A deep reliability pass across everything Skales does. Tools now do the work they claim and tell you the truth the moment something fails, instead of a confident "done" that quietly did nothing. This is the solid engine the next release builds its new look and feel on.
Highlights
- Tools tell you the truth. Every action checks the real result and reports an honest failure with the reason, instead of a cheerful fake success.
- Goals run to completion on their own. No more "Continue?" cards every dozen steps. A goal works the whole task through and only stops when it's done or genuinely needs you.
- Replies never freeze. Finished answers are kept even when a provider holds the line open, and reasoning models get the time they need to finish.
- Claude and Gemini stream word by word. The two premium providers no longer make you wait for the entire answer before showing anything.
Fixed
- Tools tell you the truth about whether they worked. Across messaging, calendar, WordPress, Notion, YouTube, smart home and more, a failed, unconfigured or partial action used to come back as success. Skales now checks the real result and reports an honest failure with the reason.
- Messages reach the right person, every time. WhatsApp matches a contact exactly and asks rather than guessing. Telegram only reports "sent" when the send went through. Long Discord messages are split instead of being silently cut off. Twitter mentions and your timeline show up again, and Signal correctly reports a failed delivery.
- Email is safer. A reply goes through the same trusted-address check as a new email, finds the original in any folder, and uses the right account. Deleting one email removes only that message.
- Reminders and scheduled goals are dependable. A reminder set for later no longer fires instantly, and a reminder that hits an error is retried instead of lost. A scheduled goal whose previous run crashed recovers and runs. An empty or malformed schedule is refused instead of creating a job that does nothing.
- Long tasks and big projects finish. A goal that writes many files no longer stops partway at a hidden limit. When a step is cut off by length, the valid parts still run and the rest is retried. A tool waiting for approval is no longer mistaken for a failure that aborts the task.
- Goals run autonomously to completion instead of asking you to babysit. A goal works the whole task through and stops only when it's done, when it genuinely needs your input, or before a consequential action like sending an email or deploying, where it still asks once (with a one-tap "always allow"). A plain chat that grows into a real multi-step task is carried on as a goal automatically. A goal that reached its limit picks itself back up while you're away. The step limit in Settings > Goals is now a safety ceiling, not a check-in cadence; 0 means run to completion.
- Replies do not freeze, and thinking models finish their answer. If a provider sends the end of an answer and then holds the line open, Skales keeps the finished answer instead of throwing it away. Reasoning models are given the time they need before Skales decides a reply has stalled.
- Claude and Gemini type their answers out as they go. Replies from Anthropic (Claude) and Google (Gemini) now stream word by word, the same as every other provider, instead of sitting on a blank screen until the entire answer is finished.
- The same behaviour across every AI provider. Tool use, sending an image by file, parallel tool calls and reasoning-model requests work consistently whether you run Gemini, OpenRouter, Kimi, DeepSeek, Anthropic or a local model.
- Your files stay inside your workspace. Editing, moving and copying a file resolve relative paths inside your workspace instead of escaping to the install folder, binary files are no longer read as text, and a saved session can no longer be opened through a crafted path.
- Web access is safer and actually returns the page. Built-in web fetching is routed through the same protection that blocks internal and local-network addresses, and a fetched page returns its real content to the model instead of only a character count.
- Media gets delivered. A generated image arrives on Telegram instead of a broken reference, a video scene actually renders into the result, a voice clip is returned even when Telegram is not connected, casting a link with special characters works, and a browser playbook reports a failed step instead of marking everything done.
- Connected services hold up. Google Drive and Docs refresh their sign-in automatically instead of breaking after about an hour, Notion reports a real failure on invalid input, Home Assistant validates a command and confirms the change happened, and a custom capability survives the next rebuild instead of being wiped.
- Your memory and goals do not get corrupted. The knowledge graph and goal schedules are written safely, so a crash or two changes at once can no longer empty or scramble them, and a partly-failed update is reported as exactly that.
- Skills are honest about themselves. Creating a skill verifies it actually loads before claiming success, deleting one matches it safely, and the built-in documentation opens correctly in the packaged app.
- Your message always shows, even when you send the same thing twice. Sending an identical message again no longer makes its bubble disappear after the reply arrives.
- The chat is ready the instant the answer lands. The message box re-enables as soon as the reply is saved instead of waiting on background bookkeeping.
Changed
- Skales IQ, the free built-in trial, is more private and more reliable. Every trial request enforces zero data retention at the provider, the trial no longer garbles ordinary words like "Google", and its daily limit survives a server restart. Activating the trial no longer turns on usage analytics on its own.
- Approvals reach you wherever you are, and the task finishes after you approve. A sensitive action started from your phone asks for your approval and waits for it, then carries on with the rest of the task. Writing or skill-authoring actions are held back in read-only Plan mode and ask first over WhatsApp.
- Simple turns stay snappy. A greeting, a thanks or a quick "ok" gets a fast, light reply, and one earlier hiccup no longer makes the rest of the conversation heavier.
- More connected integrations without the slowdown. Having many MCP servers, custom skills or Hugging Face Spaces enabled no longer weighs down every message, so you can keep more of them switched on at once. Disabling an integration still removes it entirely.