Adopting the six v0-era PRs surfaced two conflict-shaped cases the merge
leg handled only by luck. Behind-main branches are refused outright by
the repo's block_on_outdated_branch protection, so the leg now refreshes
every branch against origin/main before merging — which is also where a
textual conflict should surface, as diff text the loop resolves only
where it authored the hunks, otherwise abandoning the PR to a human
with a comment. And the one guard no mergeability check provides is the
push run on main after the merge: three PRs touching the same file can
merge cleanly and contradict each other, so a red main now halts the
loop instead of the tick reporting merged and moving on.
Closes#238
The scheduled backlog runs already claimed, fixed, verified and opened
PRs one issue at a time (~50 runs), but stopped at "PR open, CI green" —
every merge and every stale branch was human work, and the pipeline
shape existed only as one prompt file. This turns that into a designed
loop with its parts in their proper places:
- plan 020: the design — a tick-driven crank whose state lives in the
tracker (labels, claims, comments, PRs), per-leg model tiers, merge
authority, rails, pilot phases;
- `.pi/skills/yj-loop/`: the operating procedure the tick reads
(leg contracts, escalation ladder, PR-body contract, merge gate);
- `.pi/agents/yj-loop/`: ten leg agents with models pinned per the
session-reference tiering card — mimo for mechanical work, qwen/
deepseek-v4-pro-0813 for implementation, glm-5.3 for selection,
planning and consequences review, glm-5.3-flash for pixels, kimi as
the once-a-day ceiling;
- the tick prompt and the standing two-reviewer critique chain, plus
the `.pi/loop/` gitignore entry and the CLAUDE.md pointer.
The switch stays where the v0's was — `.pi/schedule-prompts.json`,
gitignored, live only while the loop's pi session is open. No code
changes. Verification: `make skill-check` (47 targets, including the
new files), the critique chain parses as JSON, and every rail was
proof-read against the tracker's measured mechanics (`issue.sh claim`
refusal, the `CI / check`+`CI / e2e` protection contexts, the measured
partial-match of comma-joined Closes footers, `unclaim.yml`).
Closes#236
Plan 019 phases 3 and 4, which finish #63. The queue panel and both
playlist detail views get tap-to-play and hold-to-select; the playlist
views get swipe-to-queue as well.
Phase 3 was not the pure wiring the plan expected, in two places.
A tap on a queue row plays that position. Copying track-list's tap --
which sets the queue to the list the row is in -- would rebuild the
queue from the queue, discarding its source, its shuffle order and
anything inserted by hand. It reads as a no-op and is not one.
And the queue panel has no swipe, deliberately. A right swipe means add
to the queue everywhere else it exists, and a queue row is already in
the queue; the only thing it could mean there is remove, which is the
same gesture with the opposite effect one screen away. Removing a queue
row is on the row, on its sheet since #60, and now on its selection
bar. The assertion is that its rows do not opt in.
The reveal became utils/swipe-to-queue.ts rather than being copied into
three lists, keyed on a data-swipe attribute so one stylesheet carries
the touch-action half of the device fix to rows that are called two
different things.
Phase 4 was already true and is now asserted: a claimed tap has its
click swallowed, so an explore-link inside a row never sees one and
tap-to-play wins with no rule of its own. Its test was vacuous when
written -- the tap helper sent no click, so there was nothing to
swallow -- which also weakened phase 1's. It sends one now.
Escape leaves selection mode, from selection-bar rather than from each
of the four hosts, since that element exists only while the mode does.
The platform's back gesture deliberately does not reach it: the shell
owns the history stack and four lists reaching for history is four
stacks. That is #200.
Verified on the reference phone: a queue row taps to its own index and
refuses a swipe, a playlist row queues on a swipe and plays its
playlist on a tap, and a hold raises the bar without the menu.
Closes#63
Plan 019 phase 2. A finger on a track row now drags a reveal out from
under it and queues the track on release, with the affordance saying
what it will do before it does it.
Two things the device said that the plan did not predict, and both
change the implementation rather than decorate it.
The gesture runs on touch events, not pointer events. Chrome 113's
WebView cancels the pointer stream ~16px into any drag whatever
touch-action says -- measured at auto, pan-y and none alike -- while
touchmove keeps firing. So touch-action: pan-y is half the fix and a
non-passive touchmove calling preventDefault is the other half, and
neither works alone: with the preventDefault in place and touch-action
back at auto the gesture died after one move. Both are correct in
Chromium either way, which is why the module's header carries the
measurement and the component tier asserts the stylesheet.
And a phase 1 defect the device found on the way past: the native
contextmenu arrives in either order and only one was handled. Our
500ms timer firing first, a component claiming it, and Chrome
delivering its own menu 50-70ms later was suppressed by nothing -- so
the context menu opened over the selection bar, two holds in four, on
the one surface this issue exists to have changed. Six holds clean
after.
draggable="true" is not a competitor: no dragstart fires from a touch
drag on this WebView at all.
Phase 1 of #63, and the design the issue asks for as one piece is
.planning/plans/active/019-android-touch-model.md.
**A finger has no second button and no modifier keys**, so the primary
action has to be the primary gesture: tap plays the row, and the hold
that opened a context menu now enters selection mode with that row
selected.
Three decisions in it, and two diverge from the report.
**The predicate is the pointer, not the platform or the viewport.**
`pointerType === 'touch'`, per event, which is already how long-press.ts
decided and is the only such test in the frontend. This is #64's rule --
named after the capability -- and it carries #64's warning: keyed on a
width, an Android *tablet* at 600px gets click-selects/double-click-plays
on a touchscreen, which is the inversion this issue exists to fix, on
the platform it exists for. A touchscreen laptop cannot be described by
a width at all. Per event, a mouse keeps desktop semantics on the very
same row, and there is no second declaration of what a phone does.
**There is no double-tap, and the number is why.** The report asks for
single tap to play *and* double tap for the menu. Those cannot both be
honoured: the first tap of a double tap is indistinguishable from a
single tap until the interval expires, so "tap plays" becomes "tap
waits". Measured on the device, the play command to TrackChanged is
155/123/85/56/91 ms -- median ~100 -- and the app's own
DOUBLE_CLICK_GRACE_MS is 250. That is 3.5x the primary interaction,
250ms of it spent deliberately doing nothing, on every track anyone
plays, to reach a menu the hold already reaches. So the menu and the
selection action bar are the same surface, which is also the platform's
convention and removes a concept rather than adding one.
**Tap-to-play and selection mode ship together**, because splitting
them is a regression dressed as an increment: a touch user selects by
tapping today and acts through the long-press menu, so moving tap to
play on its own would leave a window with no way to select forty tracks
at all.
**What lets this reassign the hold without touching one of the fourteen
context menus**: the layer announces `yj-tap` / `yj-long-press`
(composed, cancelable) and acts on nothing. A component claims one with
preventDefault. An **unclaimed long press still becomes a
`contextmenu`**, so the card grids, Explore, the playlist rows and
every other menu behave exactly as they did, and only lists that opt in
get selection mode. An unclaimed *tap* does nothing at all and the
click follows normally, which is what leaves every button in the app
alone -- only a claimed tap has its click swallowed, or playing a track
would also select it.
**And the device found the one thing no browser tier can see.**
Chrome 113's Android WebView fires its own `contextmenu` on a long
press. long-press.ts stood down when a trusted one arrived, which was
right while both paths ended in a context menu; they no longer do, so
standing down means the gesture silently does the *old* thing.
Measured, before the fix, holding a track row:
{"log":["contextmenu isTrusted=true"],
"state":{"bar":null,"menuActive":true,"selected":1}}
`yj-long-press` was never announced, the menu opened, and all 26 tests
passed -- dispatched pointer events do not make a browser synthesise
one. So the native event is a **trigger, not a competitor**: the
gesture is announced from it and only a claim suppresses it. Unclaimed
it propagates untouched, which is the same "browser wins" outcome
reached by asking instead of assuming.
The tier could not find that and can hold it, because this module has
always told its own events apart by identity rather than isTrusted, so
an untrusted one from a test takes exactly the browser's path.
Verified on the device by *performing* the gestures rather than
describing the page -- `adb shell input tap` and `input swipe x y x y
700` reach the WebView as real pointer events, which is new here and is
written down in the plan with the pixel mapping. Tap plays; a hold
raises the bar with one selected and no menu; a tap toggles to two,
back to one, and the mode ends with the last row; an album card still
opens its context menu.
29 new tests. The e2e spec is rewritten to assert **both** halves --
the row selects, and a card elsewhere still opens the real menu --
because a spec that only checked the row would pass on a build that had
silently broken the other thirteen.
Phases 2-4 (swipe to queue, the other three surfaces, and what #67
inherits) are in the plan and not in this commit.
The `page-header` paragraph already stated "the header asks for a sort,
it does not perform one"; actions now follow the same division and it
belongs beside it — the header decides what fits, the host decides what
happens.
Plan 018 moves to completed/ because #69 was the last thing it owed:
its size matrix promised "no action is ever unreachable at any
supported size" and the residual 114/162px clip was that promise
outstanding. Its recap also corrects a claim the plan made — the queue
and the actions were not the only two things competing for the header's
width, since every child of that flex row was flex-shrink: 0 and the
actions come last.
CLAUDE.md gains the three bands as a promise (Phone <600, Compact
600-899, Desktop >=900, and "no action is ever unreachable at any
supported size"), the computed queue rule and why it cannot be a media
query, and the correction that 900 — not the 800x600 minimum — is the
worst desktop width.
NOTES.md gets the measurements, including two things worth more than
the fix. My first probe for the sidebar's scroller searched
shadowRoot.querySelectorAll('*') and reported "no scroller, items are
unreachable", which reads exactly like a live Settings-unreachable bug;
the scroller is the host, and a host is not inside its own shadow root.
And the plan's first draft claimed the overlay "removes the desktop
half of #69", which the screenshot disproved: open and closed are now
identical at 900x600, so the queue's contribution is gone, but the
header's own overflow remains and is still a live defect.
Refs #24
#24 asks for a design pass, and #73 hangs the rest of Phase 2 off the
answer, so the decision is written down before any CSS moves.
Measured against the running app, and five things are not in the issue:
the Playlists header clips at 800x600 with the queue *closed* — the
minimum window is the only size this app promises; 900x600 is worse
than 800x600, because the sidebar expands at 900, so the worst desktop
case is not the minimum and every test that stops at the minimum misses
it; at 320px with the queue open the main panel is 0px wide, because
the panel is in the flow rather than over it; only Playlists overflows,
so #69 is one view's action set and not a systemic header failure; and
both reasons in MinWidth's comment describe mechanisms that no longer
exist.
The queue's mode cannot be a media query: its width is drag-resizable
between 200 and 500px and persisted, so a fixed breakpoint assumes the
default 320 and is wrong by 180px in the direction that hurts. It is
computed from the measured widths instead.
#69 stays its own PR on a finding rather than an estimate: page-header
cannot collapse actions that arrive as arbitrary light-DOM markup
through a slot, so the fix needs an actions API across all three hosts.
A very small window becomes the phone layout, which already exists and
is already tested, rather than the mini-player: #12 is a second
always-on-top window, and making it a mode of the main window would
discard navigation state on a resize and put the process-level MPRIS
question on a path a drag can trigger.
Work has been starting from a chat message and a plan file, so two
people could pick up the same thing and neither could see the other.
The tracker is where that is visible.
Search before starting, claim before the first edit -- not before the
commit, since the point is that the other person can see the work is
taken while it is being done. If no issue covers it, open one first:
that is what makes the tracker a description of the project rather than
a description of the past.
The conventions were already right and are written down rather than
reinvented -- the Kind/Area/Priority/Platform/Reviewed/Status taxonomy,
its exclusive scopes, #73 as the roadmap, real Gitea dependencies for
hard blockers, and PR #83's body shape.
What #83 also demonstrated is that a Closes list closes nothing
reliably: it listed ten and five of them sat open in main for a
fortnight. So closing is a step you take and verify, not a keyword you
trust.
.planning/ stops being a queue and keeps design documents and measured
history -- NOTES.md, the audits, the completed plans and the arguments
in them. plans/pending/ is gone, because a plan nobody is executing is
an issue; everything unimplemented in it is now #85-#91, and each
completed plan says which issue carries its remainder. autotag.md is
kept as a historical record, marked stale where the scoring overhaul
overtook it.
The commit grammar is unchanged and is load-bearing for a different
reason, so the issue number lives in the branch name and the PR body
rather than the commit subject.
Refs #92
Moves plan 017 to completed with a recap, and lifts the three findings
that generalise into NOTES.md: a preset major that renders empty notes
with everything green, a 403 that looks like branch protection and is a
token scope, and tag-triggered workflows running the tagged commit's
own definitions.
CLAUDE.md said .releaserc.yml was a config nothing ran and that there
were five workflows; both stop being true with this branch. The CI
section now names release.yml as the entry point and records the four
things in it that are load-bearing, including the two silent failure
modes worth pinning against.
packaging/homebrew/README.md and docs/android-release.md say where a
user would actually look that upgrading from 1.x needs a reinstall --
Homebrew offers nothing silently, and Android refuses outright.
A track credited to more than one artist has exactly one navigable
artist in this app and the rest are punctuation. `primaryArtist()`
string-parses the credit, strips a " feat. " clause and discards the
guest; it deliberately does not split on "&", "with" or "," because
those live inside real artist names.
Measured on a real 26,069-file library plus an 80+80 MusicBrainz
sample: 13% of recordings are multi-artist upstream, while only 0.86%
of files carry any structured multi-artist tag — mp3 carries zero
files with multiple MUSICBRAINZ_ARTISTID across 19,840. Of 1,286 files
saying "feat.", 90% have nothing structured behind it, and a sample of
80 such files was multi-artist in MB 80 times out of 80.
CLAUDE.md justified plan 013's removal of the credit tables with "3
credits of 2,823 listed more than one artist". That measured our own
*writer* — cachedLinkArtist was called once per credit, so a
collaboration could never have been recorded. Dropping the join table
was still right on cost; the evidence for "multi-artist is rare" was
not.
A credit is ordered parts and the credit string is derived from them,
so join phrases are assembly instructions, not disassembly ones.
Nothing here reconstructs a credit by searching a name inside a credit
string: the stored text may come from tags while the parts come from
the catalog, and those disagree for ~1 in 3 multi-artist credits.
Where it comes from, after two dead ends: the canonical dump CI
already streams has no join phrases and no as-credited names, and the
JSON dumps cover 153,691 recordings of ~35M with *zero* overlap
against a real library. So mbdump.tar.bz2 — 7.1 GB, ~13.7 min in
pure-Go bzip2, whose members are alphabetical, which is what lets one
pass resolve an entity's credit without buffering 35M recordings.
- artist_credit_part / artist_credit_ref, multi-artist credits only:
a single-artist credit is already explore_index's own artist_name.
- Column layouts verified against the real 20260815 export;
ErrDumpShape makes a wrong guess a failed build, not a wrong catalog.
- The pass runs on every mode, not just a build. The job picks its mode
from the index's own state, and a complete import means "refresh",
which never enters the importer — so credits could otherwise only
arrive via a rebuild that re-downloads ~205 GB. It reports whether it
populated anything, which is what flips `changed` and republishes.
- The importer asks whether an artifact carries the tables, on the
writer where `core` is attached, so the artifact already published
still imports.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AfVYUVExXsx1nSWrXN8mAh
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".
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.
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.
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.
logan
yonlu
