# valw: a region across monitors (design) Date: 2026-10-04 Status: approved in brainstorming (transparent gaps), pending written-spec review Builds on: region mode (`src/region.rs`, `src/selection.rs`), capture (`src/main.rs`, `src/frame.rs`) ## 1. Goal `valw region` today keeps a selection on the output where the drag started and clips it there. Let a drag carry on onto other outputs: each output shows its part of the selection, and the shot is one image of the whole rectangle, stitched from every output it covers. The user's layout: LVDS-1 1366×768 at (0, 0), HDMI-A-1 1920×1080 at (−277, −1080), both scale 1, so a box dragged from the laptop up onto the monitor is the main case. ## 2. Behaviour - The selection lives in global logical coordinates (it already does); Shift, Alt and Space keep working across outputs. - While dragging, every output whose area the rectangle touches shows its part of it (bright inside, dimmed outside); the others stay fully dimmed. Wayland keeps a drag's pointer events on the surface where it started, with positions beyond its edges, so the global position is still `start output origin + event position`. - The `W × H` label shows the size of the whole shot, next to the pointer, on the output the pointer is over (found from the global position; when it is over no output, on the nearest one). - On release: - inside one output: exactly as today (`Choice::Region(output, rect)`, opaque, the same pixels); - across outputs: one image the size of the whole rectangle in physical pixels at the largest scale among the outputs it covers; each output's part is cropped from its frozen frame and placed at its offset (resampled when that output's scale is smaller); areas that no output covers are transparent (alpha 0). Everything else (sound, clipboard, file, preview) as for any shot; the preview appears on the output the pointer was over at release. - A drag smaller than the click threshold (`MIN_PIXELS`) in the whole rectangle is still a click. ## 3. Structure | File | Role | |---|---| | `src/selection.rs` | `span(a, b, outputs) -> Option`: the rectangle's parts per output (output index, its `PixelRect`, where it goes and at what size in the stitched image) and the image size; `locate(point, outputs)`: the output a global point is over (or the nearest) | | `src/stitch.rs` (new) | `stitch(frames, span) -> RgbaImage`: transparent canvas, each part's crop pasted (resized when needed) | | `src/region.rs` | every surface draws its part (`selection_on` no longer anchor-only), redraw only surfaces whose part changed; label on the pointer's output with the whole size; release returns `Choice::Region` for one output or `Choice::Span(Span, preview_output)` | | `src/main.rs` | `Choice::Span` → `stitch`, source = the preview output | ## 4. Testing Unit tests (`cargo nextest`): - `span`: a rectangle inside one output → one part equal to today's `clip`; the user's layout, a box from the laptop up onto the monitor → two parts, their places, the image size; a box partly over no output → the image is the whole box; mixed scales (1 and 2) → the image at scale 2, the scale-1 part placed and sized ×2; nothing under `MIN_PIXELS` → None. - `locate`: a point on each output, in the gap (nearest output). - `stitch`: pixels land at the right offsets, gaps are alpha 0, covered pixels alpha 255, a scaled part is resized to its target size. - region: the surface where the drag did not start draws its part; the release decision (one output → `Region`, two → `Span`) as a pure function of the corners and the outputs. Manual checks (a section in `docs/test-checklist.md`): drag from the laptop up onto the monitor and back, Shift/Alt/Space across the edge, the label on the right output with the whole size, the stitched PNG (gaps transparent, seams exact), preview on the release output, single-output shots unchanged. ## 5. Definition of done `cargo clippy -- -D warnings`, `cargo fmt --check`, `cargo nextest run` and `nix flake check` pass; checklist updated; copland lock updated after the user's approval.