Two specs failed locally and passed in CI, which is the least useful
direction for a disagreement to point.
**`dev-headless.sh` was the only launcher not stubbing out the
catalog.** `seed-sandbox.sh` and `ci.yml` both send
`YJ_CORE_INDEX_URL` to a dead address; the dev launcher did not, so the
app downloaded and built the real ~1M-row Explore catalog into the
run's YJ_HOME and every local `make e2e` after that ran against a world
CI never sees. Found by reading the failure screenshot: the spec had
searched Explore for its fixture album and the page was full of real
ones. It defaults to the dead address now and takes an explicit one for
exploring by hand.
**And the shared backend carries spec state between runs.**
`explore-shelves` staged its catalog only `IfEmpty`, so one album row
left behind by `requested-badge` satisfied that gate: the shelves were
drawn from a single foreign row and the artist card the spec clicks did
not exist. It failed on the *second* local run and passed on the first,
and never in CI, where every run gets a fresh home.
"Is the catalog empty" was the wrong question and "are my rows there"
is the right one, so staging is unconditional (INSERT OR IGNORE keyed
on the MBID) and the assertion moved from *this insert wrote a row* to
*every fixture row is present*. That is both idempotent and stronger:
an MBID that fails CHECK(length(mbid) = 16) is silently dropped by OR
IGNORE, which the old per-insert count caught only on a cold catalog
and the new one catches always.
Verified by running the whole suite twice against one app: 97/3 before,
100 passed both times after.
Plan 016 B2, phase 1. Below 600px the grid drops its sidebar column,
`bottom-nav` becomes the primary navigation, and the shell fits the
viewport instead of scrolling sideways out of it.
600 rather than the sidebar's own 900, because 900 is a laptop and the
answer there is a narrower sidebar, which is still a sidebar. Under 600
there is no room for one at all: 360px of viewport over a 200px nav is
not a layout.
**The tab bar is four destinations and a way to everything else.**
Three to five is where touch targets stop being thumb-sized -- eleven
over 360px is 32px each -- so the four are the ones plan 016's subset
says a phone is for, and "More" opens the *existing* `app-sidebar` in a
drawer rather than listing the destinations a second time. Two lists is
two places to add the next view to.
That reuse has a cost this found the hard way: a shared component
brings its `data-testid`s with it, so rendering the drawer's sidebar
unconditionally put a second `nav-home` (and ten siblings) in the DOM
and **failed 30 existing specs** with "resolved to 2 elements" -- on a
desktop viewport, where this element is `display: none` and the drawer
can never open. It renders only while the drawer is open, and the
component test asserts the absence, because the failure is invisible
from inside the component and lands in files nobody touched.
**What made the shell overflow was minimums, not padding.** Measured at
360px: the body was 652px wide, because a `min-width` in a flex row is
a hard floor and a grid item's implicit minimum is its content. So
`min-width: 0` on the boxes between the viewport and the content, and
each component stands its own non-essential parts down in its *own*
stylesheet -- search-bar's 200px floor, job-indicator's label (the
visible one; the live region that announces it is untouched),
audio-player's seek bar and volume. A media query inside a shadow root
is answered by the viewport, so this is the component saying what it
drops rather than the shell reaching in.
Volume goes because the hardware keys own it on a phone, which is the
same reason mediacontrols' Android handler implements no volume
callback. Seeking goes because 4px is not a thumb target; it belongs to
the full-screen now-playing view, which is the next phase.
An existing spec therefore asserts the opposite of what it did:
layout-overflow's 320px case used to require that the 464px behind
`overflow: hidden` could be *scrolled to*, which was the remedy
available while the shell had one layout. It reflows now -- 320px in a
320px viewport, exactly -- and reflow is what WCAG 1.4.10 asked for.
Section A of plan 016 is closed and B1 is decided, so the three tenses
move together: CLAUDE.md for what mediacontrols now is, the skill for
what to run, NOTES.md for what was measured and when.
The entry worth reading is the one that disproves a claim written here
earlier in the same session. Dropping x86_64 was expected to make
make android-install fail with INSTALL_FAILED_NO_MATCHING_ABIS.
Measured, it installs and launches: Google's google_apis x86_64 images
carry arm64 translation (abilist = x86_64,arm64-v8a), so the loader
maps lib/arm64/libwails.so and runs it. It dies before any of our code
with SIGILL, and the disassembly names the reason exactly --
`mrs x0, ID_AA64ISAR0_EL1`, Go's internal/cpu reading the arm64 feature
register at runtime init, which the translator does not implement. So
no Go binary starts under it, and that is not a property of this app.
Which closes the last plausible shortcut. There are now three distinct
ways this app fails on an x86_64 Android -- seccomp on the x86_64
build, an unimplemented system register on the translated arm64 one,
and a real device still unverified -- and none of them is a bug in it.
A phone remains the only verification path.
Plan 016 also carries the B2 scope, now decided rather than
recommended: option 1's data model with option 2's surface. The phone
gets home, library browse, now-playing-as-a-view, the queue, search and
playlists; it does not get autotag, downloads, Explore or the 93-control
Settings page, and each of those has a reason written beside it. One
rule for the work: no view forks, because a phone template that copies
a view's is two templates to fix every bug in.
Plan 015 shipped a pipeline; this is what stands between that and an
app worth installing. Verified against the source and the generated
manifest rather than guessed.
Four blockers, and none of them is porting work. The manifest requests
no storage or media permission at all, so the app can read no music --
and READ_MEDIA_AUDIO would not be enough, because it grants access
through MediaStore while this app's whole model is absolute paths:
audio_files.file_path is the primary key of ownership and every
GetFilePathsBy... query exists to hand paths to the player. The
first-run wizard calls DirectoryPicker, which Wails documents as
returning an error on Android, and the wizard intercepts pointer events
until a library exists, so the app is inert rather than merely empty.
mpris_linux.go is compiled in, because android implies linux. And the
scaffold's foreground service is typed dataSync rather than
mediaPlayback, with no MediaSession and no audio focus, so playback
dies at screen lock and there are no lock-screen controls.
They are all the same question: is the Android app a librarian or a
player? The desktop app is a librarian -- it scans folders, dedupes
covers, rewrites tags on disk -- and that model rests on owning a
filesystem, which is exactly what Android declines to give. So the plan
argues that parity is the wrong target and lays out three coherent
products instead, recommending a MediaStore-backed player.
Four things are worth doing whatever is decided, and the highest
information-per-minute one needs no code: run the published APK on a
real phone. Nothing in sections A or B has been observed on Android,
because the x86_64 emulator cannot run the app and emulator 37 refuses
arm64 images on an x86_64 host.
Emulator 37 refuses cross-architecture emulation outright -- "Avd's CPU
Architecture 'arm64' is not supported by the QEMU2 emulator on x86_64
host" -- and there is no flag for it. Google dropped it.
That matters because the previous commit's finding points at arm64 as
the ABI that works, so the obvious next move is to boot an arm64 AVD,
and the obvious next move costs a 3.8 GB download before it fails.
Written down so the next session does not spend it.
The consequence is stated rather than hidden: the claim that arm64
avoids the seccomp trap rests on reading modernc's two code paths, not
on having run it. Verifying it needs an arm64 host, a physical device
or adb connect.
Two harness bugs and the finding that exposed them.
**`pidof` exits 1 when it finds nothing**, and under `set -e` a failing
command substitution killed the script before it could print anything
-- rc=1, no output. That was invisible for as long as the app
crash-*looped*, because there is always some pid in that state. It
appeared the moment the app died for good and ActivityManager stopped
respawning it, which is precisely the run you most want output from.
**And an install failure said nothing useful.** Both ways it fails are
about identity rather than the build: INSTALL_FAILED_VERSION_DOWNGRADE
when a bare `make android` (versionCode 1) meets something a versioned
build left behind, and a signature mismatch when a debug-signed local
build meets a release-signed one. Both were hit in one session, and
both are fixed by uninstalling. The target says so now instead of
leaving someone to read the constant name.
The finding: with the startup bug fixed the app reaches the database
and takes SIGSYS on the x86_64 emulator, because modernc.org/libc's
Xlstat64 issues a raw lstat syscall on linux/amd64 and Android's
seccomp filter forbids it -- bionic never issues it. arm64 has no lstat
syscall at all, so ccgo_linux_arm64.go routes Xlstat through fstatat
and is structurally unaffected; Go's own syscall package already used
fstatat on both.
So the default emulator cannot verify this app, and the skill says so
rather than letting the next session read a tombstone as a regression.
CLAUDE.md said `wails3 task common:update:build-assets` regenerates
build/ios/ and build/android/. It does not: in beta.8 that command
extracts only updatable_build_assets, which is darwin/ios/linux/windows,
and the android tree comes from `generate build-assets`. It also said
nfpm's homepage and license are left alone by the refresh -- a comment
in that file says the same -- and a refresh reset them to wails.io and
MIT. Both corrected, and the CI section now describes five workflows.
NOTES.md gains the measurements: what cross-compiles and what does not,
the emulator environment, the Wails Android documentation's own two
errors, and the one line that stops the app at runtime --
buildUserDirPath switches on runtime.GOOS and Android takes the default
branch returning errUnsupportedOS, so main() calls os.Exit(1) six
milliseconds after the JNI bridge comes up.
The fix is a documented, build-tag-free API:
application.Mobile.StoragePath() returns the app's private files
directory and returns "" on desktop, and resolveUserDirPath already
lets YJ_HOME override the path on every OS. Deliberately not taken here
-- plan 015 is a pipeline, not a port, and the larger question it does
not answer is that open-directory dialogs return an error on Android
while this app's entire first run is "choose your music folder".
21 conflicts, all from the same cause: three features were developed on
both lines and this branch's copies are the ones adapted to v3's
bindings and to the file-shaped schema. Resolutions:
- `frontend/wailsjs/` stays deleted — v2's generated bindings, replaced
by `frontend/bindings/`.
- remove-from-library, `library-status.ts`, the requested-badge spec and
its component test: took this branch's copies, which differ from
main's only in calling `pruneEmptyEntities`/`CountAudioFiles`,
importing `@go/download/models.js`, and staging a real UUID for the
catalog's `CHECK(length(mbid) = 16)`.
- `GetFilePathsByRecordingMBIDsByLibrary` dropped: it joined
`recordings`, which no longer exists, and `library_id = 0` answers
both scoped and unscoped now. `GetAudioFilesByPaths` was already here.
- The album page, the artist page and the library badge kept this
branch's versions, which supersede main's: ownership asked once from
the files, the partial-completeness ring, and the request action.
- Docs: no migration chain (013) over main's two-file column rule and
its pre-1.0 squashing note, both of which 013 retired. Kept main's
`CreateSmartPlaylist` read-pool example, which is a real second
instance of that bug.
Verified on the merge result, not on either parent: lint clean in all
three build configurations, `make test` green in all three, 776 Vitest
tests, `tsc --noEmit`, bindings-check and skill-check clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AfVYUVExXsx1nSWrXN8mAh
Plans 013 and 014, the album page that prompted them, and the smaller
fixes they turned up. Changelog, largest first.
## The local library is shaped like files, not like MusicBrainz
`audio_files` carries its own tags and points at `albums` and
`artists`; `file_genres` is the one real many-to-many. `recordings`,
`release_group_recordings`, `artist_credit`, `artist_credit_artist`,
`recording_genres`, `release_groups` and `release_to_rg` are gone from
the local side, and with them a six-way join in every read, a
`MIN(release_group_id)` subquery in eleven queries and a
first-credited-artist subquery in nine. Measured on a real 25,966-file
library, every many-to-many that model expressed was 1:1 in the data.
- Ownership is a file. `GetFilePathsByRecordingMBIDs`,
`LibraryMBIDIndex.CheckMBIDs`, `collectLibraryEntities` and
`pruneStaleLocalCrossReferences` all join `audio_files`, so the 812
orphaned recordings, 216 release groups and 260 artists that library
carried are now structurally impossible.
- One projection: every track query selects from the `track_metadata`
view, one row type, one mapper. Nine hand-rolled copies had drifted
far enough to report different years on different screens.
- `library_id = 0` means every library, so each list query exists once
instead of scoped and unscoped with a branch at every call site.
- No migration chain. `sql/schemas/` is the one description of the
shape; `sql/migrations/`, `applyMigrations` and `schema_migrations`
are squashed away, along with the drift between them that had sqlc
generating against a stale schema.
- `database.InsertTestTrack` is the one test seeder; twenty test files
had been assembling the old FK chain each in its own order.
## The catalog stores its ids as bytes
`explore_index`'s three 36-char MBID columns and its entity-type text
are 16 raw bytes and a small integer. The table and its six indexes go
780 MB to 405 MB on a real 2,052,200-row catalog, which is why a fresh
install is ~0.6 GB rather than ~1.0 GB.
- `backend/explore/mbid.go` is the only place the encoding is known;
everything above it speaks dashed strings.
- `CHECK(length(mbid) = 16)` makes a stringly write fail at the insert
rather than silently returning no rows, since SQLite does not coerce
between TEXT and BLOB.
- The importer asks the artifact what encoding it carries and converts
on the way in, so the artifact already published keeps working and no
format bump is needed.
- `indexRowColumns`/`scanIndexRow` replace four copies of a 22-column
list, and `TestStoredEncodingRoundTrips` sweeps every read path.
## An album page that says how much of the album is yours
- One question, asked once: is there a file. `filePaths` is filled by a
single batched lookup when the tracklist settles, and the badge, the
Play count, the dimmed rows and every menu item read it — replacing
four claims of decreasing confidence that could show a green tick on
an album whose every action did nothing.
- Play, Play 7 of 12, or no play button at all.
- `total_tracks` on `explore_index` (~2 bytes over 400,677 release
groups) and on `audio_files` from tags that have always carried it:
a complete MBID-matched album now makes no catalog call at all, where
it used to spend the most expensive request the app makes.
- A merged cluster shows the running order the most releases agree on,
and the version list marks the release you own rather than standing a
synthetic entry in for it.
- `AlbumReleasesFailed`: a slow fetch is no longer reported as a failed
one by a 12-second timer.
- Rows not in the library are dimmed in place (with `aria-disabled`)
instead of the owned ones wearing a green tick and a legend.
## Caches and cover art get ceilings
- Only the three tiers of a cover are stored; the full-resolution copy
nothing rendered was 1,134 MB of a 1.4 GB covers directory.
- One artist portrait is downloaded and the rest are remembered as
URLs — 4.1 GB of a 5.3 GB cache was candidates no code path reads.
- `browsedArtBudget` and `httpCacheBudget` bound what an age cannot:
the same install held art for 5,770 artists in a 1,301-artist
library.
- `OrphanedArtistImagesJob` joined a bare MBID onto a sharded
directory, so it deleted the rows that were the only record of the
files it left behind. `explore.ArtistImageDir` is that layout's one
definition now.
## The autotag queue asks whether there is work
`tagging_items` was a row per album folder, not a queue, and no query
read the `tag_status` column that held the answer. The four queue
queries ask the files, which matters most where it is least visible:
`startPrefetch` was scoring every album in a tagged library against
MusicBrainz.
## Phantom playlist tracks resolve in place
An M3U8 imported before its files leaves phantom rows; they now match
by path and fall back to position, keep their place in the playlist
when resolved, and pair best-first so two phantoms cannot claim the
same file.
## Playing a track plays the list it is in
Double-click, and Play on a single row's menu, queue the list as
displayed with `startIndex` on that row — the album page and the track
list used to queue one track and discard the album around it. A
multi-row selection still plays exactly itself.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AfVYUVExXsx1nSWrXN8mAh
CLAUDE.md gains a Packaging section for the four Taskfile facts the
recipes just needed — wails3 on PATH by bare name, no -ldflags on
`wails3 build`, bin/ not build/bin/, and bundling as its own step —
plus how build/'s platform metadata generates from build/config.yml and
what that refresh overwrites.
Its lifecycle, bindings, harness, events and CI sections were still
describing v2. The events one matters most: the rule to emit through
events.Emit survives, but its justification is now the weaker one, and
saying so is the point of the migration. v2's runtime.EventsEmit
log.Fatalf'd on any context not carrying the runtime; v3's emit takes
no context at all, so what is left to pin is that one emit path is what
lets emitStatus drop an unchanged payload for every caller at once.
README told a contributor to `go install wails/v2/cmd/wails` and
apt-get libgtk-3-dev/libwebkit2gtk-4.1-dev; the CLI is vendored and the
stack is GTK4 + WebKitGTK 6.0. Two comments claiming Xvfb and one
claiming frontend/wailsjs go with them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
`make sandbox`, `make dev`, `make build-dev` and `make build-prod` all
died with "/bin/sh: wails3: command not found". `wails3 dev` and
`wails3 task` are supervisors: they run the scaffold's Taskfile tree,
which invokes `wails3` by bare name in 54 places across four files. The
CLI is a vendored Go tool by design (plan 009, D3 — a global install
would be this build's first undeclared dependency), so that name did
not exist.
scripts/toolbin/wails3 execs `go tool wails3`, and the Makefile
prepends that directory only for the targets that start a supervisor.
Rewriting 54 scaffold call sites would be churn to redo on every
scaffold refresh; nothing global is installed either way.
The shim does not cd. The first version did, to be sure `go tool` found
the module — it does not need to — and that silently discarded the
`dir:` a task had set, so generate:icons failed with "open
appicon.png: no such file or directory" against a file that was there.
Three things the build path needed once it got that far:
- `frontend/package.json` gains `build:dev`, which build:frontend runs
under DEV=true and which did not exist.
- Vite binds 127.0.0.1. It defaulted to `localhost`, which resolves to
`[::1]` only here, while wails3 dev's asset proxy dials IPv4 — so the
first request for the dev server was refused and the first paint
raced a retry. Zero proxy errors after.
- The icons and the .desktop file are generated on every build.
icons.icns/icon.ico are deterministic from our appicon.png (verified
by regenerating), so the regenerated pair is committed and the churn
ends; .task/ and the .desktop file are ignored.
Also corrects a claim: build-prod strips and trims but does **not**
UPX-compress — that was v2's `-upx` flag. Phase 1 recorded UPX as
still working, but neither build target had been run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
Phases 2 and 3 of plan 009, plus the parts of phase 1 that could not
land before them. Nothing in the tree imports wails/v2 any more; all
three lint and test configurations are green and `go build .` produces
a running binary.
The point of the migration is one file. backend/events/emit.go probed
ctx.Value("events") — a v2-*private* context key — to decide whether
emitting was safe, because runtime.EventsEmit called log.Fatalf on a
context without the runtime and took the process down with it. v3's
emit takes no context, so that is now application.Get() == nil. D1
held: events.Emit keeps its ctx as the WithSink test seam, and all 45
call sites and 7 test files are untouched.
The bootstrap splits into application.New + Window.NewWithOptions +
Run. Ten bound services implement ServiceStartup instead of being
handed a context by hand from OnStartup, which also stops ten
SetContext methods being exported as bindings. jobs.Registry and
explore.SearchIndex keep theirs — neither is bound, so converting them
would be churn for no binding removed.
Four things differed from the plan and are written up in it: GPU policy
moved to the per-window LinuxWindow options rather than surviving on
LinuxOptions; there is no OnStartup/OnDomReady option, so app-level
wiring hangs off ApplicationStarted; application.NewService is generic,
so FEBindings []any could not survive (the binding generator is a
static analyser and would have seen nothing); and the quit veto had to
be restructured, because v3's dialog answers on a callback rather than
returning the button, so ShouldQuit vetoes, asks, and quits again from
the callback.
Window state saving moves to a WindowClosing hook — the size has to be
read while the window still exists, and v3's OnShutdown has neither
context nor window. backend/logging is deleted rather than ported:
v3 takes a *slog.Logger directly, so the v2 logger.Logger adapter had
no caller left.
Phase 1's tail rides along, now that it can: the Makefile's wails
invocations, all 50 webkit2_41 sites, lefthook, both packaging recipes
and ci.yml's apt lists. v3 builds against GTK4 + WebKitGTK 6.0, which
Arch and ubuntu:24.04 both ship, so the tag is a deletion rather than
a translation.
Phase 4 is next and the branch is not usable until it lands: the app
builds, but frontend/wailsjs/ is v2's tree and nothing regenerates it,
so the frontend cannot reach the backend yet.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
The plan assumed build/ was free and that GTK4 was a preference. It
was not free — this repo used it as ignored build output — and GTK4 is
not available on the dev machine, which breaks `go tool wails3`
outright rather than merely changing which webkit is linked.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
Three things on the Explore surfaces, all about not asking twice.
A portrait already on disk costs no network call. explore-view seeded
only from the library store — owned artists, which on a catalog search
is nearly none of the results — and sent everything else to
GetArtistImageURL, the resolving entry point, one await at a time.
GetArtistImagesCachedPaths asks the disk about every unresolved artist
in one call, and only what it does not answer reaches the resolver,
in parallel.
The artist page's two sections both wanted PrefetchReleases and each
called it, so the most expensive call the app makes was issued twice
for an overlapping set on a 1 req/s limiter. They are collected and
sent once on a microtask, and prefetchRequested stops the cold-artist
refetch re-asking for what it already asked for.
The release cards — most of the artist page — had no context menu at
all. They have one now on both release shapes, normalised to a
ReleaseMenuTarget when the menu opens so the union does not reach the
action handlers. It is a discriminated union rather than one nullable
field per kind because the panel is shared with the track menu: that is
what keeps aria-label moving with the target, which is the fault
cover-grid shipped. Which items appear is three different questions —
playback is gated on a local album id, not on "owned", and the request
needs a catalog MBID, so it is absent for a library-only release.
Note on the docs: the CLAUDE.md and NOTES.md prose here was
reconstructed after a mishandled `git stash --keep-index` destroyed the
uncommitted originals. One NOTES.md section is marked as incomplete
where its text could not be recovered.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
RemoveFromLibrary deletes the audio_files rows the way the scan's own
orphan cleanup does and records each path in excluded_paths. The
exclusion is not an enhancement: without it the next scan finds the
file, sees no row and imports it again, so the button undoes itself.
The soft scan compares files on disk against rows in the database, so
surveyAudioFiles and countAudioFiles both take the exclusion set —
otherwise an excluded path makes the two disagree forever and queues a
full scan on every launch. Deleting a row cascades to queue_tracks, so
the removal calls the same CompactQueue hook RemoveLibrary does.
Also lands the requested badge: library-status-indicator is a button
again where it can act, utils/library-status.ts states once what owning
and wanting mean, and the long-declared queued state finally has a
producer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UDCbcCZQepnpSQYJ6SxxZm
The album page asked MusicBrainz how many tracks an album has, because
the only total it had was the length of the tracklist it was already
showing — a tautology for a library copy. The denominator was on disk
all along: metadata has read the "5/12" totals off every file since
forever and discarded them. They persist to
release_group_recordings.total_tracks now, and a complete, MBID-matched
album makes no catalog call at all.
Around that:
- AlbumReleasesFailed, so a slow browse is no longer reported as a
failed one. The page inferred failure from a 12s deadline, against a
browse queued behind up to eight prefetches on a 1 req/s limiter.
- Tracks not in the library are dimmed in place rather than the owned
ones carrying a green tick, which is also what let the "loading
catalog" banner go.
- A partly-owned album draws the release, not the part, so the missing
tracks are visible and Play can say "9 of 12" truthfully.
- The version dropdown appears only when tracklists actually differ,
and the version you own is marked by name instead of being replaced
by a synthetic "Your Library" entry.
- A merged cluster shows the running order the most releases agree on,
not whichever pressing the browse returned first — which is what made
a correctly matched album claim it was unlinked from MusicBrainz.
Also carries in-progress work from earlier sessions that shared these
files: the queue source link, autotag mixed-bag grouping, the mix
feature and its schema, and the config general page.
Committed with --no-verify: every pre-commit check was run by hand and
passed, but bindings-check refuses to run while frontend/wailsjs is
dirty and counts *staged* as dirty, so it cannot pass on any commit
that updates the bindings. Verified separately by regenerating and
diffing against the staged content.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NSmYeXS3k9xw3MnMPoCjvP
Two of Phase 2's three judgement calls were answered by reading the
code rather than by choosing: there is no artist badge to make a
button, and a track badge stops reading as noise the moment it means
something. The third went the other way — `EntityRecording` reads like
a placeholder and is real work.
The plan's own framing was wrong in a way worth keeping: the badge was
not waiting on the download client, which had largely landed already —
it was waiting on somebody looking at a state nothing produced.
Plan 008 is complete and moves to completed/. The two findings worth
carrying forward are that a new table needs one schema file rather than
two (and a datamap entry, which is a gate nobody remembers), and that
excluding a path has to reach every place that counts what is in the
library — the soft scan's disk-vs-database comparison above all, which
would otherwise have rescanned the whole library on every launch with
nothing failing anywhere.
Phase 3 shipped in six landings and a11y.md is closed, which closes all
four audits from 2026-08-11.
The pass's one lesson is that an accessible name is computed on the
element carrying the role, and every tier we check with looks somewhere
else: the audit read the source and credited a name that was never
computed, an AX sweep read the tree and reported a placeholder-only box
as clean, and a component test asserted the attribute and pinned the
bug it existed to prevent.
Six of the audit's claims turned out to be wrong or smaller than
written, and one of the plan's own findings was false — the page
header's sort control is named on all nine views. All of them are
written down, which is where a third of the value of the last two plans
came from.
The two findings recorded as too big for the contrast pass are fixed, so
the plan says so. Also corrects a claim I made and did not check: the
chrome does not stay dark under the light ramp -- that screenshot was
taken before the theme propagated, which is the third time in two passes
a picture read at the wrong moment produced a confident wrong claim.
The audit's one 'borderline ~4.1:1' pair was nine of twelve failing
combinations across three ramps, 110 nodes on screen, worst 2.31:1. The
other never-measured item closed on measurement and stays dropped, now
for a reason with a number behind it. Two findings larger than either
are recorded and deliberately not fixed: the semantic colours are fixed
across ramps, and the light ramp is not a supported theme.
Three a11y findings shipped. The generalisation is the mirror of 'a
finding creates the conditions for the next one': that one is about the
code path a fix opens, this one about the path it sends people to. The
reduced-motion guard is two lines and both bugs behind it were in the
fallback it routes users into -- one of which had been wrong in every
mode, including the default, since the component was written.
All six phases of 007 shipped. The plan moves to completed/ with a recap
rather than a rewrite: its seven "where the plan was wrong" lists are
seventy-nine entries and about a third of them are the audit being wrong,
which is the material 008 is planned against.
008 is a11y.md, the only audit with open items and the least verified
material in the repo. A grep pass closes at least five findings the
coverage map still shows open, including a11y.7, which the map assigns to
phase 6 and which phase 1 fixed. The triage in the plan is recorded as
hypotheses for that reason.
Twelve corrections to a plan written before any of phases 1-5 existed,
of which the load-bearing one is that a rule written against a
mechanism does not cover what the rule is for: `home` suppresses a
repeated shelf by comparing album ids, Explore's first two shelves hold
different entity types and share none, and the page repeated itself
anyway because a person reads artists.
Each of perf.p2, H-13 and the dialog naming had a second defect behind
the one named, reachable only once the first fix made the code path run.
Also records two probe failures worth more than the fixes: the a11y
snapshot cannot see a dialog's accessible name at all, and a scroll
assertion that could not fail was hiding both a bug and a false claim.
The e2e job passes on both engines for the first time, so the three
files that describe it as red are wrong. Also records the two things
that made it findable: the CI container is reproducible under Docker,
and the app's own audio stack had to be the thing measured.
The e2e failure two sessions could not diagnose is the container's
audio clock, on both engines — 48 specs pass under Chromium and 48
under WebKit, failing the same three. Nothing in last pass's dialog,
focus or role work is WebKit-specific.
Also records what got in the way of knowing that: gitea_ci's job-log
endpoint 404s on this build while the REST API answers fine, and the
WebKit step had been skipped on every red run.
The suite passes locally on Chromium; CI also runs WebKit, which cannot
run on Arch, and this pass changed focus management and dialog
modality. The job-log endpoint is not exposed by this Gitea build and
the runner is not on this machine, so the WebKit half is unverified
rather than attributable to the known audio-clock flake.
Three of a11y.md's findings describe a build that no longer exists —
one fixed by a phase that was not about it, one whose stated mechanism
stopped being true when Phase 4 bundled the icons, and one that
reproduces as a different shape. The generalisation is that a finding
has a date as well as a magnitude and a mechanism.
Also records the two bad versions of the duplicate-shelf rule that the
*existing* tests caught, the eleven e2e specs that landing on Home
broke and the one of them that was a real bug, and the second CI e2e
failure on a commit that changed no application code.
Two reproductions in this pass were read before Lit had rendered, so
both reported the same answer on the broken build and the fixed one -
the third costume of this plan's most-repeated trap, and the first time
it has appeared in a reproduction rather than a measurement. Also
records that the audit's symptom for H-11 outlives its mechanism, that
fixing H-7's arithmetic does not remove every clipped Duration, and
that two e2e specs spend backend state they never give back.
`.planning/audits/2026-08-11-ui/` is the pass this work came from: the
app driven by hand headless plus three static reviews, ~118 findings
that are really five problems, each spread by being copied rather than
fixed. `.planning/plans/active/007-ui-reconciliation.md` sequences them
by blast radius and records what each of the six passes actually
shipped — including twenty-five entries under "where the plan was
wrong", which is the point of writing it down.
The discipline those entries add up to, now in NOTES.md: a finding is
three hypotheses — how big it is, why it is that big, and what to do
about it — and they can be independently right and wrong. Three of the
audit's recommended fixes would have shipped a bug (`m1` stops the
card grids repainting, `m6`'s index-ordered selection goes stale on
any re-sort, `m5`'s guard leaves the marquee short), all three because
they reasoned from the shape of the code and not from what the rest of
the file already knew about it. Five findings evaporated or inverted
on contact.
CLAUDE.md gains the invariants that came out of it, and the skill
gains the fourteen measurement traps, each of which produced a wrong
number first — the newest being that a longtask entry arrives after
the task that produced it, so two numbers that must agree are worth
more than one you have to be sceptical about.
CLAUDE.md gains backend/home and the two cross-cutting frontend pieces
a list or detail view now has to know about: explore-link's
always-navigate rule with its double-click grace, and
<catalog-scope-notice> with the catalogPending/catalogLoaded
distinction behind it.
main.go embeds frontend/dist, so lint, test and bindings-check all fail
on a fresh clone until pnpm build has run. Invisible locally because
anyone who has started the app has a dist/ lying around, and the
container prototype missed it because both job scripts shared one
mounted directory, so job 1 consumed a dist/ that job 2's dev-headless
had built on an earlier run.
A coding agent could develop this repo's Go packages and could not
develop the application: every path to running YellowJacket ended in a
blocking GTK window, so 265 bound methods, 46 events, 33 component
directories and 13 stores had exactly one form of verification
available — `tsc --noEmit`.
The unlock is that `wails dev`'s dev server on :34115 serves the real
frontend with the real generated bindings against the same Go backend a
desktop window attaches to, so a plain Chromium under Xvfb gets a fully
functional app. Four test tiers now exist, cheapest first:
- `make ui-test` — 313 Vitest tests in a real browser in ~2 s, no app,
no backend, no display. Works because `frontend/wailsjs/` is a pure
passthrough to `window.go`/`window.runtime`, so faking just those two
globals runs the real bindings and the real store code.
- `make test` — services in-process, asserting on the payload the
frontend would receive, via a new `events.Emit` wrapper.
- `make dev-headless` + `playwright-cli` — the real app, driven
interactively, with an event bridge on `window.__yjEvents` and a
dev-only control surface at `/__test/`.
- `make e2e` — 19 of those flows frozen as Playwright specs.
`events.Emit(ctx, …)` replaces all 35 direct `runtime.EventsEmit` call
sites: wails' `getEvents` `log.Fatalf`s on any context without its
runtime, so those paths could not run under test and a background
worker could take the app down. Four packages had each hand-rolled the
same guard; nine more guarded on `ctx != nil`, which does not help.
`TestNoDirectRuntimeEmits` fails the build on a new one.
Fixtures are generated, not committed (`make testdata`), and seeds are
built by *running the app* — never by hand-writing config and DB rows,
which would be a second description of a valid YJ_HOME.
`.gitea/workflows/ci.yml` is the first workflow here that tests
anything; the other three only package, so `gitea_ci` reported only
packaging jobs and misled anyone asking whether a push was healthy.
Both jobs were prototyped to green in a bare ubuntu:24.04 container
before the YAML was written, which immediately caught `make lint`
linting three configurations that nothing builds: all three passes
omitted `webkit2_41`, so wails resolved webkit2gtk-4.0 — which Arch
still ships and Ubuntu 24.04 dropped.
Operational instructions live in `.pi/skills/yellowjacket-dev/`,
measured discoveries in `.planning/NOTES.md`, and architecture in
`CLAUDE.md` — split by tense, not by topic, because a topical split
gives every new fact two plausible homes. `make skill-check` fails a
commit if the skill cites a make target that does not exist.
Autotag: detect "junk drawer" folders with no artist/album consensus
and split them into synthetic per-cluster groups instead of forcing
one match on an unrelated pile of tracks; repair tagging_items rows
left behind by a prior scan orphan-cleanup gap.
Explore: fix an exact artist-name search being drowned out by its own
catalog entries in intent-prior scoring, and prune stale in_library
bookkeeping left behind when a referenced library row is deleted.
Download: fix a multi-library regression where every import failed
with "no library root configured" — the importer resolved the
library root from a legacy single-library config field that nothing
populates in the current multi-library model. It now resolves the
destination library per-request from the request's own library_id.
Also widen the Soulseek search window (12s -> 20s), measured against
real request history to be missing available peers on live queries.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y2Agd9af5hE7qzti2ackiS
Ships the fresh-start schema cleanup: rebuilt explore catalog index
pipeline (dump import, artifact fetch/build, incremental listen-count
refresh), a new download subsystem (Lidarr/Prowlarr/qBittorrent/SABnzbd/
slskd/yt-dlp providers, staging, reconciliation, wanted list), and the
supporting schema/query/store changes across backend and frontend.
Also includes two smaller follow-ups: bump the central index's
rebuild-after cadence from 90 to 180 days, and remove the Explore
"library only" online/offline toggle entirely (frontend-only, no
backend counterpart) rather than carry unused UI/state.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y2Agd9af5hE7qzti2ackiS
Add a central job registry that library scans and search index builds
report into, so background work is visible instead of buried in the
settings page.
- backend/jobs: registry with per-job ring-buffer logs, capability-driven
controls, and one coalesced JobsChanged snapshot at 4Hz
- pause survives restart via a job_state table; a paused scan is adopted
back on launch and skipped by the soft scan
- top-bar indicator, popover, details drawer and a Jobs page replacing
the config page's scan UI; per-library start/stop retained
- scan timing breakdown moves into the job log, Full rescan to the Jobs
page; delete the orphaned library-manager component
Also add cmd/indexbuild and cmd/indexexport so the explore index can be
built once centrally rather than by every install, which today streams
~205GB from the ListenBrainz spark dump on first run. indexbuild picks
build/refresh/rebuild from index state; the Gitea workflow runs it on
push, weekly, or manually and publishes only when content changed.
fresh-install no longer defaults YJ_HOME under /tmp: it is tmpfs on most
distros, and the import needs ~6GB of real disk.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Consolidates in-progress work across autotag, explore, and library:
- autotag: beets/Picard-informed scoring engine — ID-first matching, VA
handling, recommendation tiers, and a merged distance/rank cascade, with
an eval harness for regression tracking.
- explore: offline MusicBrainz dump import/incremental refresh replaces the
legacy tier crawl; index-first local search with fuzzy matching and a
dedicated ranker; disk-free guards for dump downloads.
- library: artist-credit extraction and matching.
- lyrics: owned-library lyric search (FTS) with LRCLIB backfill.
Also: rewrite README to be user-focused, and migrate upstream to
git.ljones.me/yonlu/yellowjacket.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
logan
yonluandClaude Opus 5