Desktop Environment

ShapeBit’s desktop runs on Hyprland, the compositor, and Quickshell, the shell that draws the bars, panels, Settings, and the lock screen, in a session run by UWSM. Its design follows the demo; where this page and the demo disagree, the demo decides. The source is in desktop/ of the monorepo.

This page has two parts: the desktop as built, and the future design of what is not built yet.

The desktop as built

Session

The login screen offers one session, ShapeBit. The login password unlocks the person’s encrypted home and their keyring.
The shell is the session’s own service for what applications ask of the desktop: notifications, administrator passwords, network passwords and Bluetooth pairing, the file chooser, screen sharing, screenshots, and permission questions such as using the camera, or which app opens a file. Questions show centered over everything, with the screens dimmed, and wait their turn.
Copied text stays available after the app it came from closes; only the last copy is kept, in memory, never on disk.
Printing and email are not offered yet.

Bars

Two bars: the system bar (the clock and the status items) and the dock (Apps, the workspace maps, and the quick actions). Each can sit on any edge, and their items can move between the bars’ zones or be hidden (Settings > Shell).
The system bar’s items each open a panel: Notifications, Microphone and Sound, Brightness (where the computer has a backlight), Network (Wi-Fi and Bluetooth, joined and paired in place), Input language, Battery with the power profile (where there is a battery), and the System menu (Settings, Lock, Suspend, Restart, Shut down, Log out). One panel is open at a time.

Notifications

New notifications show as banners on the focused screen and stay in the Notifications panel until dismissed. Critical ones stay until dismissed. Do Not Disturb holds back all banners but critical ones, and keeps the notifications all the same.

Workspaces and Windows

There is no Overview. The session starts with one workspace; dropping a window or an app on the “+” that the dock shows while dragging makes another, up to eight. A workspace without windows goes away.
Each workspace has a map on the dock, showing its windows where they are. With the pointer, windows move only on the maps: drag a window onto another to swap them, or onto another map to move it there. There are no title bars to drag; the keyboard moves windows too.
Windows tile in the workspace’s layout: Lead, Dwindle, Columns (the default), Center, Grid, Rows, Tabs, and Accordion. Tabs and Accordion show one window at a time. Handles between tiles resize them.
Apps show no close, minimize, or maximize buttons where they follow the system’s setting; full screen covers the bars, with a button to end it.

Keyboard

Every action has a shortcut, and Super+/ shows them all on a picture of the keyboard. The list is in Shortcuts. Shortcuts work by the position of the key, whatever the layout; on the lock screen only the sound, media, and brightness keys work.

Apps and Quick Actions

Apps lists the installed apps, with a search. A click opens an app on the current workspace; dragged onto a map, it opens there.
The quick actions sit on a bar or open from a hot corner: the assistant, capturing, and dictation.
Capturing opens one overlay for screenshots and recordings, of a selection, a window, or the screen. Screenshots are saved at once, or first open for markup; recordings can have a timer and the microphone. When an app shares the screen, the same overlay asks what to share, and the capture button shows while anything is shared.

Dictation

Dictation starts from its quick action or Super+D and writes into the app the person is in; it is separate from the assistant. Speech recognition runs on the computer, live while the person speaks, in any of 25 languages. Its model is an extra: the image ships no speech model, so dictation works once the extra is downloaded, and until then its button leads to the download.

Text goes straight into the text field only if the same field still has the focus when the text is ready, and goes in as text, not keys, so it can’t trigger shortcuts. Otherwise it waits in a card, to insert, copy, type, add to, or discard.
Password fields never get dictated text.
An open microphone is always visible on the bar. Audio stays on the computer and goes when the dictation ends. Locking the screen discards everything under way.

Lock, Idle, and Sleep

The lock screen looks like the login screen. The screen locks after the idle time chosen in Settings > Account and before the computer sleeps; the screens turn off after their own idle time.

Settings

Settings is part of the shell and has only what works: Account, Assistant, Language, Time, Desktop, Shell, Windows, Display, Mouse, Network, Sound, Apps (Applications), Downloads, Updates, and Licenses, with a search across them. Settings for the whole computer, such as the system language and the time zone, ask for an administrator’s password.

Not built yet: the first day of the week, dragging tabs, and keeping tile sizes across sessions.

Updates

The system updates itself in the background (Lifecycle). Once an update is ready, a notification says so, once, and Restart in the system menu reads Restart to update. Apps update daily; a running app gets its update the next time it opens.

Downloads

Everything the desktop downloads and installs goes through one queue: the browser and apps from Welcome, installs and removals in Settings > Apps, extras, and the daily app updates. One job runs at a time, outside the shell, so it goes on if the shell restarts, and what a logout stopped starts again at the next sign-in. Settings > Downloads shows the queue, where each job can be cancelled or tried again; a download goes on where it stopped.

Licenses

Settings > Licenses names the license of ShapeBit OS, opens the license texts of Fedora and the other programs in /usr/share/licenses, and lists each work that an extra holds: who made it, its license, any changes, and a link to its source, also before it is downloaded. Each downloaded extra also keeps the same in its NOTICE file.

Extras

Extras are what is too large for the system image and not everyone needs, such as speech and language models. They are downloaded once for everyone on the computer, checked against known hashes, and only administrators may add or remove them. They are kept apart from the sealed system, so the system needs none of them and updates without them. Welcome offers them; Settings > Assistant adds and removes them later.

Assistant

The assistant answers questions and acts on the computer when asked. As in the demo, it has three levels:

The command line, from its button, the hot corner, or Super+A, for a quick question.
The sidebar, which holds one conversation beside the windows.
Conversations, the list of all of them, with a search.

The assistant acts only through typed actions, never a command it writes. Each action shows as a card, and has a risk:

Read (the computer’s status, the installed apps, finding files by name) runs at once.
Change (volume, brightness, keyboard layout, Do Not Disturb, Wi-Fi, Bluetooth, opening apps, moving windows, screenshots, locking) asks on its card: once, always for that group, or not at all. Settings > Assistant shows the groups, to allow or take back. A change can be undone from its card.
Confirm (shutting down, restarting, installing or removing an app) asks every time.

One assistant serves the whole session, with one queue of requests, and runs outside the shell. Locking the screen hides it, and nothing new starts until unlock. Conversations and an activity log stay on the computer, the log for as long as Settings > Assistant chooses.

A model on the computer and an online one work the same way. The built-in model, Gemma 4 E2B, is an extra: on by default in Welcome where the computer has memory enough, and it runs only while in use. A person can add their own models, online ones too, with their keys in the keyring.

The assistant always says that an AI answers, as the EU’s AI Act asks (article 50): “AI · on this computer” (or “online”) beside the empty command line, and that answers can be wrong at the start of a conversation.

Welcome

Each person’s first login opens Welcome. It asks only what the installer didn’t:

Network, only if the computer was offline when Welcome opened.
Browser, only when online (Internet Browsers).
Recommended apps, only when online (Recommended Apps).
Optional downloads (Extras), only when online, for administrators.
Installing, which queues what was chosen and shows how it goes.
Getting around: the main ways to move around the desktop.

Skip all ends Welcome on any step, and what is installing goes on. Settings > Account opens Welcome again. If Welcome closed while the computer was offline, a notification offers to finish setting up once it is online.

Future design

Nothing in this part is built yet, except dictation and the assistant’s chat, actions, and models. The assistant’s chat follows the demo, which replaced the chat designed here.

Radial Context Menus

The shell’s and ShapeBit’s own apps’ context menus are radial (“pie”) menus. Third-party apps keep their own menus.

They open on the secondary button, or the Menu key or a long press, around the pointer or the focused object. A press, a move into a wedge, and a release runs that action; releasing in the center keeps the menu open.
A ring holds at most eight wedges, each in a fixed direction for its kind of object; menus never reorder. The primary action is at the top, and destructive ones in a consistent lower sector, marked by icon and color. Unavailable actions stay, disabled. At most two secondary levels.
Near a screen edge the whole ring shifts inward; the pointer never jumps.
Destructive, privileged, or consent-requiring actions still go through their usual confirmation. The menu checks its target again before it acts.
The keyboard works throughout, nothing depends on color, motion, or speed alone, and the menu never times out.

Shelf

The Shelf is a temporary place to keep things at hand while working across workspaces: global to the person, and cleared at logout.

Things enter only by an explicit action, such as a drag or Send to Shelf; the clipboard is never captured. Files and folders stay references; text, images, and links become the Shelf’s own copies.
It opens from a screen edge (the bottom by default), first as a compact preview, then as the full Shelf, over the windows without changing their layout, and from a shortcut, above full screen too.
Items keep their order, can be dragged in and out, and opening one leaves it there. Removing one, or clearing the Shelf, is immediate, with Undo.
Copies share a size limit; nothing is evicted silently.

Assistant Framework

The assistant framework is the shared foundation for assistant and agent interaction.

Voice and chat are two ways into the same assistant, with the same access and confirmations, and one queue of requests in arrival order. Stopping a request lets the next one start; it never rolls back what is done.
The assistant may read the person’s files, window contents, and clipboard. Changes are off until allowed per group, and destructive, privileged, or outward actions always ask. Permissions are per agent and per capability, in one place that shows and takes them back. Approvals are given on the card, with the pointer or keyboard, not by voice.
Memory stays on the computer and can be viewed, edited, and cleared; raw audio and transcripts never persist.
Open: how window contents differ from screen pixels, which each need a fresh grant under the media and privacy policy; detailed capability scopes; and isolating future agents.

Voice Interaction

Talking to the assistant, as dictation already does for text:

Only push-to-talk or a click; never always listening. Speech and requests stay on the computer, and audio and transcripts are discarded after each interaction. Never on the login screen, the lock screen, or password prompts.
No modes: the sparkle button, held, sends speech to the assistant; the waveform button dictates into the app.
Answers are spoken by default, on the computer, ducking other sound. A click on the indicator or a new push-to-talk stops speech. A request that needs a permission shows the same card as chat.

Gaze Input

An optional way to move the pointer to where the person looks:

Looking alone never moves or clicks anything: a key or button moves the pointer to the latest stable gaze point, without snapping to controls.
Camera frames and gaze data stay on the computer and in memory; no app or agent can read them. A privacy indicator shows while tracking, and tracking stops on the login and lock screens.
It needs calibration per display, and pauses when calibration is unreliable. Without supported hardware, Settings shows it as unavailable.