Deskemy

Approach: the first version was a Tauri app with a SvelteKit interface in a WebView2 window. mpv rendered into its own child window, raised above the WebView.

Problem: those are two separate surfaces, and their resizes cannot be atomic. The WebView resizes instantly on the GPU compositor; the video window is repositioned through JavaScript, IPC, and a Win32 call, landing one or two frames behind. On every fullscreen toggle and panel open, the picture visibly chased the layout. Two mitigations were tried and rejected: a reflow that rubber-banded, and a mask that blacked the video out during resizes, which was worse than the bug.

Approach: stop having two surfaces. mpv's render API draws into a Direct3D texture owned by the app, set as the content of a DirectComposition layer beneath a transparent WebView2 on the same window. The OS composites both, so resizing becomes a single atomic operation.

The gate: the whole plan hinged on one unknown — whether a composition layer would show through a WebView that Tauri hosts as a window rather than as a composition visual. If not, the fix meant forking Tauri's web view layer. So the first step was a solid magenta layer, with no mpv involved at all. It showed through, and the approach went ahead without a fork.

Shipping it safely: a side-effect-free GPU probe runs once at startup. If it passes, the player composites; if it fails — old GPUs, some virtual machines, Remote Desktop — it falls back automatically to the child-window player. The rule was never to ship a broken player, and the fallback is what kept it.

Problem: the compositor worked, but it was Windows only, and it still relied on JavaScript reporting the video's rectangle on every layout change. That combination was the blocker for Linux and macOS.

Approach: in Slint, mpv's OpenGL render API draws into a texture inside the interface's own GL context, displayed by an ordinary image element. Overlays are simply elements declared after it — the same code on every platform, with no rectangle plumbing at all. The layering problem stops existing instead of being solved again.

Alternatives: Iced was ruled out for having no accessibility support, no API for foreign textures, and yearly breaking changes. Flutter meant a Dart rewrite plus an FFI bridge. A plain HTML <video> element only plays browser-friendly codecs, and people's course libraries are not all H.264.

Problem: downloaded course videos are whatever the uploader exported — arbitrary containers, codecs, and subtitle formats. A browser-grade video element supports a narrow slice of that, so a web-based player would either refuse to play files or force people to transcode their library.

Implementation: playback goes through libmpv, loaded at runtime through FFI with libloading rather than linked at build time, so codec coverage is mpv's. libmpv-2.dll ships inside every installer and the portable zip, and the loader falls back to a copy on PATH or at DESKEMY_LIBMPV. The playback tests run against a headless mpv.

Problem: replacing the entire front end of an app people use every day is the classic way to spend months with nothing shippable.

Implementation: the first step was extracting a core crate with no knowledge of any interface — database, import, configuration, backup, and the libmpv bindings — and making the shipping Tauri app depend on it. Both apps then shared one core while the new one caught up, and the old one kept shipping. The two were deliberately not joined into a single Cargo workspace, because that would have moved the Tauri app's build directory and broken its build script and release flow. The Slint app can also render any screen offscreen to a PNG with no window, which made layout checks scriptable.

Problem: every 1.x install checks for updates. A rewrite that loses someone's library, or quietly installs alongside the old version, breaks trust with the exact people who made the project worth continuing.

Implementation: 1.x installs are offered 2.0 as an ordinary update, and the MSI upgrades an existing 1.x MSI install in place. The first launch migrates the library database forward in a way 1.x can still open. Every installer is signed, the app only accepts updates that verify against its embedded public key, and a verification step checks both installers against that key before a release is published. The whole check, download, verify, restart and install sequence can be rehearsed locally by pointing the app at a manifest served from disk.

Problem: importing a course means probing every video for duration and track metadata. Probing is slow and I/O-bound. Doing it while holding the database connection would serialize the whole import behind media inspection and leave the interface unresponsive on a large folder.

Implementation: the importer runs in two phases — probe, then persist. Probing runs on a worker thread off the database connection, with live progress, and only the resulting metadata is written in the persist phase. The user sees a preview of exactly what will be imported first, and the probed plan is kept, so confirming does not probe everything a second time.

Problem: if a library row is keyed by file path, reorganizing your folders throws away your history. Rename a directory and every lecture underneath it looks brand new — progress gone, bookmarks orphaned.

Implementation: lectures are matched by content rather than path, so a moved or renamed file keeps its watch progress and bookmarks, and adding a course folder again updates it rather than duplicating it. The thumbnail cache is content-addressed on the same principle, and an optional folder watcher picks up changes on its own.

Problem: title search only finds a lecture if you remember roughly what it was called. Often you remember a phrase the instructor said and have no idea which of two hundred lectures it was in.

Implementation: SQLite FTS5, bundled through rusqlite so there is no external database, indexes course, section and lecture titles. Optional subtitle search indexes the spoken text from your subtitle files and jumps straight to the matching timestamp, with the matched words highlighted.

Problem: "portable" is a promise about more than the executable. An app that claims to run from a USB stick but quietly writes to %APPDATA% and the registry has left traces where it promised not to.

Implementation: the installed build keeps its library under %APPDATA% and deliberately leaves it behind on uninstall, so a reinstall resumes where you left off. The portable build detects a .portable marker beside the executable and redirects all state into a data/ folder next to it. Since a portable copy cannot safely replace itself, it points you to the release page instead of self-updating.