feresaria/spec.md

Ferese Window Manager Specification

Status: Draft for implementation
Target: Ferese v0 core and desktop-tier roadmap Language: Rust
Display protocol: Wayland

1. Purpose

Ferese is a Wayland tiling compositor for a conventional, keyboard-friendly desktop. It combines predictable window management with compositor-owned rounded geometry, shadows, animation, transparency, and semantic glass materials.

This document is the implementation source of truth for Ferese v0 and records the gated desktop-tier roadmap. The terms MUST, MUST NOT, SHOULD, SHOULD NOT, and MAY are normative within the tier being specified.

2. Product principles

1. Window-management behavior is predictable and familiar.
2. The compositor owns composition effects; clients request semantic intent.
3. Core state has one authoritative representation.
4. Rendering and animation never block input.
5. Optional shell processes may fail without terminating the compositor.
6. Effects degrade gracefully when performance requires it.
7. The nested compositor is the primary development environment.
8. Security-sensitive operations are unavailable to untrusted clients.

3. Scope

3.1 v0 core requirements

Ferese v0 is the smallest independently usable native-Wayland release. It MUST provide:

• native Wayland compositing;
• automatic and manually adjustable tiling;
• floating and fullscreen windows;
• multiple workspaces and outputs;
• keyboard and pointer input;
• native Wayland clipboard, primary selection, and drag-and-drop;
• client-side and server-side decoration negotiation;
• rounded clipping, borders, shadows, and transparency;
• frame-timestamp-driven window and workspace animations;
• damage-aware rendering;
• validated live configuration reload;
• authenticated local IPC and a command-line client;
• direct application launching;
• standard layer-shell support under the policy in Section 8;
• a direct DRM/KMS session; and
• the required native protocol baseline in Section 8.1.

The completed v0 MUST support routine native-Wayland software-development work without regular compositor restarts. The graphical Ferese shell is not required for basic window management or application launching.

3.2 Desktop tier

The desktop tier builds on the v0 core and includes:

• backdrop blur and semantic glass materials;
• the Ferese overview, launcher UI, workspace strip, OSD, and notifications;
• the graphical settings application;
• portal-backed screenshots and screen sharing; and
• X11 compatibility through xwayland-satellite unless an architectural decision record selects another design.

These features MUST NOT delay the v0 core. Each MUST degrade or fail without terminating the compositor or invalidating managed-window state.

3.3 Non-goals

The following are outside v0:

• a file manager, display manager, or lock screen;
• network, Bluetooth, or package-management UI;
• a general plugin, scripting, or theme system;
• desktop widgets;
• advanced dynamic layout algorithms;
• exhaustive support for every Wayland protocol or toolkit edge case; and
• pixel-for-pixel imitation of another operating system.

Ferese MAY display a configured color or image, but a standalone wallpaper daemon is not part of v0.

4. Architecture

4.1 Processes

Executable	Responsibility
ferese	Compositing, policy, input, scene construction, and rendering
ferese-shell	Overview controls, launcher, workspace UI, notifications, and OSD
ferese-settings	Graphical configuration editor
feresectl	Command-line IPC client

ferese MUST remain usable if another Ferese process exits. Shell and settings code MUST NOT run inside the compositor.

Wayland clients (plus optional X11 satellite)
          |
          v
  protocol adapters
          |
          v
 authoritative core state
          |
          +----> layout targets
          +----> animation state
          |
          v
      scene graph
          |
          v
 damage-aware renderer
          |
          v
 nested window or DRM/KMS output

Protocol handlers MUST translate requests into domain operations. Business logic MUST NOT accumulate in Smithay callbacks.

4.2 Technology baseline

The initial implementation uses Rust, Smithay, calloop, Smithay's GLES renderer, glow, targeted GLSL shaders, DRM/KMS, GBM, EGL, dma-buf, libinput, xkbcommon, libseat, serde, TOML, serde JSON, clap, tracing, and GPUI.

Dependencies MAY change when an architectural decision record explains the reason and migration impact.

4.3 Repository layout

ferese/
|-- Cargo.toml
|-- crates/
|   |-- ferese-core/
|   |-- ferese-layout/
|   |-- ferese-render/
|   |-- ferese-animation/
|   |-- ferese-config/
|   |-- ferese-ipc/
|   |-- ferese-protocols/
|   `-- ferese-shell-protocol/
|-- compositor/
|-- shell/
|-- settings/
|-- feresectl/
`-- protocols/
    |-- ferese-effects-v1.xml
    `-- ferese-shell-v1.xml

ferese-layout MUST be pure Rust and MUST NOT depend on Smithay or Wayland objects.

4.4 Domain operations

Wayland callbacks, input handlers, IPC handlers, configuration reload, and shell requests MUST enter core policy through explicit domain operations rather than mutating unrelated managers directly. A representative shape is:

pub enum Command {
    Focus(Direction),
    Move(Direction),
    Resize(Direction, f32),
    SwitchWorkspace(WorkspaceId),
    MoveToWorkspace(WindowId, WorkspaceId),
    ToggleFloating(WindowId),
    ToggleFullscreen(WindowId),
    SetConfig(ConfigChange),
}

pub enum Event {
    WindowMapped(WindowMapData),
    WindowUnmapped(WindowId),
    OutputConnected(OutputDescriptor),
    OutputDisconnected(OutputId),
    SurfaceCommitted(SurfaceCommit),
}

pub struct WindowMapData {
    pub window: WindowId,
    pub surface: SurfaceId,
    pub app_id: Option<String>,
    pub title: Option<String>,
    pub parent: Option<WindowId>,
    pub requested_fullscreen: bool,
    pub constraints: SizeConstraints,
}

A command or event is applied as one state transaction. Commands express requested policy changes; events describe facts received from protocol or backend adapters and include the metadata needed to apply policy. Neither type MUST be used for emitted notifications.

The transaction MUST either complete with all Section 5 invariants restored or leave authoritative state unchanged. Rendering, protocol configure requests, IPC events, and persistence are consequences emitted after the state transition; they MUST NOT become a second source of policy state.

4.5 Threading and execution

The calloop thread owns the Wayland display, input dispatch, authoritative FereseState, policy transactions, configure scheduling, and scene publication. No other thread may mutate compositor state or Wayland objects.

Shader compilation, image decoding, configuration file I/O, IPC serialization, and other bounded background work MAY run on workers. Workers receive immutable snapshots or owned messages and return results through bounded channels. The calloop thread validates each result against current identifiers and generations before applying it. Rendering MAY run on the calloop thread initially; if moved to a render thread, scene snapshots and buffer ownership MUST cross an explicit handoff and never require input dispatch to wait for GPU completion.

Blocking filesystem, process-wait, D-Bus, portal, and client IPC operations MUST NOT run on the calloop thread.

5. Authoritative state

The compositor MUST maintain one state root:

pub struct FereseState {
    pub outputs: OutputManager,
    pub workspaces: WorkspaceManager,
    pub windows: WindowManager,
    pub seat: SeatState,
    pub layout: LayoutManager,
    pub animations: AnimationManager,
    pub config: Config,
}

Smithay objects SHOULD remain behind adapter types rather than become the domain model.

5.1 Identifiers

Outputs, workspaces, windows, layout nodes, and the logical seat MUST have stable internal identifiers. An identifier MUST NOT be reused during one compositor lifetime.

pub struct WindowId(u64);
pub struct WorkspaceId(u64);
pub struct OutputId(u64);
pub struct NodeId(u64);
pub struct SeatId(u64);
pub struct SurfaceId(u64);

Ferese v0 exposes exactly one logical seat. Multiple physical keyboards, mice, and touch devices MAY feed that seat. Multi-seat policy is outside v0 and MUST NOT be implied by the domain model.

5.2 Windows

pub struct Window {
    pub id: WindowId,
    pub app_id: Option<String>,
    pub title: Option<String>,
    pub workspace: WorkspaceId,
    pub placement: WindowPlacement,
    pub fullscreen: bool,
    pub constraints: SizeConstraints,
    pub urgent: bool,
    pub material: WindowMaterial,
}

pub enum WindowPlacement {
    Tiled,
    Floating { rect: RectF },
}

Fullscreen is a presentation state layered over the preserved tiled or floating placement. Entering fullscreen MUST NOT remove a tiled window from its layout tree or a floating window from its floating list. At most one window may be fullscreen on a workspace. While fullscreen is active, that window is the sole normal managed-window content presented for the workspace, although permitted shell overlays and popups may still appear. Leaving fullscreen restores the preserved placement without reconstructing it.

A window MUST NOT independently store an output. Its output is derived from its workspace, preventing stale ownership after workspace migration. Authoritative keyboard focus belongs to SeatState and MUST NOT be duplicated on every window or workspace. Layout hierarchy MUST NOT be embedded in Window.

5.3 Workspaces

pub struct Workspace {
    pub id: WorkspaceId,
    pub name: String,
    pub output: Option<OutputId>,
    pub root: Option<NodeId>,
    pub floating: Vec<WindowId>,
    pub last_focused: Option<WindowId>,
}

pub struct SeatState {
    pub id: SeatId,
    pub keyboard_focus: FocusTarget,
    pub pointer_focus: Option<SurfaceId>,
    pub focused_output: Option<OutputId>,
    pub focused_workspace: Option<WorkspaceId>,
}

pub enum FocusTarget {
    None,
    Window(WindowId),
    Shell(SurfaceId),
}

Workspace::last_focused is restoration/history state only. It MUST NOT be interpreted as current keyboard focus.

These invariants MUST hold after every state transaction:

1. Every managed window belongs to exactly one workspace.
2. Every attached workspace belongs to exactly one connected output; only a workspace awaiting an output may contain None.
3. An output exposes exactly one active workspace.
4. A window with WindowPlacement::Tiled appears exactly once in its workspace tree and not in the floating list.
5. A window with WindowPlacement::Floating appears exactly once in the floating list and not in the tree.
6. Fullscreen does not alter the window's underlying placement membership.
7. A FocusTarget::Window refers to a managed window on seat.focused_workspace; a FocusTarget::Shell refers to a live, authorized interactive shell surface.
8. Workspace::last_focused, when present, refers to a window belonging to that workspace.
9. At most one window is fullscreen on a workspace at a time.

Workspace migration MUST be atomic. If an output disappears, its workspaces move to a deterministic remaining output. If none remains, they stay detached until an output appears.

Numeric workspace names are unbounded, but the default bindings directly address 1 through 9 and create them lazily. Named workspaces MAY be added after v0.

5.4 Focus semantics

The single v0 seat owns authoritative keyboard focus, pointer focus, focused workspace, and focused output. A workspace stores only last_focused so focus can be restored when that workspace becomes active.

When switching to a workspace that is already visible on another output, Ferese MUST focus that output and workspace instead of moving or swapping the workspace. Explicit workspace-move commands are responsible for relocating workspaces.

Mapping a normal application while another window is fullscreen MUST preserve the fullscreen window and place the new window into the underlying layout or floating list without revealing or focusing it. An urgent indication MAY be shown. A transient with a valid parent is floating by default, centered over and constrained to the parent's visible output, and follows the parent's workspace. Activation policy, not mapping alone, determines whether a transient receives focus.

Window rules MAY match normalized app_id, title, and transient metadata and set initial workspace, floating state, dimensions, or fullscreen state. Rules MUST be evaluated deterministically before insertion. Title matching SHOULD be avoided when a stable app_id exists.

6. Layout

pub enum Node {
    Window(WindowId),
    Split {
        axis: Axis,
        ratio: f32,
        first: NodeId,
        second: NodeId,
    },
    Stack {
        children: Vec<NodeId>,
        active: usize,
    },
}

pub enum Axis {
    Horizontal,
    Vertical,
}

Split ratios MUST be finite and clamped so descendants satisfy their minimum sizes where possible. If all constraints cannot be met, the focused window takes priority and the conflict MUST be logged.

The engine consumes a layout tree, usable logical geometry, size constraints, and tiling configuration. It returns a deterministic WindowId -> RectF map and constraint warnings. Identical input MUST produce identical output.

The following tree invariants MUST hold after every layout transaction:

1. The tree is acyclic and every reachable node has exactly one parent except the root, which has none.
2. Both children of a split exist and are distinct.
3. A stack contains at least two children and its active index is in bounds.
4. Removing a child collapses a one-child split or stack into its remaining child; removing the last child removes that branch.
5. Every tiled window is represented by exactly one reachable leaf and every reachable window leaf refers to a tiled window in the same workspace.
6. The focused leaf inside a stack is the active child's recursively focused window; directional operations enter a stack through its active child.

6.1 Automatic insertion

1. The first tiled window becomes the root.
2. A later window splits the focused leaf.
3. For default_split = "automatic", use a horizontal split when the focused rectangle is wider than tall and a vertical split otherwise.
4. The new window becomes focused after its first mapped commit.

6.2 Required operations

Ferese MUST support directional focus, movement and resizing; explicit split direction; floating and fullscreen toggles; tiled-divider resize; floating move/resize; and tiled drag-and-drop.

Drag targets are left, right, top, bottom, and center. A drag MUST show a live preview. Center creates or joins a stack; an edge creates a split.

6.3 Gaps

Gaps belong exclusively to [tiling]. inner_gap separates tiles and outer_gap separates tiles from output bounds. With smart_gaps = true, a lone tiled window has no outer gap.

7. Geometry and animation

Ferese MUST distinguish buffer, surface-local, logical-output, and physical coordinates. Layout and input policy use logical coordinates. Rendering performs physical conversion. Fractional scaling MUST avoid cumulative rounding errors between adjacent tiles. The renderer MUST snap shared logical boundaries once per output and derive adjacent physical rectangles from those same snapped edges; it MUST NOT independently round each tile's origin and size.

A managed window has three geometry domains that MUST remain distinct:

1. logical geometry: the authoritative layout destination;
2. visual geometry: where the compositor currently presents the window while an animation is in progress; and
3. client geometry: the most recently configured and committed client content size.

A representative model is:

pub struct AnimatedRect {
    pub current: RectF,
    pub target: RectF,
    pub velocity: RectF,
}

pub struct ClientGeometry {
    pub last_configured_size: Size<i32>,
    pub committed_size: Size<i32>,
}

pub struct WindowGeometry {
    pub visual: AnimatedRect,
    pub client: ClientGeometry,
}

visual.target is the authoritative logical layout rectangle. The target becomes authoritative immediately after the state transaction even if the client has not yet committed a buffer matching its new size.

For a non-interactive layout transition Ferese MUST:

1. compute and store the final target geometry;
2. send a configure for the final client size rather than one configure per animation frame;
3. continue presenting the most recent committed client content while visual geometry interpolates toward the target;
4. map pointer coordinates through the inverse visual transform before delivering surface-local input; and
5. accept a newly committed target-sized buffer without restarting the visual transition.

The renderer MAY use a short-lived snapshot of the previously committed content to avoid a visible source-size discontinuity. Whether using a live committed buffer or snapshot, client content MUST remain aligned with the visual geometry used for hit testing.

During a non-interactive resize, committed content is scaled uniformly to cover the visual content box and clipped to that box, anchored at its top-left content origin. Ferese MUST NOT expose uninitialized regions. When a matching commit arrives it replaces the scaled source without restarting the animation. A commit is matching when its acknowledged configure and effective buffer geometry match the requested client size after scale and transform.

If a client has not acknowledged and committed the final configure within 500 ms, Ferese MUST stop waiting for visual completion, retain the target layout, and continue presenting the latest valid content using the same scale-and-clip policy. The timeout is diagnostic rather than permission to send repeated configures. A client commit larger than its assigned tile is clipped to the tile; it MUST NOT overlap adjacent managed windows or change layout geometry.

During an interactive resize, configure events MAY follow pointer motion subject to normal coalescing and client responsiveness. Keyboard navigation MUST use the target layout topology, not transient visual geometry.

When animation is disabled or reduced motion is active, visual.current snaps to visual.target. Geometry uses spring animation; opacity, dimming, blur intensity, and overlays use bounded easing. Updates MUST use presentation or monotonic frame timestamps and MUST NOT assume a fixed refresh rate.

The default geometry spring uses mass 1.0, stiffness 320.0, and damping 36.0, which is approximately critically damped. Geometry defaults MUST NOT overshoot shared tile boundaries. User-configured underdamped geometry springs MAY overshoot numerically, but their rendered content remains clipped to the animated allocation and MUST NOT cover adjacent windows.

Animations MUST stop within configured tolerances. Hidden animations SHOULD stop requesting frames. Opening, closing, or moving a window MUST never block input.

8. Wayland behavior

8.1 Required native protocols

Before the native desktop is usable, Ferese MUST support the following core protocols and widely deployed extensions:

• compositor, subcompositor, shared memory, and outputs;
• xdg-shell;
• xdg-activation;
• seats, keyboard, pointer, and available touch devices;
• data-device clipboard and drag-and-drop;
• primary selection;
• xdg-decoration;
• relative pointer and pointer constraints;
• presentation timing;
• viewporter and fractional scaling;
• layer-shell;
• screencopy with the capture authorization policy from Section 13;
• cursor-shape-v1;
• text-input-v3 and input-method-v2;
• keyboard-shortcuts-inhibit-v1;
• idle-notify-v1 and idle-inhibit-v1;
• xdg-foreign-v2;
• single-pixel-buffer-v1; and
• Linux dma-buf where supported.

Layer-shell is a public standard role and MUST NOT require a Ferese private capability. Ferese MUST enforce exclusive-zone, focus, namespace, and layer policy so ordinary panels, launchers, and notification daemons can use it without receiving unrestricted compositor authority. Input-method access, screencopy, and shortcut inhibition MUST use explicit user/session policy; binding a public global alone MUST NOT silently grant capture or persistent keyboard interception.

Virtual input and session-lock MAY be added after their security and failure policies are specified. Portal integration is not a Wayland protocol: the desktop tier MUST separately provide an XDG Desktop Portal backend and PipeWire integration for interactive screenshots and screen sharing.

8.2 Decorations

Ferese MUST support client-side decorations. If a client negotiates server-side decorations, Ferese MUST provide usable move, resize, and close interactions. Shadows, masks, and focus borders are compositor effects and may apply in either mode.

8.3 XWayland

The desktop tier SHOULD use xwayland-satellite, launched on demand, so X11 clients become ordinary xdg-shell toplevels from Ferese's perspective. The layout engine and Window abstraction MUST remain protocol-agnostic. Clipboard and primary-selection interoperability depend on the corresponding public Wayland protocols remaining available to the satellite.

An embedded X11 window manager requires an architectural decision record and MUST justify the additional selection, focus, stacking, and override-redirect policy. Tests MUST cover X11 menus, dialogs, games, and applications that request absolute positioning; the satellite approach reduces compositor complexity but does not guarantee every X11 positioning convention.

9. Scene and rendering

Ferese MUST build an explicit scene from authoritative state and current animation values. Nodes SHOULD cover surfaces, popups, decorations, shadows, masks, borders, materials, shell UI, drag icons, and cursors.

Scene order is background, windows back-to-front, shell/overlays, drag icons, then cursor. Within a window, the shadow is behind clipped background and client content; the border is above. Rounded clipping MUST be established before drawing clipped content or through an equivalent offscreen pass. Client buffers MUST NOT be modified to implement effects.

Rendering MUST be damage-driven. Damage includes commits, movement, animation, shadows, borders, material invalidation, shell, cursor, and output changes. Moving or removing a node MUST damage both old and new bounds. Expansion MUST include shadow, blur-kernel, and sampling extents.

Direct scanout SHOULD be possible when one compatible fullscreen surface covers an output and no transform, overlay, capture, conversion, or effect requires composition.

Rounded corners use a GPU mask or equivalent. Fullscreen windows use zero radius. Shadows SHOULD use analytic shaders, cached textures, or reduced-resolution intermediates rather than full-resolution per-frame blur.

10. Material system

This section specifies the desktop-tier material system. It is not a v0 release gate.

Initial semantic materials are:

normal
glass.panel
glass.popover
glass.menu
glass.hud

The compositor maps semantic names to theme-controlled parameters. Clients SHOULD request names rather than shader values.

ferese-effects-v1 is private to trusted Ferese components. It MAY set or clear a semantic surface material. It MUST NOT expose shader code, unrestricted capture, or unbounded effect values. Its global MUST be advertised only to Wayland clients granted the effects capability by the private-client mechanism in Section 13.

Glass renders as:

lower scene content
  -> capture expanded region
  -> downsample
  -> dual Kawase blur
  -> upsample
  -> saturation and brightness
  -> tint and noise
  -> glass foreground

Blur MUST exclude the glass surface and all content above it. Sampling bounds MUST include the kernel radius.

A blur cache entry MUST identify its output and scale, visible and sample regions, material parameters, contributing lower-layer generations, and result texture. Changes to content, stacking, geometry, scale, transform, material, workspace, or output MUST invalidate affected entries. A global generation MAY be used initially; region-aware invalidation SHOULD replace it before performance tuning is complete.

10.1 Built-in visual profile

The desktop tier provides one restrained dark profile rather than a theme engine. Its defaults are a subtle blue focus accent, 14 logical-pixel window radius, 10 logical-pixel tiling gaps, soft shadows, 24 logical-pixel backdrop blur, and mildly increased glass saturation. Focused and foreground content MUST remain distinguishable without requiring transparency. Blur, glow, borders, and shadows SHOULD establish hierarchy rather than compete for attention. Settings, overview, notifications, launcher, and OSD MUST use the same semantic materials and spacing system.

11. Shell and overview

This section specifies desktop-tier behavior. ferese-shell owns overview controls, launcher, workspace strip, notifications, and volume and brightness OSD. Screenshot UI MAY be added after capture policy and implementation are complete. The shell uses a capability-scoped private Wayland connection plus authenticated IPC.

Overview MUST NOT require copying each managed window into the shell process. The compositor owns the live window presentation and enters an overview scene mode in which existing window scene nodes receive temporary visual transforms. ferese-shell supplies controls such as labels, selection affordances, workspace UI, and later search.

The overview binding is Super+Tab. Entering overview MUST animate windows from their current visible geometry into overview geometry. Pointer and keyboard selection MUST target the compositor-owned live window nodes. Dismissing overview MUST animate those same nodes back to the current layout targets. Workspace switching from overview MUST obey Section 5.4. Search MAY wait for the launcher milestone.

11.1 Shell surfaces

ferese-shell MUST use standard layer-shell for launcher, workspace UI, notifications, and OSD surfaces. Such a surface is not a managed application window and MUST NOT enter tiling, workspace focus history, or normal application task switching. Ferese MAY use a minimal private ferese-shell-v1 protocol for overview control and compositor-owned live-window selection, but it MUST NOT duplicate the standard layer-shell role or geometry model.

Initial semantic roles are:

launcher
overview-controls
workspace-ui
notification
osd

The compositor enforces layer, exclusive-zone, focus, input-region, and output policy. The shell provides content and requests a standard layer-shell role; it MUST NOT receive unrestricted arbitrary z-order or geometry authority. Interactive shell surfaces MAY receive keyboard/pointer focus while active, but Ferese MUST restore the prior managed-window focus when the interaction ends. Destroying or disconnecting the shell client MUST remove its shell surfaces without changing managed-window state.

11.2 Application launching and notifications

The launcher MUST execute configured argument vectors directly and MUST NOT pass them through a shell. Before launch, it MUST obtain an xdg-activation token tied to the triggering user interaction and make the token available to the child. The child receives the public Wayland display and normal session environment. It MUST NOT inherit a private Wayland descriptor, IPC capability, or bearer secret. A private descriptor deliberately inherited when the compositor starts a trusted component is the sole exception during that component's initial exec; the component MUST mark it close-on-exec before it can launch any child.

ferese-shell MUST implement the standard org.freedesktop.Notifications D-Bus service for basic notifications. The desktop tier MUST support summary, body, icon metadata, expiry, replacement, close, and actions. Notification content is untrusted: markup, image paths, and action identifiers MUST be parsed and rendered without command execution. If the shell restarts, it MUST reacquire the service and continue without affecting managed windows.

12. Configuration

The primary file is $XDG_CONFIG_HOME/ferese/config.toml, falling back to ~/.config/ferese/config.toml.

[appearance]
glass = true
blur_radius = 24.0
glass_opacity = 0.82
corner_radius = 14.0

[appearance.shadow]
enabled = true
radius = 32.0
offset_y = 8.0
opacity = 0.28

[appearance.active_border]
style = "subtle"
width = 2.0

[animations]
enabled = true
reduced_motion = false
speed = 1.0

[animations.spring]
stiffness = 320.0
damping = 36.0
mass = 1.0

[tiling]
inner_gap = 10.0
outer_gap = 10.0
smart_gaps = true
default_split = "automatic"

[input]
focus_follows_mouse = false
xkb_layout = "us"
xkb_variant = ""
xkb_options = []
repeat_rate = 25
repeat_delay_ms = 600

[input.touchpad]
tap = true
natural_scroll = true
disable_while_typing = true

[commands]
terminal = ["foot"]

[[bindings]]
keys = "Super+Enter"
action = "spawn"
argument = "terminal"

[[bindings]]
keys = "Super+Q"
action = "close"

[[bindings]]
keys = "Super+Tab"
action = "show-overview"

[[bindings]]
keys = "Super+H"
match = "keysym"
action = "focus"
argument = "left"

[[bindings]]
keys = "Super+Space"
action = "show-launcher"

# Example: remove a built-in binding.
[[bindings]]
keys = "Super+Q"
disabled = true

[[window_rules]]
app_id = "org.example.DialogApp"
floating = true

Command values are argument arrays, never shell command strings. Binding records use a normalized key chord, match mode, action, and optional argument. keysym matching is the default and follows the active XKB layout. physical matching uses XKB physical key names and is independent of the selected layout. The complete built-in set is listed in Section 14; when a user supplies a record for the same chord and match mode it replaces that default. A record with disabled = true and no action explicitly unbinds the matching default. Duplicate normalized bindings, unknown actions, invalid arguments, and references to missing commands MUST fail validation. Every built-in binding MUST be expressible, replaceable, and removable through this schema.

Unknown fields SHOULD warn. Invalid types, non-finite values, negative sizes, and out-of-range opacity MUST fail validation.

Reload MUST debounce duplicate events, read the entire file, parse, validate, construct, and apply the difference as one transaction. Invalid input MUST NOT replace active configuration. Errors MUST be logged and emitted through IPC.

Settings and IPC MUST use the same validation path. Persistent changes MUST be written atomically using a temporary file and rename.

13. IPC and security

13.1 Transport and trust

The user-control socket is $XDG_RUNTIME_DIR/ferese/control.sock. Its parent directory MUST be owned by the user with mode 0700; the socket MUST use 0600. The compositor MUST reject peers whose effective UID, read from OS peer credentials, differs from its own.

Same-UID checks authorize user control only; they do not identify a trusted UI client. Privileged Wayland access is capability-scoped per client. Initial capabilities are:

effects
overview-control
global-metadata
capture
virtual-input

Trusted UI processes MUST receive a private Wayland connection created by the compositor. Ferese creates a socket pair, registers the server end as a Wayland client, and passes the client end only through an inherited descriptor such as WAYLAND_SOCKET. Privilege MUST NOT depend on PID, executable name, command-line arguments, environment secrets, or a discoverable private socket path.

The compositor associates an explicit capability set with each private Wayland client and filters private globals and requests against that set. In particular, ferese-effects-v1 requires effects and the private overview requests in ferese-shell-v1 require overview-control. ferese-shell receives only the capabilities required by enabled shell features. ferese-settings receives effects when spawned as a trusted Ferese component; its configuration changes still use the normal validated IPC path.

The private capture capability permits unattended capture only for a trusted component. Public screencopy clients remain subject to the interactive or configured capture policy and do not acquire this capability merely by binding the public screencopy global.

If a trusted component launches another privileged Ferese component, it MUST ask the compositor to create and pass a new capability-scoped private connection; it MUST NOT forward its own Wayland descriptor or capabilities.

13.2 Framing

IPC v0 uses unsigned 32-bit big-endian length-prefixed UTF-8 JSON:

4-byte payload length | JSON payload bytes

The maximum payload is 1 MiB. Oversized, truncated, invalid UTF-8, or invalid JSON frames close the connection after a rate-limited diagnostic.

{
  "version": 1,
  "id": 42,
  "type": "command",
  "command": "focus",
  "args": { "direction": "left" }
}

Responses echo id and contain exactly one of result or error:

{
  "version": 1,
  "id": 42,
  "result": {}
}

Errors have stable codes. Unknown fields MAY be ignored; unknown commands and unsupported versions MUST return errors.

13.3 Commands and events

User-control commands are:

focus                 move
resize                workspace
move-to-workspace     toggle-floating
toggle-fullscreen     close
reload-config         get-focused-window
get-workspaces        get-outputs
get-config            set-config
show-overview         hide-overview
show-launcher         hide-launcher

Enumerating all window titles or capturing pixels requires shell privilege.

Events include window-opened, window-closed, window-focused, window-moved, workspace-changed, output-added, output-removed, config-changed, and config-error. Payloads MUST omit titles and application IDs without shell privilege.

Subscribers have bounded queues. The server MUST disconnect a subscriber whose queue overflows rather than block the compositor.

14. Input

Ferese MUST support click-to-focus, configurable focus-follows-pointer, floating move/resize, tiled rearrangement and divider resizing, pointer constraints, relative pointer, cursor-shape, text input, configurable XKB layout and repeat, touchpad tap/natural-scroll policy, and configurable bindings. Default bindings match keysyms, so Super+H/J/K/L follows the active keyboard layout. Users MAY select physical matching per binding when spatial key positions are desired.

A focused client MAY inhibit compositor shortcuts only while its requesting surface is visible and holds keyboard focus. Ferese MUST retain an emergency escape binding that cannot be inhibited. Input-method activation MUST follow the focused text-input request and MUST NOT grant the input-method client general shortcut or capture authority.

Super+Enter             terminal
Super+Q                 close window
Super+H/J/K/L           focus left/down/up/right
Super+Shift+H/J/K/L     move left/down/up/right
Super+1..9              switch workspace
Super+Shift+1..9        move window to workspace
Super+F                 toggle fullscreen
Super+Shift+Space       toggle floating
Super+Space             launcher
Super+Tab               overview

The launcher and overview bindings are desktop-tier defaults and MUST be installed only when their actions are available. Their absence from a v0-only installation is not a binding error.

Bare Super MUST NOT activate overview. Interactive grabs MUST have explicit cancellation. Losing a device, seat, surface, or output during a grab MUST end it safely.

15. Outputs

Each connected output has a mode, scale, transform, logical position, usable region, and one active workspace. v0 MAY choose preferred modes automatically but MUST handle hotplug without losing windows.

On startup, the first output creates and activates workspace 1. Selecting an uncreated numeric workspace from 1 through 9 creates it on the focused output. A newly connected output receives the lowest unused numeric workspace name and activates it; this MUST NOT silently move an existing workspace.

Output removal selects a target from the remaining outputs in this order:

1. the v0 seat's focused output, if it remains connected;
2. nearest output to the removed output by logical-center distance;
3. lowest stable OutputId as a tie-breaker.

A workspace MAY move between outputs but MUST NOT appear on two outputs at once. Switching to a workspace already visible on another output changes the seat's focused output/workspace to that existing presentation and MUST NOT implicitly move or swap workspaces.

When an output is removed, the migration target keeps its current active workspace unless the removed output held the seat focus. In that case, the removed output's active workspace becomes active on the target and the seat's managed-window focus is restored there where possible.

Ferese MUST derive a persistent output identity from stable connector and EDID properties and record the workspaces evacuated during the compositor lifetime. When that output returns, those workspaces return to it atomically if they still reside on their automatic migration target and have not been explicitly moved since disconnection. Its previous active workspace is restored when available. Explicit user movement wins; Ferese MUST NOT take a workspace back from a new assignment. If no recorded workspace can return, normal new-output assignment applies.

16. Performance and observability

The target is sustained output refresh at 60 Hz on integrated graphics during ordinary use. Animation MUST also work at higher and variable refresh rates.

Ferese SHOULD track frame, render, layout and blur duration; damaged pixels; missed presentation deadlines; surface and glass-region counts; and active animations. Performance tracing MUST be optional. Suggested targets are ferese::wayland, ferese::layout, ferese::render, ferese::input, ferese::ipc, and ferese::xwayland.

17. Testing

17.1 Layout

Unit tests MUST cover insertion, deletion, focus, movement, ratios, constraints, resize, stacks, floating transitions, fullscreen placement preservation, workspace migration and restoration, gaps, shared-edge rounding, and deterministic geometry. They MUST NOT require Wayland. Property tests SHOULD prove arbitrary operation sequences preserve acyclic trees, valid parentage and stack cardinality, and do not duplicate or lose windows.

17.2 Configuration

Tests MUST cover defaults, partial files, unknown fields, invalid and non-finite values, duplicate key chords, unknown binding actions, missing command references, keysym versus physical matching, default unbinding, window-rule ordering, reload rollback, serialization, and atomic updates.

17.3 IPC and security

Tests MUST cover partial reads, multiple frames per read, oversized frames, unsupported versions, subscriptions, slow clients, peer-UID rejection, and denial of privileged commands without capability. Launch tests MUST prove that children receive the public display and activation token but cannot inherit any private Wayland or privileged IPC descriptor.

17.4 Rendering and integration

Pure tests MUST cover clipping, shadow bounds, damage expansion, blur sampling, cache invalidation, transforms, animated hit testing, and refresh-rate-independent interpolation. Geometry tests MUST cover mismatches between logical target, visual geometry, configured client size, and committed client size. Nested integration tests SHOULD cover configure/commit behavior, inverse-transformed input during animation, clipboard, popups, drag-and-drop, focus, fullscreen, xdg-activation, output removal, private-client capability filtering, overview scene transitions, notification replacement/actions, and shell restart.

18. Milestones

A milestone is complete only when its acceptance criteria pass. M0 through M6 form the v0 release path. M7 through M10 are independently gated desktop-tier work and MUST NOT delay the v0 release.

M0: Native window on screen

Implement the nested backend, public Wayland socket, output, seat, xdg-shell, basic renderer, and commit handling.

Acceptance: foot opens and redraws; keyboard, pointer, and popups work; closing a client leaves the compositor running.

M1: Essential native protocols

Implement clipboard, primary selection, drag-and-drop, decoration negotiation, xdg-activation, relative pointer, pointer constraints, presentation timing, and scaling.

Acceptance: text copies between native clients; drag-and-drop completes; decorated and undecorated clients can move and resize; scaled hit testing matches rendered content; a user-initiated test launch receives a valid activation token.

M2: Tiling and workspaces

Implement identifiers, layout trees, insertion, focus, movement, resize, gaps, floating, fullscreen, and workspaces.

Acceptance: five terminals tile without overlap; navigation works in all directions; close reflows layout; modes restore correctly; randomized tests preserve all invariants.

M3: Animated geometry

Implement logical, visual, and client geometry separation plus timestamp-driven animation.

Acceptance: layout changes do not jump; one non-interactive reflow does not send per-frame configure events; a client may commit its final size mid-animation without restarting the transition; inverse-transformed hit testing matches the visible surface; input stays responsive.

M4: Early hardware session

Implement the smallest useful libseat, DRM/KMS, GBM, libinput, and session- lifecycle path for one output. Record page-flip presentation timestamps and missed-deadline metrics. Keep the nested backend for development.

Acceptance: Ferese starts from a TTY, renders the M3 workload directly, survives session deactivate/reactivate, and drives animation from measured presentation timing. Frame pacing and animation are measured on the reference integrated GPU before blur or shell work begins.

M5: Native desktop interoperability

Implement layer-shell, cursor-shape, text-input/input-method, shortcut inhibit, idle notify/inhibit, xdg-foreign, single-pixel-buffer, screencopy authorization, multi-output hotplug, and the remaining Section 8.1 protocols.

Acceptance: representative native panels, launchers, notification daemons, IMEs, VM/remote-desktop clients, and screenshot clients exercise their protocols; output removal and return obey Section 15; security-sensitive requests are denied without the required policy decision.

M6: Shippable v0 core

Implement rounded clipping, shadows, active borders, transparency, damage expansion, validated textual configuration, input settings, window rules, direct application launching, IPC, and feresectl. Complete native-client crash, soak, and performance testing.

Acceptance: masks and damage are correct; configuration reload is transactional; defaults can be replaced or unbound; feresectl commands work reliably; a configured terminal launches without shell parsing; malformed clients do not affect availability; a 24-hour native-client soak has no compositor crash or unbounded resource growth; all Section 3.1 requirements pass.

M7: Glass materials

Before choosing a shell toolkit, build a minimal wayland-client probe that creates a layer-shell surface and sets a semantic material through ferese-effects-v1. Then implement capture, dual Kawase blur, tint, saturation, noise, caching, and invalidation.

Acceptance: the probe blurs only content behind its surface; backdrop changes invalidate correctly; the effects global is visible only to a client with the effects capability; a public client cannot bind it; the demo meets its frame budget on the reference integrated GPU.

M8: Overview and shell

First spike GPUI layer-surface creation, output discovery, focus, configure, and shutdown behavior. If the spike fails its lifecycle tests, use a thin native Wayland shell adapter or record a toolkit decision before proceeding. Implement the compositor overview scene, minimal ferese-shell-v1 overview control, launcher UI, workspace strip, OSD, and the standard notifications D-Bus service.

Acceptance: normal shell windows use layer-shell; Super+Tab controls overview; managed windows remain compositor-owned live scene nodes; selection follows visual geometry; notification replacement/actions work; shell restart does not disrupt windows; each trusted process receives only its assigned capabilities.

M9: Settings and portals

Build Appearance, Tiling, Shortcuts, Input, Workspaces, Displays, and About. Implement the portal backend and PipeWire integration for screenshots and screen sharing.

Acceptance: controls preview live; invalid changes do not replace active config; changes survive restart; shortcut editing can replace and unbind defaults; the Displays page represents hotplug and return behavior; portal requests require an interactive authorization path and produce working screenshots/streams.

M10: X11 compatibility and desktop stabilization

Integrate xwayland-satellite on demand and complete desktop-tier crash, soak, compatibility, and performance testing. Replacing the satellite with an embedded X11 window manager requires the architectural decision from Section 8.3.

Acceptance: representative X11 applications tile and float; clipboard works both ways; menus, dialogs, games, and absolute-position requests have documented test results; a 24-hour mixed-client soak has no compositor crash or unbounded resource growth; all Section 3.2 desktop-tier requirements pass.

19. First implementation slice

nested Smithay compositor
  -> two xdg-shell terminals
  -> deterministic two-window tiling
  -> logical/visual/client geometry separation
  -> spring reflow with coalesced final configure

The slice succeeds when A and B tile automatically; closing B immediately updates the logical layout while A expands visually; A receives no per-frame configure flood; pointer input remains aligned through inverse transforms; shared edges round identically; terminal changes damage the correct regions; and all interaction remains responsive.

M4 follows this nested slice immediately so frame pacing is validated on real hardware. Glass, GPUI, Settings, XWayland, notifications, portals, and theming MUST NOT delay it or the v0 core.

20. Definition of v0

Ferese v0 is complete when a user can:

1. start a direct Ferese session;
2. run ordinary native Wayland applications and layer-shell utilities;
3. copy, paste, drag, and use application popups;
4. tile, focus, move, resize, float, fullscreen, and close windows;
5. use multiple workspaces and outputs;
6. configure input, bindings, rules, gaps, corners, borders, shadows, and animations through validated live reload;
7. launch configured applications directly;
8. use an IME, inhibit shortcuts where allowed, take an authorized screenshot, and resume after output removal or session switching;
9. control permitted behavior through feresectl; and
10. perform routine native-Wayland programming work without regular compositor restarts.

The intended v0 result is a low-latency, independently usable tiling compositor. The desktop tier adds a coherent Ferese shell, glass, graphical settings, portals, and X11 compatibility without redefining core correctness. Visual effects support hierarchy and usability; they do not replace correctness, responsiveness, or security.