was:
+ every surface, from Chat and Code to Studio, Flow, Iris and Settings, was
+ re-checked against the new foundation before this release.
-- **Vision Provider with OpenRouter returned a 404.** OpenRouter needs vendor-prefixed model ids, and the vision path sent the bare id the settings carried. Bare ids of known families get their prefix automatically now, switching the vision provider prefills that provider's known-good default model instead of carrying the previous one along, and the model field shows a provider-aware example.
-- **A text-only local model no longer silently becomes llava.** When the active Ollama model cannot see images, Skales now looks at which vision models are actually installed and uses one of those; with none installed it says exactly that, with the pull command, instead of failing with a confusing missing-model error. Detection also recognizes cloud-suffixed and differently-spelled tags of the multimodal families.
-- **A #trigger routes to its API connector, deterministically.** Typing a connector trigger in chat now pins the turn to that connector instead of hoping the model picks it over web search or an MCP tool, a trigger saved with uppercase resolves too, and a connector whose docs yielded no endpoints says so in Settings instead of failing quietly.
-- **The command highlight keeps its place.** When the colored command overlay appears mid-draft in a scrolled composer, it aligns to the text in the same frame instead of showing the caret on the wrong line, in the chat session and on the starting page.
-- **The reasoning-effort tooltip is shorter.** The dial says what level is active without the click-to-cycle lecture.
-- **The per-modality model pickers actually fetch now.** A key-handling fault made every fetch quietly fall back to a static text list, so the pickers looked empty. Fetching works now, the public catalog even loads without a key, and the pickers moved from a native browser datalist onto the app's own dropdown, loading their catalogs automatically when the panel opens.
-- **PDF attachments work in the packaged app, not just in development.** Packaged builds shipped without the PDF text extractor, so every extraction failed silently while development worked fine. The packaged app now carries it.
-- **Long generations are never cut off anymore.** The per-request timeout used to stay attached to a streaming response, so a model that was actively writing a large file for minutes was aborted mid-work with a timeout error. The timeout now only bounds the wait for a response to begin; once a model streams, only genuine silence on a dead connection ends a run, and the Stop button remains yours.
-- **The Flow chat can be read while the agent works.** Auto-scroll only follows the newest message when the view is already at the bottom, instead of fighting an upward scroll.
-- **Fixed-format compositions fit the preview.** A 1920x1080 motion graphic (or a 9:16 reel) scales to fit the pane instead of running off screen, and Motion gained format and duration controls that the composition and the renderer respect.
-- **Flow projects stay out of the chat history.** Flow sessions no longer appear as regular chats in the sidebar, History or the command palette; they reopen from Flow's own project grid.
-- **Small but annoying, fixed in Flow.** The code editor no longer collapses to a strip, a long model name no longer pushes Start to the next line, cloud-generated images and videos show real thumbnails in the project grid, the mode chip rail lost its stray dark box, and the Flow window in the desktop app wears the same hidden-inset chrome as the main window.
-- **Flow works on a narrow screen.** Opened from a phone browser (Tailscale on the go), the workspace stacks chat above the live preview instead of hiding the preview, dropdown menus clamp to the viewport, and the template gallery wraps into rows. On desktop, the gallery breaks out of the column so no card is cut off and the hover lift is never clipped, and the mode chips shrink to fit one line.
-- **Picking a template no longer types into your prompt.** A chosen template shows its brief as the placeholder (exactly what hovering previews) instead of inserting text that survived switching or clearing the template; with the box left empty, Start uses the template's brief as is.
-- **LLM Profiles no longer flatten Flow designs.** Profiles keep weaker models precise while they operate tools, which is right for ordinary work but wrong for design: DeepSeek and MiniMax were composing Flow layouts in that precise mode, and the results came out flat and generic. Flow now keeps the model's full creative range while it designs; everything else stays precise.
-- **Agent-written documents reach the Document panel again, proactively.** Documents the agent wrote were invisible to the panel, a fresh chat never told the model the panel existed, and weaker models wrote documents as plain chat text instead. All three fixed: ask for a summary, article or report in any new chat and it lands in the panel, opens rendered on a wide window, and lights the header dot on a narrow one.
-
-## v12.0.0 - Solid
-
-Two steps ahead. (This entry covers the work landed so far; the release is still in build.)
+- **A picture the agent goes back to now says it is the same picture.** A run
+ that takes a screenshot and later puts it back on screen drew it twice, with
+ nothing to tell the two apart, so a long run read as if the agent had looked
+ again when it had only shown you what it already had. The second one keeps
+ its picture, because taking it away to save a few lines would hide that the
+ agent went back to it, and carries a small "Shown again" chip beside the one
+ that says it came from a tool. Two real looks at a screen that did not move
+ are still two looks, and are never labelled that way, however identical they
+ are.
-### Added
+### Removed
-- **A format for role bundles.** Skales now has a defined format for a role bundle: one folder that carries the skills, connectors and slash-commands for a job role, with a loader that validates it before anything uses it. A bundle is only sound if every connector it ships is reachable through a skill or command, so a role arrives complete instead of as a connector with no skill that uses it. There is no install screen yet; this is the groundwork, with an example bundle in the developer kit.
-- **Skales learns from your finished work, and the learning is yours.** When a goal hits friction and finds a way through, Skales keeps what worked as a short, reusable approach and notes what to avoid, so the next job of that kind starts ahead instead of from scratch. The approaches it leans on get sharper the more you work, since it favours the ones that actually reach a good result and quietly drops the ones that do not, and a quick "that's perfect" or "that was wrong" on a finished goal counts. The memory page shows each lesson with a quality mark, lets you pin the ones you rely on, and delete anything you do not want kept. It is a plain file on your computer that you own and can read.
-- **Tell Skales to remember.** Say "remember this" or "from now on...", and it keeps that as a first-class note that surfaces first the next time it is relevant.
-- **Skales stays a step ahead.** It surfaces a meeting that is about to start or a scheduled task that did not run, quietly prepares for a meeting that has an agenda before you ask, and gathers everything else into one calm daily briefing instead of a stream of pings. The memory page shows what it got ahead of and why, and you can mute it in notification settings.
-- **Skales names your chats for you.** A new conversation gets a short, specific title drawn from its first exchange, the way you would name it yourself, instead of the first line of your message. The placeholder appears at once and is replaced when the title is ready.
-- **Show an image you already have.** Skales can display an existing image file from your workspace right in the chat, so a picture you or a connected service saved appears inline instead of being described or pasted as raw data.
-- **A dot shows when a document is waiting.** The document button in the chat header now carries a small mark whenever the current chat has a document but the panel is closed, so an artifact Skales wrote, or one in a chat you reopened, is easy to find instead of hidden behind a closed panel.
-- **Delete your Skales IQ trial data yourself.** A new control in Settings, under Skales IQ, removes your trial record from our server and switches off its key in one click. The signup now also says plainly why your email is kept, to prevent abuse and to offer you a later upgrade or credit, and that you can remove it whenever you want.
-- **Codework templates are back in the library.** The Templates page has a Codework category again, with ready-made coding tasks, fixing a bug, writing a test suite, refactoring, scaffolding an API endpoint and more, that open straight in Codework with the task already filled in.
-- **Find your real models, not a fixed list.** The model switcher now shows the models you actually have. Your favourites and recently used sit at the top, and a search field finds any model across every provider you use, your real OpenRouter, Gemini, Anthropic and Ollama models, instead of a short hardcoded lineup. A star keeps any model one tap away, the same in the in-chat switcher and on the new-chat screen, and a local model server that is not running drops out instead of listing dead entries.
-- **Pick who answers and which model before you start.** The new-chat screen has its own row under the box for the agent and the model. Choosing an agent fills in its model for you, and you can switch to any other model just for this chat without changing your default. The choice sticks to that conversation and is still there when you reopen it, and a later switch inside the chat sticks too.
-- **A new in-between coding mode for Chat.** Between Code, which asks before each edit, and Auto, which works on its own, there is now an Edits mode: it approves your file edits as it goes but still asks first before a shell command, a git push or a deploy. Pick it from the same Chat, Code, Plan, Auto strip.
-- **Plan mode hands you the plan to approve.** A folder-bound chat in Plan mode now writes out its finished plan and waits for you to approve it before anything runs, instead of switching to Code and starting on its own. You read the exact plan, then approve to build it.
-- **Choose how much Codework may touch.** A new file-access setting runs a project read-only, lets it edit, or lets it edit but never delete. Deleting a file always needs an explicit yes, whatever the setting.
-- **Chat reads your project's own rules.** A folder-bound chat in a coding mode now reads an AGENTS.md (or CLAUDE.md) at the folder root and follows it, so your project's style, test runner and conventions carry into every reply without you repeating them.
-- **Codework remembers your project across runs.** Before it starts, Codework reads an AGENTS.md (or CLAUDE.md), memories.md or branding.md at the project root and follows it, so its work keeps to your conventions and tone every time instead of relearning them.
-- **Codework tells you when it is done.** A finished or interrupted Codework task now raises the same toast, chime and sidebar mark as a chat, so a task you left running reaches you even when you are on another page.
-- **Your activity recap covers the tools you actually reach for.** Using an MCP tool or a Hugging Face Space, creating a template, and exporting any Studio video (not only a text animation) now register in your Discover activity and your weekly Wrapped, instead of going unrecorded. Wrapped also keeps the model you leaned on most for the week and tells your Chat, Code, Plan and Auto time apart.
-- **A home you choose: Calm or Power.** The dashboard opens in one of two layouts and remembers which one you picked, so it looks the same the next time you open it. Calm keeps it personal: pick up where you left off, jump back into recent chats, what Skales has on its mind, your most active surfaces, and quick ways to start. Power is for getting work done: your active goals with their real progress, Autopilot status with this hour's API calls and anything waiting on your approval, your provider balance, your connections and paired devices, and live runtime memory, CPU and uptime. Your own avatar emoji greets you, the current weather sits under the greeting, and your weekly Wrapped badge rides in the header with a tap straight to the recap.
-- **Your briefing, in a single line under the greeting.** When you have joined a Briefing in Discover, its latest items rotate through one calm row and open in the built-in browser on click. When you have not joined yet, that same row invites you to, so the home screen is where a briefing begins.
-- **A balance you can trust, or nothing at all.** Running on Skales IQ or OpenRouter, the Power view shows your real remaining balance with a gauge. Skales IQ shows the percentage of your trial left, exactly like the Settings box; OpenRouter shows your remaining credit in dollars. For a subscription or a key that exposes no balance, the card simply does not appear, rather than inventing a percentage.
-- **Your week, at the top of the home screen.** A seven-day strip under the greeting lays out the current week, with the events and planner tasks for each day, and today marked. Empty days stay empty rather than guessing. It is there in both Calm and Power, and any day opens the Planner.
-- **Pick up where you left off, in both views.** The one-click way back to your last session now sits at the top of Power too, not just Calm, so getting back to work never means hunting through History.
-- **Real runtime health, measured not guessed.** The Power view reads the backend's actual memory, CPU and uptime live and labels them as the runtime's own, instead of the browser-heap placeholder the old system tile showed. Nothing on the home screen is a stand-in number.
-- **Home widgets that all show real data.** The optional-widget strip gains your latest Studio work, your running goals, the next scheduled goal, your latest Briefing items, your Wrapped badge, your online team devices, and your most-used surfaces. The empty Quick Chat box, the duplicate memory cloud, the buddy-mood tile and the placeholder system tile are gone.
-- **Choose how deeply the model thinks, per chat.** A small dial next to the message box steps through four levels of reasoning effort, and its circle fills and shifts colour as you go up: an empty yellow-green ring for the lightest, then half blue, a full orange, and a full red for the deepest. The dial follows the model you have picked. On models with a real reasoning control, Claude Sonnet 5 and the newer Opus, and reasoning-capable models through OpenRouter, it drives that control directly. On models that have none, your local models, custom endpoints and most direct providers, it shows a muted "none" ring and steps aside, so it never claims a depth the model cannot deliver and the message box never shifts when you switch models. The level stays with that conversation, and the dial is on the new-chat screen too, so a level you pick before the first message rides into the session. Once you raise it, it takes over from the global deep-reasoning switch for that chat; a fresh chat, or the dial back at its lightest, follows your global setting.
-- **Claude Sonnet 5 is ready to use.** Sonnet 5 is in the Anthropic model list and runs with its adaptive thinking on by default, with its output length and long-conversation compaction sized for it, so a long answer no longer cuts off early.
+- **The install count, the monthly-active count and the star count.** The install figures came from the public aggregate's `installs_lifetime` and `installs_30d`, which are the count of DISTINCT IP ADDRESSES that asked the update endpoint for a version - the same source reports 639 devices active in thirty days. A figure nobody can explain when asked what it measures does not belong on a page that asks a company to sign something. What is left is countable: tools from the generated manifest, providers and models from the aggregate, countries, and languages from the locale files.
-### Changed
+- **The Revenue Dashboard tile and the "lead scoring" claim.** Neither exists anywhere in the tree. The usage dashboard arrives with Skales Business and is listed there instead.
-- **Trivial turns feel instant.** A greeting, a thanks or a quick "ok" skips the heavy setup a real task needs and answers right away, with no thinking card sitting above a one-word reply.
-- **Deep reasoning reaches more of your work, and stops wasting itself.** With the xhigh deep-reasoning boost turned on, a Codework run now plans as carefully as a chat or a goal already did, so a coding job gets the same care. And a greeting or a quick thanks no longer triggers a full reasoning pass, so those pleasantries stay instant.
-- **The reply box frees up the moment your answer lands**, instead of waiting on background bookkeeping, so you can keep typing without a pause.
-- **A tidier message box.** The magnifying-glass and slash buttons are gone from the message box. Both only opened things you already reach by typing "/", which lists every command as you type, so the reasoning-effort dial takes the slash button's old spot. The small labels on the message-box buttons now appear the instant you hover instead of after a pause.
-- **A document Skales writes opens ready to read.** It opens rendered, the finished look, instead of as raw markdown; switch to the editor when you want to change it.
-- **It feels like a native app, not a web page.** The pointer is the system arrow over buttons and links instead of the web hand, and every dropdown is a real, keyboard-friendly menu that matches Skales rather than the raw operating-system control.
-- **Clearer empty and error screens.** When a list has nothing in it yet or something fails to load, Skales shows a short explanation and a way forward, like try again, adjust filters or start a chat, instead of a bare spinner or a raw error line.
-- **Every tile on the home screen goes somewhere.** Each card, stat and list on the dashboard now opens the thing it stands for, a chat, a goal, a memory, a setting, the schedule, the briefing, instead of sitting there as text. The old capability grid that led nowhere is gone.
-- **Your recap counts the work you do outside chat too.** Finishing a Codework project, running an Organization, holding a Group Chat, running a custom skill, and rendering a Studio video now register in your weekly Wrapped and the dashboard's most-active card, the same way in-chat work already did, so the picture of how you use Skales is no longer chat-only.
-- **A cleaner, more human voice.** Replies avoid em-dashes and marketing filler, both in how Skales is guided and in a final pass over the answer.
-- **A smaller, faster download on Mac and Linux.** The installer no longer carries build tooling it never runs, so it is meaningfully smaller and updates arrive faster.
-- **Background and team work runs leaner.** Unattended tasks, agents and team plans load only the tools a step needs instead of the whole set.
-- **The free trial no longer ties itself to usage analytics.** When you start Skales IQ, anonymous usage statistics stay on by default to help improve the product, but onboarding now has a clear switch to turn them off on the spot, and the trial starts either way. Choosing not to share them never blocks or changes your trial.
-- **The native, consistent look now reaches the rest of the app.** Settings, Chat, the sidebar and History share the same rounded cards and inputs, snappier motion that only animates what should move, the last raw dropdowns replaced by the same keyboard-friendly menu, and disabled controls that dim again, so the whole app feels of a piece. History's provider and channel filters sit on one tidy row with the refresh beside the conversation count, and the sidebar's bottom bar reads more calmly.
-- **The new-chat screen has more to say.** It opens with a wider, rotating set of lines, including a few meant to make you smile, instead of the same handful, and the input keeps a little breathing room when you focus it.
-- **Codework asks before writing by default.** New file writes wait for your approval until you say otherwise; turn auto-approve back on in the approval panel whenever you want the faster flow.
-- **Full disk access takes a deliberate second step.** The full-disk option in the Code access prompt now says plainly that it applies to every chat and stays on until you change it in Settings, and asks you to confirm once more, so a single tap can no longer open your whole disk everywhere. Granting just the one folder stays one tap and is the recommended choice.
-- **Codework follows the same folder-safety rules as the rest of Skales.** It uses the shared protected-paths list and re-checks the real location of a file before writing, so a project that tries to reach outside its own folder through a link is stopped.
-- **Codework works to your full step budget, not a fixed limit.** A run now follows your goal step-budget setting all the way (0 means run to completion, like a goal) instead of a hardcoded ceiling, so a real autonomous job is no longer cut short after a handful of steps; a stuck model is still stopped early by the progress guardrails.
-- **Codework's autonomy settings moved out of the header and onto the start screen.** You now set auto-approve, file access, preview and the test command before a run, on the start screen, the way a coding agent should. During a run the gear opens those same settings as a clean centered panel that matches the rest of Skales, instead of a strip wedged into the header that could sit on top of a pending approval.
-- **Codework's approval looks like the one you know from chat, sits where you are working, and waits for you.** A pending action now shows as the same clear approval card you see in chat, with Approve, Deny and Always approve this session. It is pinned to the bottom over the terminal, not jammed into the header, so the step it is asking about stays readable above it. "Always approve this session" now actually gives the run a free pass, so the rest of the run stops asking instead of prompting again on the next command. And it is patient: the approval waits 30 minutes instead of 5 and a run may take up to an hour, so pausing to read it or opening settings no longer aborts the task with a timeout.
-- **Codework reads and searches your files through the same engine as the rest of Skales.** Reading a file and searching the project in a Codework run now go through the shared tool path that Chat uses, so an improvement or fix to it reaches both at once and there is one set of file-safety rules to trust. Writing files, running commands, the file tree, diffs, the write preview and undo are unchanged, and a run still works only inside the folder you chose.
-- **Auto pauses before it reaches outside the work.** In Auto mode, an action that pushes to a remote, deploys, runs a destructive command, or writes outside the folder you bound now stops for a single approval instead of going through unattended. Ordinary edits and commands inside the folder still run without asking, so Auto stays fast where it is safe and checks in only where the stakes are higher.
-- **Codework shows an accurate diff, and you can open any file to read it.** The change view now lines up moved code correctly instead of marking every line after an inserted one as both removed and added, so a real edit reads cleanly. Clicking a file in the tree opens it in a read-only viewer, so you can check the code on the right without leaving the page. It stays a viewer, not an editor: Codework works by doing, not by hand-editing.
-- **A coding chat gets the same project map as Codework.** The structural index of your project now uses the same, larger budget in Chat as in Codework, instead of a tighter chat-only cap, so on a big codebase the model starts with more of the picture. When the map is still too large to fit, it says so and points the model at the rest to open on demand, rather than quietly handing over a partial map.
-- **A mistyped file mention tells you, instead of failing quietly.** In a coding chat, when you mention a file with @ that is not in the bound folder, Skales now says it was not attached, so a typo no longer looks like it was simply ignored. A bare word that is not a path, like a skill mention, is left alone.
-- **Safe commands stop asking, risky ones get a clearer prompt.** A short list of read-only and test commands, like npm test, git status and listing files, now runs without a confirmation, so a command you repeat all day stops nagging. This skips the prompt only inside a folder-bound coding session (Code, Edits or Auto on a folder you chose); a plain chat still confirms every command. Installing packages or pushing to a remote now asks with a message that says what it does, instead of a blanket "Run command?". A force-push that rewrites remote history, or a download piped straight into a shell, is refused outright with an explanation. The block on system-wrecking commands is unchanged and still cannot be turned off.
-- **Commands Skales runs no longer see your secrets.** When Codework or a coding chat runs a shell command, the command now runs with API keys, tokens and Skales' own internal values stripped from its environment, so a script in a project you are working in cannot read them straight out of memory and send them away. The command still runs only in the folder you chose, and new file writes still ask first unless you turn that off.
-- **A folder-bound run cannot read outside its folder through a link.** When a coding run reads a file in the folder you chose, it now checks the file's real location, so a link inside the project that points somewhere else can no longer be used to read a file outside the folder. The write side already did this; the read side now matches.
-- **Codework reads several files at once.** When Codework decides to read or search more than one file in a single step, it now does that reading in parallel, up to four at a time, instead of one after another, so the part where it gathers context to understand your code is quicker. Writing files, running commands and everything else stay one step at a time, in order. A step that mixes a read with a write or a command runs sequentially too, so a read never returns content from before a write in that same step.
-- **Your most-used tools live in the main menu now.** Workflow, Planner, Tasks and Wrapped moved up from Tools into the main sidebar, Discover sits right under History, and Codework moved into Tools, so the surfaces you reach for every day are one click away. The order is the same in every theme, and anything you have switched off in Add-Ons still stays out of the sidebar.
-- **Features that have settled no longer wear a Beta tag.** Teams, LLM Profiles, Agent Swarm, the Always-On Agent, Call Mode and memory Dreaming have been steady for a while, so their Beta labels are gone, in every language. The genuinely early ones, like the Skales IQ trial and the Business preview, still say so.
-- **The home screen fills in faster.** The dashboard now runs its status checks together instead of one after another, holds the weather and your goal and memory lists for a few seconds, and no longer scans the same files twice on a single open, so it settles quicker without going stale, and a memory you delete still disappears at once.
-- **The update screen puts its links last.** The social links now sit at the very bottom of the update page, below the updater's own log, so the log is easier to reach.
+- **"A growing library of business plugins - included."** It promised plugins that do not exist, in the row of things you get. The five that do ship - Open Invoices, Receipt Inbox, Notice Radar, Meeting Prep, Page Watch - have a section of their own now, on the free side of the line where they belong.
-### Fixed
+- **"...and nothing to bill"** from the Skales IQ line. Beside something a company pays for it read as "free", and it was not true.
-- **A connected tool cannot smuggle hidden instructions into Skales.** A tool from an MCP server is third-party, and its description is shown to the model, so a malicious or careless server could hide commands in invisible or control characters inside that text. Skales now strips those characters from every MCP tool description and caps its length, so a connected tool cannot quietly steer the model through text you never see.
-- **One misbehaving MCP tool no longer breaks the rest.** A tool that reports a malformed input schema is now coerced to a safe shape instead of breaking the model's tool-calling for the whole message, so a single bad server can no longer take the others down with it.
-- **Skales works with MCP servers on the newer protocol.** It now remembers the protocol version a server agrees on and sends it back on every later request, the way the current spec requires, so a stricter or newer server no longer rejects Skales' calls after connecting.
-- **The balance card shows a number you can place.** On OpenRouter the home screen now shows only your real remaining credit, without the confusing lifetime-total figure and gauge beside it that did not map to any budget you set.
-- **A stopped background engine comes back on its own.** If the Skales server stops while the app is open, for example because its background helper was force-quit (a common way to close Skales on Windows, where shutting the window only hides it to the tray), Skales now quietly restarts it and reloads you right where you were, instead of a red "stopped unexpectedly" box. A calm restart prompt appears only if the engine cannot stay running after several tries.
-- **On Windows, selecting Ollama no longer freezes the app or flashes a console window.** If Ollama is not installed Skales says so at once instead of opening a stray window and hanging for ten seconds on every check; when it is installed it starts quietly with no window.
-- **A model that got stuck loading its tools now just works.** Some models would keep announcing that a tool is "available next step" instead of using it; Skales now hands those models every tool directly and adds a switch under Settings, Advanced to turn on-demand tool loading off if you ever run into it.
-- **The memory page reads cleanly again.** The nightly Dreaming summary and the knowledge-graph line showed a placeholder instead of the real count, and Dreaming could list raw bits of page data as if they were facts. The counts now read normally, only real sentences are kept as memories, and any old stray fragments stop showing.
-- **Studio templates fill themselves in.** Picking a template for Skales Studio now opens it with the prompt already in place on the right tab, instead of an empty screen.
-- **What you set about yourself stays yours.** The emoji, occupation and goals you enter on the Memory page no longer get overwritten when Skales does its background tidy-up of what it knows about you. The cleanup used to let its model rewrite those user-set fields with its own guesses; now your own answers are kept every time, while a field you left blank stays open for Skales to fill in as it learns. Only the AI summary is meant to update on its own.
-- **Plan mode is genuinely read-only.** It used to let a writing connector or an external tool slip through because they were judged only by their name; Plan now reads what the tool would actually do, so a chat set to Plan investigates and proposes but changes nothing until you switch to Code.
-- **Codework approvals are steadier.** A pending approval or a proposed change is now also saved to disk so it is easier to keep track of in the background, and a stale internal reference that could send a tool down the wrong path was removed.
-- **A long chat that becomes a goal now proves it is finished.** When a long task quietly turns into an autonomous goal, Skales works out what "done" means for it and checks against that before saying it is finished, instead of taking the model's word, so a weaker model can no longer stop a converted task early.
-- **Codework re-checks the project folder before it reads anything.** The folder you point Codework at is now validated on the server as well, not only in the window, so a run always works inside a vetted, writable, non-system folder before any file is read.
-- **Plan mode catches more external write tools.** A connected tool whose name mixes a read word with a write one, like a "query and delete" tool, is now held back in Plan mode instead of slipping through on the read word, so a read-only investigation stays read-only.
-- **Wrapped reports an honest "most-used tool".** Studio, Organization, Playbooks, Templates and MCP activity used to fall into the Chat bucket, which inflated Chat and hid those tools; each is counted on its own now. A busy week is no longer trimmed early either, so the activity total reflects everything you did instead of stopping at an old internal limit.
-- **A shell command cannot slip a second action past the safety check.** A safe command that runs without asking in a coding session can no longer carry a riskier one hidden on a new line, and Auto mode now pauses for a single approval on a destructive command however its flags are spaced or spelled, and on a command that writes to a file outside the folder you bound. A force-push that rewrites remote history is still refused unless you have deliberately set safety to Unrestricted.
-- **A coding run stays inside its folder when it reads, not only when it writes.** A file reached through a link, or through a path that points up and out of the project, is no longer read from outside the folder you chose, so the read side now matches the protection the write side already had.
-- **The commands Skales runs never see your secrets, including the ones it runs for itself.** The internal version, status and search commands now also run with your API keys, tokens and Skales' own values stripped from their environment, and a few more credential shapes (a database URL, a webhook) are recognised and removed.
-- **A long chat that becomes a goal finishes in one go again.** A task that quietly turned into a goal at the step cap could get stuck asking you to keep going instead of completing on a clear, finished answer. It now completes unless a real check actually finds something unfinished.
-- **Your daily briefing still reaches you when the first look of the day was quiet.** If nothing was waiting at the first morning check, the day stays open so anything that comes up later still arrives in one calm briefing, instead of the slot being spent on an empty one.
-- **One reminder for a meeting, not two.** With a check-in style turned on, a meeting about to start no longer pings you twice from two different places.
-- **Skales keeps only what you meant to teach it.** A one-off request phrased like an instruction ("from now on, can you...") is no longer stored as a permanent rule, and pinning or deleting a saved lesson on the memory page can no longer collide with the background learning and lose your change.
-- **A new chat is never named "No response", and a real title is kept.** A title that happens to begin with "We", "I" or "The" is no longer mistaken for leaked reasoning and dropped, while an empty or thinking-only reply no longer becomes the chat's name.
-- **Every most-used surface on the home screen opens its own page.** A row for Codework, Studio, Organization, Playbooks or Templates now opens that surface instead of always opening Chat.
-- **Premium replies stay quick across a long session.** Replies from Claude reuse the stable part of the prompt from one turn to the next again, so a long conversation does not slow down as it grows.
-- **Role bundles load safely.** An empty or malformed bundle file no longer stops the rest of a bundle from loading, and a connector named with a scope or a symbol is recognised correctly.
-- **The dark and light mode control is labelled correctly** again, instead of reading "Appearance".
-- **Typing a command in a long message keeps the cursor where you are.** When your draft held a /command or an @mention and grew tall enough to scroll, the box could stop following your cursor and drop your taps in the wrong spot, so you could not fix a word. The box now scrolls with you and the highlighted text lines up with the caret, so a long message edits normally.
+- **"WHAT PEOPLE ARE BUILDING".** Six unsourced claims about customers, under a heading that presented them as real.