Calabrass Files

Calabrass v1.0.2 — The agent asks you, the app says what happened

This release is about the loop between the agent and you — it can now ask you a question properly, and it can tell you when it is running — and about the app saying out loud what just happened, instead of writing it in red on the page.

Download Calabrass-setup-v1.0.2-x64.exe (12.94 MB, SHA256 5ca242519305cfc91f31e8edf5c7661473560ec2991a0139f9ced88899893c09) from the Assets below.

✨ New — the Autonomous page

A new screen, /autonomous, where the four agents sit in the four corners of the window, each with its own mark and its own colour:

• Each card opens by growing taller than it is wide, so the open one takes the room it needs without pushing the others around.
• A connector line runs from every card to the mark of the app in the middle, and while a card is open its line carries that card's colour along it — you can see at a glance which agent is working.
• A Stop button sits across the bottom.
• The Settings page is gone; Autonomous takes its place in the sidebar.

❓ New — ask_questions, so the agent can ask you things

The model had no way to ask a question and wait for the answer: it would guess, or write the question into the chat and carry on. It has a tool now.

• A question comes back as a form above the composer — a shadcn questionnaire, docked flush on top of the input, with the answers sent back as your next message.
• The prompt now says, in every turn, that anything only you know must be asked with this tool rather than in plain text, and shows it the shape of a call.
• The reader is forgiving about the names a model writes on its own — description, values, type, a list that is not under a key — so a slightly different answer still lands as a form.
• /ask shows a fake questionnaire with every kind of question, to look at, and it makes no model call at all.

🔔 New — every message comes as a toast

• shadcn's toast component, and a small notice layer over it: notifyError, notifySuccess, notifyInfo, notifyWarning.
• Every error is a toast now. The error fields, the clear*Error() functions and the chat error selectors are gone from the stores: a failure is a notice, not a state. Providers, connectors, agents, workspaces, schedules, usage, models, chat and voice all say what went wrong, where it happened.
• Successes are said as well — a provider, connector, agent or workspace added, saved or removed; a task booked, updated or removed; the account created.
• A keyed notice is only said once until it is cleared, so a list that keeps failing to load does not nag you every five seconds.
• The panel is a little see-through — a faded background with a blur behind it, so it sits over the page rather than on it.

🧩 New — provider templates

Twelve providers, pre-configured and waiting for you, seeded when you sign up or sign in for the first time. All of them are off until you add your key, and every one of them opens with the endpoint, the model list and the quirks already filled in.

Google Gemini (its native adapter) · Groq (max_completion_tokens and the reasoning fix) · OpenAI · OpenRouter · Mistral · DeepSeek · NVIDIA NIM · Cloudflare Workers AI · Ollama (local) · Together AI · Cohere · Perplexity

• Cloudflare Workers AI is supported properly: its models are listed, and the key is read as account_id:api_token.
• The power button on a provider with no key opens the edit dialog instead of toggling, and turns amber with the tooltip "Add API key to enable".
• Anthropic (via OpenRouter) and xAI (Grok) are out of the presets.

📊 Improved — usage is counted per step

• Usage is now recorded on every LLM call, not once per message. An agent loop that calls the model five to ten times is now counted five to ten times, and the numbers in the usage page are the real ones.
• The overall card on the usage page is capped in height, so the page does not grow without end.

🐛 Fixed

• The Stop button works while a reply is streaming. It used to inherit the disabled state of the send button and could not be pressed; stopping also clears the whole streaming state now, not just the text.
• Navigating no longer flashes the previous page. The shell follows the committed router matches rather than the target of the click, and the session is checked once per run instead of on every navigation.
• Long chats no longer hang — a conversation keeps only the last twenty messages.
• Provider templates show up immediately after you sign up or sign in, instead of only after a restart.
• The top bar badge is a link to the repository now: ★ Star on GitHub, and the way to give one without leaving the app.
• The account at the foot of the sidebar is a row rather than a menu, and the icon on the right of it is the way out.

📦 Installer

• Calabrass-setup-v1.0.2-x64.exe — per-user, no UAC prompt, installs over v1.0.1 in place (the app identity never changes).
• Bundled MCP servers (excel-mcp, win-control-mcp) are still switched off until you turn them on: driving Excel or the keyboard is a deliberate choice.
• The version reads 1.0.2 in the installer, in the executable, in Apps & features and on the desktop shell.
• Your data is untouched by the upgrade: everything still lives in %LOCALAPPDATA%\Calabrass, and the uninstaller still asks before it deletes it.

⚠️ Notes

• The installer is not code-signed yet — Windows SmartScreen will show "Windows protected your PC"; choose More info → Run anyway.
• Requirements: Windows 10/11 x64, and the WebView2 runtime (the installer checks for it and offers Microsoft's download page if it is missing).

Full Changelog: https://github.com/CodewithMubasher/Calabrass/compare/v1.0.1...v1.0.2