You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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 call, 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.
