The button advanced one step round the untouched → doing → done cycle,
so pressing "Mark done" on a fresh event produced "doing it" and needed
a second press. Worse, the press after "done" cleared the status
outright rather than undoing anything.
Three states need three targets — that is the control directly above it.
This one is the commit, so it sets done and back. cycleStatus had no
other caller and is gone.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Two symptoms of one bug. The "next to expire" headline counted events
the reader had marked done or ignored, and the dailies strip kept a
tickable chip for a repeating event they had already finished.
showCompleted and showIgnored decide what a reader can *look at*. The
headline and the strip are *instructions*, so they answer a different
question — what is still on your plate — and both now go through one
`outstanding` lens.
Also fixes a second bug in the same line: `next` took the head of the
list, which under "doing first" sorting is whatever you're partway
through, not the soonest deadline. It reads the minimum now.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The one moment in this app that is unambiguously good news, and it read
the same as any other tick.
Fires only on a tick that completes the set — the list also shortens when
the reader filters, and a burst there would be congratulating them for
filtering. Nothing on mount either, so arriving at an already-finished
day is not treated as having just finished it.
Decorative, aria-hidden and under a second; the "All done" text beside it
is what a screen reader gets. Fully off under prefers-reduced-motion
rather than merely reduced — nothing here carries meaning, so the honest
reduced version is nothing at all.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The hue only appeared on completion, so the half of the strip a reader
actually scans — the unfinished half — was a row of identical grey pills
with nothing to say which game each belonged to.
Soft border and wash while outstanding, full strength once ticked. Done
is still legible without the colour: the tick, the stronger border and
the deeper wash all carry it.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Endfield has two server groups, not three: Europe is served off the
Americas machine on a fixed UTC-5, so a European player's day rolls at
09:00 UTC. We were resetting it six hours early, at 03:00, which ticked
the wrong box every morning between those two instants.
Adds GameMeta.resetOffsets, a sparse per-region override, and threads an
optional `game` through every day-key function. Per region rather than
per game on purpose: a blanket offset would drag Asia — which does have
its own Endfield server — onto the Americas clock, moving day keys for
readers who never had the bug. A regression test pins Asia's output as
identical to before.
Day keys are localStorage keys, so this re-labels ticks logged between
03:00 and 09:00 UTC by European Endfield players, one day backward. No
tick is deleted and past days stay editable, but a streak can read as
broken for a day. That is the cost of correcting a wrong reset; leaving
it wrong is worse.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The tree had drifted: progress controls, the daily strip and checklist,
the undo toast, and three state modules have all been in the code for a
while without appearing here.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Guessing from a source's wording is right most of the time, and someone
who finds it wrong often enough should be able to stop it rather than
dismiss the same false positive every patch.
Off, only events the reader marked themselves get a checklist. It
silences the guess rather than deleting anything: overrides, ticks and
streaks all survive, so switching it back on restores exactly what was
there. With it off, marking an event stores an explicit yes, since there
is no longer a detection for an override to agree with.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Dailiness was read off the source's wording alone, which is wrong in both
directions: a grind that resets every day but whose page never prints
"daily" got no checklist, and a banner whose blurb mentions "daily login
rewards" got one nobody could dismiss.
The reader's answer now wins. The control sits exactly where the
checklist goes — the one place the answer visibly matters — so marking an
event and ticking today off are the same gesture in the same place.
An override is stored only when it disagrees with detection. Recording
agreement would freeze today's guess into the reader's own data, so a
later parser improvement could never reach that event.
Marked events also join today's dailies at the top of the page, beside
the per-game chores: at 23:50 a login campaign and a commission run are
the same job, and ticking one should not mean opening a sheet to find its
checklist.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The suite gained three tests that assert on .github/workflows — the
freshness-cache restore, the per-attempt cache key and the rebase-and-
retry push, all defects that live in YAML and fail silently. But
.dockerignore excluded .github, so those three tests failed inside the
image build and took the whole build down with them.
Put .github in the build context. It reaches the build stage only; the
runtime stage still copies nothing but public/ and serve.ts, and the
image is unchanged (verified: /app holds public and serve.ts, nothing
else).
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
`lastConfirmedAt` lives in gitignored bookkeeping that only refresh.yml
restored, so the workflow that actually builds and deploys never saw it:
every source reported its last *content change* as its last success, and
the UI flagged anything whose bytes had not moved in two days as stale —
which is most wiki pages most of the time. ci.yml now restores the same
cache read-only before building the feed.
The refresh push was a bare `git push`, so a human push landing in
between made it non-fast-forward: the job failed and threw away pages it
had just fetched, while the bookkeeping had already been saved, so those
sources would not be re-asked for six hours. Rebase and retry instead —
never force.
The cache save key used run_id, which is stable across re-runs, so a
re-run saved nothing and the run after it restored stale bookkeeping.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
A review of code that has never made a live request. Each of these was
reproduced before it was fixed.
robots.txt matched a group by testing whether our whole User-Agent
contained the group's name. Our contact URL carries "StereotypicalCat",
so a wiki writing `User-agent: cat` matched us — and because a named
group *replaces* the wildcard group, that silently discarded every rule
under `User-agent: *`. Match the RFC 9309 product token instead. A 200
carrying an HTML "not found" page also parsed to zero rules and read as
permission; it now fails closed, while a genuinely empty body still
means "no restrictions".
The response body was read outside the try that guarded the request, so
one truncated body aborted the whole cycle: later sources were never
fetched, and the failed source never recorded its check, meaning a
re-dispatch would ask that wiki again minutes later.
Bodies were decoded as UTF-8 unconditionally and stored re-encoded. A
page served in a legacy charset became replacement characters with the
original bytes gone — and mojibake in a title flows into slugify and
moves every localStorage key for that source. Decode by the declared
charset, keep the served bytes verbatim, and hash those.
Unchanged bytes skipped the metadata write, so once a server rotated its
ETag we sent a stale validator forever and it served full bodies instead
of 304s.
Also: snapshot writes go through a temp file and a rename, and `--only`
with no value is an argument error rather than "all sources".
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
A day key is a game-day, not an instant, but the pip's title and
aria-label were formatted in the browser's timezone. Every reader west
of UTC saw a pip reading "12" whose screen-reader label said "Aug 11" —
including the whole `america` region these checklists are scoped to.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The cut was marked with U+2026, which NFKC decomposes into three dots —
so re-sanitising a truncated string grew it by two characters and re-cut
it at a different word boundary. A title would quietly rewrite itself
every time the event was re-ingested, and the module promises
idempotency in its own docstring.
Append what normalisation would produce instead. The existing corpus
missed this because its only over-length entry has no space in its last
40%, so it happened to re-truncate to the identical string; the new test
uses prose.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The status sections claimed the feed was generated from fixtures and the
scheduler unbuilt, which stopped being true. Also documents the
sanitisation stage and the two new key spaces — `dailies:<game>` and
game-day keys — beside the existing warning about event IDs, since they
carry the same "no server-side recovery" property.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
"What runs out first" is the question this app exists for, but it is not
the only one a reader arrives with. The other is "what was I in the
middle of?", and until now the list could not answer it.
Two modes, toggled from the list's own header rather than from settings,
because ordering is something you reach for while looking at a list.
Ticking a day off a daily counts as doing it without the reader also
setting the status — the tick already said so.
Sorting only ever groups. Both modes fall back to endingSoonestFirst
inside a group and live still beats upcoming, so choosing an order can
never cost the reader the deadline order the product is for.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
A login campaign is not one job with a deadline. It is twenty small jobs
on twenty separate deadlines, and a day you miss is gone whatever you do
afterwards — which a single "done" tick cannot express.
Repeating events now get a checklist: today's tick, a strip of every day
in the run showing what you got and what you missed, the streak, and how
many chances are left. Past days stay editable, because people tick up
later and a checklist you cannot correct stops being trusted after the
first mistake. Alongside it sits today's dailies — commissions, sanity,
daily training — one tick per game, keyed `dailies:<game>`, since no
source publishes those and they are the only thing on the page that
expires tonight rather than next patch.
Days roll at 04:00 server time per region, not midnight: finishing at
02:00 is still yesterday, and a naive UTC date would tick the wrong box
for four hours every night.
Dailiness is read off what the source published — a login event type, or
"daily"/"check-in"/"7-day" wording — never from a game's habits or an
event's length. That adds no schema field, so the feed contract and every
event ID are untouched. An unannounced end yields a tick count, not a
checklist of invented length, and a tick is never removed except by the
reader, including ticks outside the window the feed now claims: a source
quietly moving a date must not erase a fortnight's streak that exists
nowhere else.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The bar tested for an entry in the progress store, but an entry appears
there the moment a reader records an effort or a note. Recording "this
looks like a grind" greyed the event out as though it were already done.
Ask whether it is done instead.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The row already opened the event; the tick sat on top of it, so the two
smallest targets on the page did different things a few pixels apart.
"Done" was never the only thing a reader wants to say about an event
either — status, effort and notes all live in the detail sheet.
The tick becomes a chevron: decorative, aria-hidden, and leaning towards
where the row is about to take you on hover. The full-bleed button is
the only control, so there is no second stop for anyone tabbing or using
a screen reader. Undo on a revealed ignored row stays a real button —
it is a real action with nowhere better to sit.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The feed was generated from checked-in fixtures and only moved when
somebody captured a page by hand. This fetches.
bun run refresh caches each page raw under snapshots/ and rebuilds the
feed from what it cached; build-feed prefers a snapshot and falls back to
the fixture, so a clean checkout and the container build stay offline and
reproducible. The workflow runs it twice a day and commits only when a
page's bytes actually changed — a 304, an identical body or a rejected
parse all leave the tree clean — then dispatches ci.yml, which already
knows how to test, build and deploy.
Scraping conduct is enforced in code rather than left to good
intentions: one request per source per cycle, a six-hour floor checked
per source, conditional requests, a User-Agent with a contact URL, and
robots.txt honoured — failing closed, because a permission we could not
read is not a permission we have. No retries; a retry is a second
request.
A body that yields zero events is rejected and the previous snapshot
kept, so a redesigned wiki shows up as a stale timestamp rather than an
emptied calendar. One source down is a warning; all of them down fails
the run, so a cycle that learned nothing is never committed.
Tested entirely offline against an injected fetch and clock — no request
has ever been made to a live wiki from this code.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Scraped text becomes React content, JSON on disk and eventually SQLite
rows, so it gets cleaned at one seam: the parse wrapper in toAdapter().
Every source passes through it, a source added tomorrow is covered
without its author doing anything, and no parser can opt out — parsers
stay pure readers of one site's markup.
Removes script/style/comment content and residual tags, decodes entities
to a fixed point so an encoded tag cannot resurrect in a later decoder,
NFKC-normalises, strips control, zero-width and bidi-override characters
(an RTL override visually spoofs a title), bounds each field to the cap
the schema already declares, and requires sourceUrl to be absolute
http(s).
Three things it will not do: touch a date, drop an event it could clean
instead, or repair in silence — every repair and drop is logged by
default. Event IDs are localStorage keys, so an ID is recomputed only
when a sanitised title actually changed and the ID was minted the
standard way; all seven fixtures pass through unrepaired and
byte-identical, which is the regression guard.
Also stops decodeEntities throwing RangeError on an out-of-range
numeric reference, which would have taken a whole source's events down.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The done checkbox now cycles untouched → doing → done, in the order people
actually move through. Effort, status and a note are set from the detail
sheet.
Rows carry chips for what the reader said, and a "tight" or "running out of
time" flag when the remaining time no longer covers the effort they recorded.
The detail sheet shows the arithmetic behind that rather than just asserting
it, and calls it a rough guide.
Imports from before this change are mapped forward rather than dropped.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Membership in a set can only say "done", so "started" was inexpressible.
Progress carries a status instead, plus an optional effort estimate and a
note.
The old completions key is read once to seed status: "done", and is never
written to or deleted. Someone who last opened the app six months ago still
has their marks under it, these live only in the browser, and nothing else
holds a copy to restore from.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Four buckets named the way a player would describe them — quick, short, long,
grind — each carrying rough working hours.
The hours exist so effort can answer the question the calendar cannot: can I
still finish this? The same two days is comfortable for a quick event and
hopeless for a grind. The runway heuristic assumes about an hour of play a
day, which is deliberately modest — a warning that only fires once something
is already impossible is useless.
An event with no recorded effort returns "fine" and never warns. Inferring an
estimate in order to warn about it would be fabricating the reader's input.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
It still thanked only Game8 and hardcoded the studio list, so wiki.gg went
uncredited and unmentioned in the disclaimer after being added as a source.
Sources and studios are now read from the feed and the game registry, so
adding either credits the right people automatically. A hardcoded credit goes
stale the moment someone adds a source, and out of date is the one thing a
credit must not be.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Ignoring made an event vanish with no way back except knowing the reveal
toggle existed — and the row you would click to undo is the row that just
disappeared.
Now: an undo appears the moment you ignore something, and revealed ignored
rows are labelled and carry a restore control where the done checkbox
normally sits, which is the most direct place for it.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Clamping the window to now made the past unreachable — you could not see when
a running event began. Rendering from the earliest start instead buried today
off-screen, which is what the clamp was fixing.
The window now covers both and the container opens scrolled to today, so the
default view is unchanged and history is a scroll away. It reaches a week
past the oldest running event, so the range stays bounded by what is actually
live. Nothing is clipped at the left edge any more.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Dark Reader inverts an already-dark palette, which wrecks the interface: the
colours here carry meaning — the urgency ramp is how you read how soon
something ends — so having them re-mapped is not a cosmetic problem.
darkreader-lock is its documented opt-out. color-scheme: dark is the standard
signal alongside it, so browser chrome and form controls render dark too.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The image build ran the tests without serve.ts present, so the server the
tests spawn never started and the hook stalled until it timed out. The
runtime stage would have failed on the same missing file immediately after.
The test now bails the moment the process exits and reports its stderr, so
this shows up as "serve.ts exited with 1 before listening" in 53ms rather
than an unexplained hook timeout after five seconds.
serve.ts is also in tsconfig's include now — it was outside it, so a type
error in the file that serves the app would only have surfaced at runtime.
Adding it immediately caught one in the test.
Verified by replaying the build stage against exactly the copied file set.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
configure-pages fails with a 404 when Pages has never been switched on for
the repository. It can enable it given the pages: write permission the job
already holds, which beats requiring a trip through settings before CI can
go green.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
github.repository preserves the owner's casing and GHCR rejects any
uppercase in a repository name, so the image job failed on tag validation
before it built anything.
Also drops the review-quarantine skill's references to evidence spans and the
extraction-evaluator agent, both of which went away with the LLM layer.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The docs had drifted in ways that would mislead: DATA-MODEL documented a
localStorage shape the code stopped using (completedAt, no ignored store),
INGESTION claimed three Game8 templates when five are known, ARCHITECTURE
still listed the whole client and time.ts as unbuilt, and the review-quarantine
skill described a pipeline that does not exist yet without saying so.
Adds the parser roster and the six date formats as tables, documents the
subpath/base-href and offline behaviour, and records the new product surface
(first-run picker, ignore, offline, credit) as PRD features.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
They read from public/, which does not exist on a clean checkout — CI runs
bun test before any build, so the four of them failed there while passing
locally against stale build output.
serve.ts now takes PUBLIC_DIR, and the tests serve a temporary tree they
create themselves. They exercise serve.ts rather than the build, which is
what they were always meant to do.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Mirrors the GitLab pipeline — typecheck, tests, feed sanity — and adds a
Pages deploy.
Pages serves from /<repo>/, and the app used absolute asset paths, so that
job would have shipped a site that 404s on everything. Assets now resolve
against a <base href> the build substitutes, the feed URL resolves against
document.baseURI so deep links work too, and the service worker derives its
paths from its own registration scope. Root-hosted builds are unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The server is a placeholder for the one in docs/ARCHITECTURE.md — it serves
public/ and a health endpoint, nothing more. Reads are confined to public/ by
resolving the path and checking it stays inside the root; string-matching
".." is not enough, since encodings and URL normalisation both change what
the string looks like and only the resolved path says which file would open.
The image runs typecheck and tests during build, ships no source or
toolchain, and runs unprivileged.
CI's feed job fails if the event count collapses. A source that quietly stops
yielding events is what a parser-only pipeline is most prone to, and nothing
else would surface it. Everything is offline, so a red pipeline always means
the code changed rather than a wiki being down.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Ignoring is not completing. "Done" keeps an event visible and counted;
"not interested" removes it from both views, which is the entire point.
Completions and ignores are the same shape and want the same guarantees, so
they now share one mark-set implementation — union merge on import, never a
removal, since nothing else holds a copy. Both ride in the export file.
Ignored events stay recoverable: the count and a reveal toggle appear in
settings once there is something to reveal.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
First non-Game8 parser, and the first source that states region-scoped ends:
wiki.gg emits ISO timestamps with one timer per server region, so Endfield
events now carry exact times and separate Asia / Americas-Europe ends. That
gap is up to 13 hours, which is what regionEnds was built for.
Two sources for one game also exercised merge for real. It caught both
overlaps — including "Bedazzling Dawnstar" against Game8's "Bedazzling
Dawnstar Sign-In" — and flagged that the two disagree on the end date by 70
hours rather than averaging them into a date neither states.
Matching an appended qualifier needed a prefix test, not a subset test: a
subset would also fuse "Gold Clash" with "Gold Rush Clash Royale".
Also fixes build-feed picking fixtures by game rather than by source, which
handed the wiki.gg page to the Game8 parser once Endfield had two.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The meter encoded two things at once and neither was guessable from looking
at it — obvious to whoever built it, opaque to everyone else.
Every row now carries a plain caption under the bar ("Aug 10 – Aug 17 · 3 of
7 days left", "started Apr 22 · no end date announced"), which explains the
bar and doubles as the exact dates people wanted anyway. A legend above the
list explains the colour ramp, which no per-row text can — shown once,
because a legend repeated fifty times is noise.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Endfield was previously written off as undatable. That was wrong: a first
pass only inspected its Duration rows (all "Permanently Available") and its
year-less version grid, and missed an "Event | Schedule & Summary" table
whose cells read "Period: 08/09/26 - 08/30/26 During the event...". Two real
events, with a year.
Adds MM/DD/YY range parsing for that shape. Month-first ordering is not
assumed lightly — Endfield's own version grid reads 01/22, 04/17, 07/16 for
versions 1.0, 1.2 and 1.4, chronological only if the month comes first.
Also stops presenting unlock conditions as descriptions. Several templates
put "Reach Union Level 8" where a blurb would go; filling the summary slot
with text that never says what the event is is worse than leaving it empty.
Wuthering Waves correctly drops to zero summaries as a result.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The window ran from the earliest event start, which is months back, so today
sat far off-screen and every visit began with a scroll to find yourself.
It now opens at now. Events already running are clipped at the left edge and
faded there, the same way an unannounced end frays on the right — both say
the truth extends past what is drawn. The left edge also carries a month
label, since the window opens mid-month and the first real boundary can be
weeks away.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The reader's question is answered entirely by data already on the device and
countdowns run off the local clock, so losing signal should not lose the app.
The shell is cache-first; the feed is network-first falling back to the last
copy seen, because stale events beat a blank screen. Webfonts are cached too
— without them an offline load silently drops to system faces and the whole
thing changes character.
Offline is surfaced in the header and above the footer rather than hidden:
stale data must never be presented as current.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Game8's editors compile the calendars this depends on, and the studios make
the games — both belong on the same screen as the data rather than one
navigation step away.
States plainly that this is an unofficial fan-made tool with no connection
to Game8, HoYoverse, Kuro Games, Hypergryph or Hotta Studio, and that the
source page is the authority when the two disagree.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Three more Game8 templates. ZZZ spans each event name across two rows with a
rowspan, so a flat cell reader sees [title, "Start", date] then ["End",
date]; losing that pairing would silently halve the calendar.
Section matching also had to widen — these pages head their sections
"Featured Events", "List of Events and Schedule" and "Ongoing Events".
Verified independently of the parser: HSR's year-less summary table lists
six ongoing events and all six are captured via their detail rows, so
nothing is silently dropped.
28 events added across the three games.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Honkai: Star Rail and Wuthering Waves state a year on each side of a range
("Aug. 14, 2026 - Aug. 24, 2026"), which the year-on-the-end parser cannot
read.
Both also publish genuinely open-ended events — "Jul. 24, 2026 - End of
4.6", "July 10, 2026 - Permanent". Those keep their real start and report no
end, rather than being dropped or given a guessed one.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
A calendar full of games you don't play is worse than an empty one — it
buries the thing you came for.
Nothing is preselected; the button stays disabled until a game is picked,
which is clearer than guessing and hoping the reader notices. The choice is
stored as hiddenGames, the inverse, so a game added later shows up by
default rather than staying invisible.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The whole row is now one click target, which gives a generous tap area and
a single unambiguous hover region.
Hover transitions are fast and eased out (110-160ms) so a row responds on
the way in and settles on the way out — a slow symmetric fade is what makes
an interface feel laggy even at full framerate. The meter is the element
that visibly wakes up, since it is the signature. Press feedback lands on
touch as well as mouse, and every hover rule sits behind `hover: hover` so a
tap never leaves a row stuck looking hovered.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Column tables already carried a description column, but label/value detail
tables did not — Genshin events had no blurb at all. The page follows each
detail table with a sentence of prose, so take that, skipping the
call-to-action paragraphs, which are navigation rather than description.
Summary coverage goes from 13/22 events to 22/22.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Covers what's built versus specified, how to run it, the parser/adapter/
merge layering, and why Arknights: Endfield has no source.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
