https://oni.sqzer.com/ommp/atom.xmlOh My Music Player — a standalone terminal music player built with Rust and ratatui2026-11-12T12:56:48+09:00urn:git:b27cc1a2e688ca93c931dfd84201b12d7b7da21b2026-11-12T12:56:48+09:002026-11-12T12:56:48+09:00debuman110329244+debuman@users.noreply.github.comUpdate landing page identity and links for debuman urn:git:0f46823fb0aaa2e7dd4df2ad99173b5c4b50ae872026-08-02T00:59:07+09:002026-08-02T00:59:07+09:00sqzersqzer-x@users.noreply.github.comUpdate the AUR package to 0.1.3 urn:git:cfd5e8f7f89e43f02ca951a0115f99f9c54d734b2026-08-02T00:53:32+09:002026-08-02T00:53:32+09:00sqzersqzer-x@users.noreply.github.comRelease 0.1.3 The help modal's key column is sixteen characters wide and Ctrl+drag border is sixteen characters long, so the field padded to nothing and the description began where the label ended: Ctrl+drag borderResize panels Shortening the label to Ctrl+drag restores the gap. The binding is unchanged. urn:git:ef862c9961b719aed39c1e1fd8e0815d42bfe88a2026-08-01T22:51:16+09:002026-08-01T22:51:16+09:00sqzersqzer-x@users.noreply.github.comPut the copy button inside the box, and centre the strike on the text The button was a second bordered element beside the command, so a command and its button read as two things. The border and the background belong to one box now, with the command and the button inside it, which is how the reference site does it. The rule was positioned at a fixed .62em from the top of the list item, which put it above the middle of the letters. It is drawn on a span around the text instead: an inline box is the height of the text itself, so 50% of it lands on the letters rather than on the line box, which sits lower and taller. urn:git:d4f81cca51acf4dc3ffef39a5a016a3f8e9693382026-08-01T22:40:20+09:002026-08-01T22:40:20+09:00sqzersqzer-x@users.noreply.github.comGive the commands a copy button, and let the setup cross itself out Clicking the command box grew it by the width of the word copied and then shrank it again. The button is its own element beside the box now, and the feedback is the icon swapping to a tick, so nothing moves. The prompt stays in a pseudo- element, which never reaches the clipboard. Install reuses the hero's command component instead of a card grid that stretched an eleven-character command across half the page. The setup you skip is crossed out a line at a time as the section comes into view. The rule is a drawn pseudo-element rather than text-decoration, which cannot be animated, and the markup ships already struck: the script un-strikes it on load and puts it back, so a reader without JavaScript sees the finished state rather than four live-looking instructions. Reduced motion strikes all four at once. Cut the three paragraphs of fine print under Install. The binary size, the Nerd Font requirement, the graphics protocols and the platform caveats are all in the README, which is where someone goes for them. urn:git:2335747957fa73d727cc9d8c7a765203e4df6a982026-08-01T22:24:52+09:002026-08-01T22:24:52+09:00sqzersqzer-x@users.noreply.github.comRewrite the page as a landing page rather than a README The page led with a spec list and then enumerated features: seven cards for the seven browse tabs, three benchmark numbers, a search section, a keybinding table, and a troubleshooting block, in that order. Nothing on it said what the program is for or who would want it, and one line defended the methodology behind its own benchmark, which nobody had questioned. Four directions were written and judged against three lenses — a stranger, the person this is actually for, and a copy editor. The one that won opens on the thing that makes someone want it. The hero says what it plays and where. The first section you scroll into is the album art with a sentence naming what you are looking at, so the screenshot has a reason to be there. Then the setup you do not do, struck out, against the one line you type. Then browsing, with the tab strip set as type in the app's own seven accents rather than seven cards. Then the small decisions — the untagged file that still has a home, the playlist entry that survives an unmounted drive — which are the reasons someone keeps a tool. Then the keys. Then install. Cut: the whole benchmark band; the feature card grid; the keybinding table that retyped the screenshot beside it; the requirements section, whose two real warnings now sit inside Install as prose, where the person about to run the command is; the GPU-acceleration aside about Alacritty and xterm; the git clone command nobody uses to install a music player. Every claim was checked against the source again: b opens the playlist modal only while something is playing and adds the playing track, h and l move between panes, the binary is 7.0M against libasound, libc, libgcc_s and libm, and the info panel has thirteen labels. urn:git:19203c2b68db46a34653e4f2420abdc9fcca99be2026-08-01T21:50:07+09:002026-08-01T21:50:07+09:00sqzersqzer-x@users.noreply.github.comReset the deploy branch to the remote before building on it GitHub commits to gh-pages itself whenever the Pages custom domain changes — it wrote a Delete CNAME and a Create CNAME during the move to ommp.sqzer.com. The deploy script fetched but checked out the local branch, which had not moved, so the next deploy built on a fork and its push was rejected. urn:git:dbf58134484d32b88e4011ba46730e85923e7b232026-08-01T21:48:25+09:002026-08-01T21:48:25+09:00sqzersqzer-x@users.noreply.github.comGive each section the measure its content wants One width for everything meant a section built around a 3840px screenshot was held to the same line length as a paragraph, which left the copy pinned to one edge with dead margin beyond it. There are two measures now, assigned per section the way the reference site does it: 68rem for prose-led sections, 78rem for the ones built around a grid or a screenshot, and 84rem for the hero shot, which is the product and is allowed to be the widest thing on the page. The split sections take the wide measure and a 26rem copy column instead of a fixed 22rem, so the text has a comfortable line and the outer margin at 1920 drops from 416px to 336px. Both containers now carry a 92% floor. Without it a viewport between the two caps — 1128 to 1288px, which is most laptops — pins the wide sections 20px off the screen edge while the prose sections stay 96px in, and a 76px stagger between neighbouring section headings reads as a mistake rather than a system. With it the two track each other and nothing touches the edge. The install boxes are stretched to a common height. The clone URL wraps to two lines, which left one box hanging 24px below its neighbours. Reviewed section by section at 1280, 1440 and 1920 against real renders. urn:git:f9e4785a349236d1d88697bc4c322c9c15fffe542026-08-01T21:15:44+09:002026-08-01T21:15:44+09:00sqzersqzer-x@users.noreply.github.comShoot every screenshot at the same size, and stop shrinking the second one The modal shots were taken at font-size 14 like the wide shot, not 20. Twenty was chosen to make the keybindings read once cropped, but the crops are gone and all it did was make the app look stretched next to the shot above it. Every screenshot on the page is now the density the program actually runs at here. The reversed section keeps its own column tracks. Reordering alone left the figure in the 22rem track, so the keybinding shot came out half the size of the search shot for no reason other than which side it sat on. urn:git:8748f0017cdb9da8ff7730465e40d3e643fcc9d32026-08-01T21:08:24+09:002026-08-01T21:08:24+09:00sqzersqzer-x@users.noreply.github.comOpen the page on the splash screen, and put the screenshots back as they were The hero fills with ommp's own boot screen. The splash scatters music notes across a dark field while the library loads; tools/note-field.py reproduces that scatter from the same hash the app uses, the same four glyphs and the same three colours. A note falls where (7x + 13y) mod 37 == 0, and shifting either axis by 37 adds a multiple of 37 to the hash, so a 37x37 tile repeats without a seam. It is rendered here rather than shipped as text because the glyphs live in the CJK fonts, not in the page's latin-subset JetBrains Mono. The two radial washes that stood in for it are still there, over the top. The screenshots go back to whole frames. Cropping to the modals was meant to make them readable and instead made them the loudest thing on the page. The layout goes back with them: copy in a fixed column, screenshot in the rest, alternating sides. Only the hero shot earns the full width — below it the screenshots illustrate a point the copy already makes. Both modal shots are now taken with playback running. One reading Stopped beside two reading Playing looks like a mistake rather than a screenshot. urn:git:3ce49f6d15fd49e599e790ec60f9c541fd5e956f2026-08-01T20:42:56+09:002026-08-01T20:42:56+09:00sqzersqzer-x@users.noreply.github.comCorrect two false claims on the page, and fix keyboard copy, the card and the icon An audit of the live site turned up two sentences that describe software this is not. The Artists card said tracks are grouped by album artist where tagged and track artist where not. Index::build keys by_artist on the track artist alone; album_artist_of() is called once, for the by_album key. A compilation tagged album_artist=Various Artists, artist=Yorushika files under Yorushika in the Artists tab. The sentence was true, but about the Albums tab — so it has moved there, and Artists now says what it does. The lede promised it picks up where you left off. SavedState holds volume, shuffle, repeat, pane widths, playlists, info view and right split; no queue, no current track, no position. Startup calls ReplaceQueue over the whole library. The claim is now the one the program can keep. The install commands were click-only, which left the page's primary call to action unreachable by keyboard. They take a tab stop, a button role and Enter/Space now. The role is assigned from the script rather than the markup: copying only works once the script has run, so without it they are plain text and should not claim otherwise. og:image was a raw 4K screenshot scaled to a card — at the size a feed renders it, nothing was legible and the name appeared nowhere. It is now composed: the wordmark, the headline, and the app's own progress bar. Legible at 240px. The favicon was the same 723x421 wordmark, which is a smear in a tab, and /favicon.ico had nothing behind it. Both are answered by one square mark cut from the wordmark's own 'O', carrying the logo's pink-to-cyan, shipped as an .ico at the site root. urn:git:0a5cb58b1ddb3da82913743d863277534d2d69492026-08-01T20:19:16+09:002026-08-01T20:19:16+09:00sqzersqzer-x@users.noreply.github.comReshoot the screenshots and move the site to ommp.sqzer.com The help modal's key column was widened to 18 characters, which pushed the longest row past the modal's inner width and clipped 'Focus prev / next pane' and 'Info panel: clock / art'. Back to 16, with a shorter label for the one binding that needed the extra space. The wide shot was taken at font-size 20, where the info panel stops at 'Disc #' and album names truncate mid-word. At 14 the full thirteen fields fit, sample rate and bit depth included, and nothing is cut. The modal shots stay at 20: a modal is a percentage of the terminal but its contents are a fixed number of characters, so a wider terminal only adds empty space inside it. Screenshots now ship at the captured 3840-wide scale instead of downsampled to 2400. Resampling terminal text both blurs the glyphs and inflates the PNG by inventing intermediate colours — the 4K originals were smaller than the downscaled ones. As AVIF the three come to 428K against 1.8M as PNG, with no visible difference at 1:1. og:image gets its own 1200x630 PNG, which is the size the card renders and the one format crawlers can all read. The two modal close-ups moved out of the half-width split and under it at the full width of the wrap. Squeezed beside the copy they were unreadable, which defeated the point of showing them. Also stops the clone URL scrolling out of sight inside its install card. urn:git:47ab1a509bebdb9bf944215d73aacb89bf5be6be2026-08-01T19:26:00+09:002026-08-01T19:26:00+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comTrim the page title and script the deploy The title concatenated the site description, which already repeats in the meta tag. deploy.sh does the build, the doctor run and the gh-pages mirror in one step so the branch never drifts from site/. urn:git:a1f1b09da51f16ff9ac87619d8205ec96b78b1df2026-08-01T19:24:00+09:002026-08-01T19:24:00+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd the landing page, and refresh what the README got wrong site/ is a sqzass project that builds the page at sqzer-x.github.io/ommp. Every colour in the stylesheet is lifted from the application's own source — BG_DARK, the logo's pink-to-cyan gradient, the progress bar's violet, and one accent per tab — so the page cannot drift away from the thing it advertises. Prose is set in Bricolage Grotesque and anything that is literally the app (commands, keys, tab names, the status bar) in JetBrains Mono, the font ommp is designed for. Both are self-hosted latin subsets, 128 KB in total. The screenshots are new. The one in the README dated from February and showed a UI that no longer exists: identical borders on every pane before focus was made visible, a nine-field info panel where there are now thirteen, and the tab colours before they were fixed. site/tools/shoot.sh reproduces them — Ghostty is the only terminal installed here that speaks a graphics protocol, so it is the only way to capture real album art rather than the block-character fallback. Three README claims had gone stale: album art has not been read from a directory scan since it started coming out of the file's own tags, the bottom-right pane renders track info rather than lyrics, and the crate is installable from crates.io and the AUR now rather than only from git. Also widens the help modal's key column, which was exactly as wide as "Ctrl+drag border" and so left no gap before the description. urn:git:b316a9813b04c64fa287a4cb96219b142cddb7672026-08-01T18:31:38+09:002026-08-01T18:31:38+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comIgnore the whole .claude directory Two entries listing individual files did the same job as one directory rule, and would have needed a third the next time something landed there. urn:git:a707cf53e205f98957e32db1a01a98722a26160a2026-08-01T18:27:56+09:002026-08-01T18:27:56+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comKeep the local worklog out of the repo urn:git:9143476093204319e5063dad7a3d1c40c185b1312026-08-01T18:21:11+09:002026-08-01T18:21:11+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comState the licence in the badge instead of deriving it The badge asked shields.io for GitHub's detected licence, and GitHub's image proxy cached the answer for four hours — so the README kept showing MIT well after the relicense, which is the one thing a licence badge must not get wrong. A literal badge cannot go stale, and it is edited alongside LICENSE and Cargo.toml when the licence actually changes. urn:git:4d09cf9e82d048943d84b28f2ea2e6443b0016922026-08-01T18:17:44+09:002026-08-01T18:17:44+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comUpdate the AUR package to 0.1.2 urn:git:cbb8040e1886c2b62b63d49db43a97c2e4860bfe2026-08-01T18:14:27+09:002026-08-01T18:14:27+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comRelease 0.1.2 Documentation only: the README now describes terminal support by graphics protocol rather than recommending Kitty by name, and carries the AUR packaging files. Bumped so the crates.io page shows the corrected text. urn:git:59e693fe46cf76c4bc1b762e8fa305af4cb73b1f2026-08-01T18:08:40+09:002026-08-01T18:08:40+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comDescribe terminal support by image protocol, not by name The README told people to use Kitty. What actually decides whether album art is a real image is the graphics protocol the terminal implements — Kitty graphics, Sixel or iTerm2 — and Ghostty, WezTerm, Konsole, foot and others all qualify. The protocol in use is detected at startup. GPU acceleration is not the criterion and the README now says so: Alacritty is GPU-accelerated and implements none of them, while xterm is not and supports Sixel. The AUR package drops its kitty optdepend for the same reason; naming one terminal was misleading and a terminal is not really an optional dependency of a terminal application. urn:git:e2a94beb25e643bcea35a1a30eaa24f3a2ce24c92026-08-01T18:00:17+09:002026-08-01T18:00:17+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd the AUR packaging files PKGBUILD builds from the tagged GitHub tarball with --frozen against the committed Cargo.lock, runs the test suite in check(), and installs the binary, LICENSE and README. Verified with makepkg: the package builds and lists only /usr/bin/ommp plus the licence and doc. Runtime dependencies come from ldd on the release binary — alsa-lib via rodio, plus glibc and gcc-libs. Nerd Fonts and kitty are optdepends because the interface degrades rather than fails without them. urn:git:fe95286fa8a69df1316763866a60dbb0762b00f42026-08-01T17:55:09+09:002026-08-01T17:55:09+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comSay what to do when stdin or stdout is not a terminal Raw mode failed with a bare "No such device or address", which says nothing about the cause. Check both streams up front and explain that ommp needs an interactive terminal and cannot run through a pipe or a redirect. Version bumped to 0.1.1: crates.io versions are immutable, so 0.1.0 cannot carry this. urn:git:87ad44bba361533c12ca96c32c1dab87c81d115a2026-08-01T17:40:14+09:002026-08-01T17:40:14+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comRelicense under Apache 2.0 LICENSE is the canonical text from apache.org verbatim, with only the appendix placeholders filled in: Copyright 2026 sqzer (hyukjin an) Cargo.toml, the README and the About box follow. The crate also excludes the 1.2 MB screenshot, which everyone installing would otherwise download for an image crates.io cannot resolve from a relative README path. Packaged size drops from 1.5 MiB to 349 KiB. urn:git:d6d7962221054204361e29cdaac2e66984a5c3eb2026-08-01T17:32:26+09:002026-08-01T17:32:26+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comRead the About box version from Cargo.toml It was a literal, so it would have kept claiming 0.1.0 after the next version bump. urn:git:3a6429990ef360ddc4ae6df626181f656590d55a2026-08-01T17:26:06+09:002026-08-01T17:26:06+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comTell the user the splash can be skipped, and skip it outright The splash could always be dismissed with a key, but nothing said so. A dim italic "Press any key to skip" now sits at the bottom, fading in and out with the rest rather than appearing at full strength. Terminals too short to fit it below the centred block simply do not get it. The key now drops the splash immediately instead of jumping to the fade-out. Half a second of animation after the prompt says the key skips is not what it promises. urn:git:25db58e89789322ca2be5a7e91aecc5573e4e7a02026-08-01T17:20:26+09:002026-08-01T17:20:26+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comLet the splash always play its full timeline Cutting the hold short when the scan finished saved about a second of startup, but the animation is there to be watched, not to fill time. The timeline is fixed again at half a second of fade-in, a hold, and half a second of fade-out. A keypress still skips ahead to the fade-out, which it always did. Starting the scan before the terminal query and the audio device open is kept: it costs nothing and only means the library is ready sooner behind the animation. urn:git:9e51afb3605b901755bcad54865dded790170e762026-08-01T17:14:47+09:002026-08-01T17:14:47+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comEnd the splash once the library is ready The scan already ran on its own thread alongside the splash, so the two were overlapped — but the splash held for a fixed two seconds regardless. With the tag cache the scan finishes in tens of milliseconds, which made the animation essentially the whole startup wait. Completing the scan now skips the splash forward to its fade-out, the same jump a keypress already performed, so a warm start takes about a second instead of two and the fade still plays. A scan slower than the hold is unaffected: the timeline runs to its normal end. The scan is also spawned before the terminal image-protocol query and the audio device open, both of which round-trip outside the process and neither of which it needs to wait for. urn:git:5309fc5c99d371a2d035996dd6a4d6b21bd3d9e02026-08-01T17:06:59+09:002026-08-01T17:06:59+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comScan in parallel and cache parsed tags between runs Scanning opened and header-parsed every file on every launch, one at a time. Measured on 5,000 files: serial, no cache 362 ms parallel, no cache 137 ms parallel + cache 18 ms (31 ms with a cold page cache) Files are now parsed across all cores with std::thread::scope, and parsed tags are stored in $XDG_CACHE_HOME/ommp/library.json. An entry is reused when the file's modification time and size still match, so a repeat scan costs one stat per file instead of an open and a header parse; retagging or replacing a file changes both and re-reads just that one. The cache is never load-bearing: an unreadable, corrupt or older-version file yields an empty map and everything gets parsed, and it is written via a temporary file and renamed so an interrupted save keeps the previous one. All 5,000 tracks were verified byte-identical between a cached load and a fresh parse. The benchmark numbers understate the cold-disk case: the test set was hardlinked, so it shared 50 inodes. Against genuinely distinct files the uncached path measured 0.72 ms per track, about 3.6 seconds for 5,000. urn:git:d4e77dde3a0dd010e63847bd941a9b748e6e336a2026-08-01T16:57:11+09:002026-08-01T16:57:11+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comPrecompute library views instead of rebuilding them every frame The derived views — artists, albums, genres, formats, directories — were recomputed from scratch on every access, each walking all tracks and cloning strings, and the render loop asks for several of them per frame. Measured on a synthetic library at 160x48: tracks before after 1,000 0.815 ms 0.409 ms 10,000 1.981 ms 0.313 ms 50,000 7.966 ms 0.337 ms Frame cost is now flat in library size rather than linear: the work happens once when the library is built, and lookups are a hash and a slice. path_to_index was a linear scan, so restoring a 1,000-entry playlist against a 10,000-track library cost ten million path comparisons; it is a map now. Extension search became an index lookup (1.594 ms to 0.008 ms at 50k), and the general search no longer lowercases five fields of every track per keystroke (4.62 ms to 1.98 ms). The queue pane asked playlists.iter().any(contains(..)) per visible row per frame; App keeps a membership set. Track::from_path had lofty decode and allocate every embedded cover — around a megabyte a track — only to drop it, since album art is read separately when the view needs it. Skipping it takes 50 real FLACs from 3.8 ms to 0.6 ms. The event loop redrew on its 50 ms timeout whether or not anything had changed, and drew between individual mouse-motion events so the unbounded channel grew faster than it drained. It now redraws only when an event changed something and the queue is empty, with the splash exempt because it animates. Event::LibraryReady is boxed so the channel element is not sized to a whole library with its index. urn:git:bb40b5bd6c6dbda79cc1a3e82422c74e6ea1571d2026-08-01T16:41:15+09:002026-08-01T16:41:15+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comReplace the Ctrl+E chord prefix with direct Ctrl shortcuts Search, help, about and library sync are now Ctrl+S, Ctrl+H, Ctrl+A and Ctrl+L instead of a two-keystroke chord. About moved off i because terminals send Ctrl+I as the same byte as Tab (0x09) and Ctrl+M as Enter (0x0D), so neither can be bound; Ctrl+H is fine, since Backspace is 0x7F rather than 0x08, and Ctrl+S is safe because raw mode clears IXON. The Ctrl arms have to be matched before the unmodified ones: the existing global bindings match on any modifier, so Ctrl+S would have toggled shuffle. Keyboard resize mode is gone — dragging a panel border already did the same job. The drag now requires Ctrl, so an ordinary click one column from a border selects a list row instead of silently starting a resize. Ctrl+C now quits from anywhere including an open modal, and Ctrl-modified keys no longer type a literal character into the search and playlist name fields. urn:git:3e8f73864576e9125d2be9f35286fbf7086d9ddd2026-08-01T15:58:34+09:002026-08-01T15:58:34+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comFix keybinding docs and modals that silently clipped content The help modal never listed p (toggle the info panel), h/l, Esc or Ctrl+C, and its 23 entries did not fit an 80x24 terminal — with no scrolling, the tail including "q: Quit" was simply cut off. It is two columns now. The README was missing c and Ctrl+C. The playlist modal drew every playlist from the top and let the selection walk past the bottom edge, so on a short terminal Enter toggled a playlist that was not on screen. The search modal reported its full height as the result viewport even though title_bottom reserves a row, which stopped keyboard auto-scroll one result short and mapped clicks on the count footer to a row that was never drawn. The tab bar had six colours for seven tabs, so the modulo wrapped and Playlists rendered in Queue's blue; hit-testing checked only the x range, so clicking the bar's border switched tabs, and the divider columns were attributed to the following tab. lyrics_pane renders track metadata, not lyrics — the name outlived the feature. Renamed, and its scroll offset, written by seven handlers and read by none, is now applied: the field list grew to 13 rows and did not fit the pane's default height. urn:git:40ce1fb2337c46b7c1e7423070f40851ebf38e902026-08-01T15:54:22+09:002026-08-01T15:54:22+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comCollapse the five duplicated list panes into one artists, albums, genre, format and playlists were 942 lines across five files. Normalising identifiers made three of them byte-identical over all 185 lines; the other two differed only inside the row-render closure and one call in the Enter arm. They are now one ListPane parameterised by rows and an activation, with the per-tab differences expressed as rows_for/activate. The seven-arm matches in handler.rs and ui/mod.rs collapse with them. That duplication was actively hiding bugs. Only one of the five clamped the selection at render time, and it survived only because a separate clamp ran on library reload. Every list clamped its scroll offset to count - 1 rather than count - viewport, so scrolling to the end left a single row above a screen of blank space — the correct formula already existed in the search modal and had never been back-ported. The format pane called get_tracks_by_format inside its render closure, rescanning and lowercasing every track for every visible row on every frame; it is one pass now. The albums pane emitted unpadded columns, so a CJK title and an ASCII one put the artist column at different offsets. get_tracks_by_album matched on the album title alone while get_albums pairs each album with its artist, so two different records named "Greatest Hits" merged and clicking either queued both. fit_to_width, format_time and centered_rect had two to four copies each and now live in ui::text. urn:git:3de0dadf2030be54bcea7dfd04a7a4a9a8366aaa2026-08-01T15:48:11+09:002026-08-01T15:48:11+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comDelete code that was never compiled or never read album_artists_pane.rs and search_pane.rs were not declared in panes/mod.rs, so 393 lines were never compiled — which is why nothing ever flagged that search_pane.rs references three AppAction variants that do not exist and would fail to build if it were wired up. They were also the only reason three Library album-artist methods carried allow(dead_code). Their backing state went with them: search_mode was never assigned true by compiled code, so search_query stayed empty and search_results stayed empty forever, making two branches unreachable and a remap in replace_library pointless. QueueState::scroll_offset was written in three places while the queue pane used its own. Six of the theme's sixteen fields were never read, hidden behind a blanket allow(dead_code). Track::lyrics survived the removal of the lyrics feature. The queue pane's mouse handler computed a hit test whose body was a single comment, duplicating what handler.rs already does. Event::Resize carried dimensions nobody used (the next draw re-queries the terminal) and AudioEvent::TrackError a message nobody displayed. No allow(dead_code) remains in the tree. urn:git:9a4c1e8c642bd141661d2a23fe884a1bd00143be2026-08-01T15:44:37+09:002026-08-01T15:44:37+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comMake focus follow the click, and make it visible border_focused and border_unfocused were both Cyan, so every pane computed a focus-dependent border colour and got the same answer — there was no way to see which pane had focus at all. They are now distinct. Focus itself was re-derived from the mouse on every 200ms tick, from a position latched on the first mouse event and never cleared. Pressing Tab moved focus for one frame before the next tick dragged it back under the pointer, and the same clobber fired on any mouse event including the scroll wheel, which also cancelled the auto-focus after adding to the queue. Focus now changes on a left click and nothing else; hover still highlights rows. The hit-testing was duplicated verbatim between the mouse handler and the tick refresh, so it is one PaneHit now. A click on the library pane's border or on the blank space below a short list synthesized Enter unconditionally and activated the previous selection; it is ignored unless the click landed on a real row, and hover is bounded the same way. Double-click compared only the row, so a click in one pane paired with a click at the same height in another; it now requires the same cell and clears once it fires. Mouse events during the splash were processed against the invisible UI. urn:git:e3b4a7ee271606ab38fba4d9ef678cc16816d62b2026-08-01T15:40:46+09:002026-08-01T15:40:46+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comFix playback: volume, queue semantics, shuffle, and repeat The player thread ran two loops: an idle one matching only Play and Stop that dropped every other command through a catch-all arm, and a playing one that recursed into itself per track. Volume lived only in the playing loop, so the restored volume sent at startup was discarded and every track that ended on its own started the next at rodio's default full volume. The recursion also left a stack frame and a still-connected Sink behind per track. Both are now one flat loop with the sink, volume and clock as locals. AddToQueue replaced the queue rather than appending, resetting the current index to 0 without telling the engine — so picking anything mid-song wiped the queue while the old track stayed audible and the UI named a different one. Splitting it into AddToQueue and ReplaceQueue makes the name true; only the startup load replaces. Shuffle picked a random index on every advance, so it could repeat a track immediately, skip others, and never terminate with repeat off. It is now a real permutation with a position, anchored on the current track when toggled, and previous walks back along it instead of stepping to current_index - 1. RepeatMode::One was applied to the manual next action too, leaving no way to skip forward while it was on, and a decode error routed straight back into next — under repeat-one an infinite retry that saturated the event channel. Advancing now distinguishes a user skip from a track ending, and repeated failures give up. Removing the playing entry left current_index pointing at the track that slid into the slot while the removed one kept playing, so that track was silently skipped; the selection was never clamped either, which killed both j and d until you pressed k. Clearing the queue and reaching its end set the UI to stopped without telling the engine, leaving audio running and the clock climbing. A failure to open the output device was printed into the alternate screen and swallowed; it now reaches the status bar. urn:git:f9d25e6adbbf8a682ad9833869da6eeb9d6175ca2026-08-01T15:34:38+09:002026-08-01T15:34:38+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comStop panics and silent data loss from destroying a session Terminal restore only ran on the normal return path, so any panic left raw mode, the alternate screen and mouse tracking on with the message scrolled away — install a panic hook that restores first. A backend write error hit the same class of problem, returning from the event loop before state was saved; record it and fall through instead. Playlists were pruned rather than preserved: replace_library dropped every path the new library lacked and the pruned result was written to disk on exit, so one unmounted drive erased them for good. Refuse a scan that comes back empty against a non-empty library, and park unresolved paths so they come back when the files do. The watcher seeded its debounce timestamp with the current time and never reset it after scanning, firing two full rescans about two seconds into every session with nothing changed — each one yanking the list cursor to the top. Only arm the debounce on a real event. Other crashes: the search modal cached raw library indices that no rescan path remapped, panicking on the next draw; play_next, play_prev and the queue renderer indexed library.tracks unchecked, now funnelled through one guarded play_queue_position; the border-drag math underflowed u16 when dragged past the dashboard edge and called clamp with min > max on a corrupt pane_widths, both replaced by split_widths; and the splash key handler subtracted from Instant without checked_sub. State is now written via a temporary file and renamed, every field carries a serde default so one unknown key no longer discards the whole file, and pane_widths is validated on load like right_split already was. urn:git:c28ef671890883609ba3c7c50818c28c9b228f132026-08-01T15:27:39+09:002026-08-01T15:27:39+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comRead album art and metadata from the audio file itself Album art was looked up by scanning the track's directory for a jpg/png, so files carrying an embedded cover in a directory without a loose image showed nothing. Decode the picture out of the file's tags instead, keyed on the track path rather than its directory, so per-track covers in a shared directory also resolve. Also surface the tag and stream fields that were already in the file but unused: year, track/disc totals, comment, sample rate, bit depth, and channel count. Fall back to the overall bitrate when a lossless format reports no audio bitrate. urn:git:33727f4edb996d630343dfee79d170bc3534b3bc2026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comWire up entry point and add README with assets Terminal setup and teardown, background library scan, and the main event loop dispatching into the action handler and renderer. README covers install, requirements, usage, and keybindings. urn:git:5aeea01f7c1878fa70c2636db895c0307e73f07a2026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd TUI panes, widgets, layout, and theming Resizable three-column dashboard with tabbed library browsing by artist, album, genre, format, directory, and playlist. Widgets cover the status bar, progress bar, info pane with album art and clock, and the search, playlist, help, and about modals. urn:git:439bfcd6f486260b0eb247c030157ca2420556f12026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd application state, action handling, and session persistence Central App struct holding library, queue, playlists, and playback state, driven by an AppAction enum. Key and mouse handlers translate input into actions. Volume, shuffle, repeat, pane layout, and playlists are saved to disk on exit and restored on start. urn:git:53fbcf0468bac8e2754a5e0613e3d34d5e44a8d22026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd audio playback engine and event plumbing Playback thread wrapping rodio with play/pause/seek/volume control and position reporting. Input and tick threads feed a single crossbeam channel, so key, mouse, tick, audio, and library events share one queue. urn:git:e7fc72c56f69f3deb70e1e6248a910bfaae83f172026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comAdd music library scanning, metadata, and filesystem watching Recursive scanner over the music directory, track model with tags read via lofty, embedded/sidecar cover art lookup, and a notify-based watcher that reports library changes back to the event loop. urn:git:c60eaede041dea6f87d12021f57e619b18c1fb002026-08-01T14:31:13+09:002026-08-01T14:31:13+09:00sqzer-x110329244+sqzer-x@users.noreply.github.comInitial commit: project scaffold, dependencies, and license Rust 2021 edition binary crate built on ratatui + crossterm for the TUI, rodio + symphonia for playback, and lofty for tag reading. MIT licensed.