§30 says the app stops asking "did your period start?" after a few
unanswered check-ins, and starts again next cycle. It could not.
The count was reset by a flow watching the newest confirmed start.
distinctUntilChanged only dedups within one collection, so its first
emission always passes -- and PeriodApplication starts the coordinator in
every process, including the one WorkManager spawns to run the reminder.
The count was wiped moments before the worker read it. A user who ignored
the check-ins kept being asked, daily, which is the behaviour §30 exists
to prevent and the kind people uninstall over.
The count now carries the row id of the period it was asked about and
reads as zero for any other. A new period starts it over by arithmetic
rather than by an event that has to fire at the right moment in the right
process. The reset chain is deleted outright -- there is nothing to race
and nothing for a second process to get wrong -- and the notification
handler no longer resets anything either.
An id, never a date. The rule that keeps dates out of
PeriodRecord.toString() applies to anything at rest a backup or a crash
reporter could pick up: an id says a record exists, a date says when
somebody bled. Delete My Data leaves an id matching nothing, which is
correctly no count at all.
The coordinator test that covered the deleted chain is replaced by one
that starts the coordinator twice on the same history and asserts the
tally survives -- the regression itself, rather than the machinery that
used to cause it.
Also raises the app-lock test's await budget from 5s to 30s. It went red
once in a full parallel run and passed alone: each PIN there costs a real
210,000-iteration PBKDF2 derivation, and one test asks for three. The
budget is for catching a hang, not for measuring the crypto -- and a flaky
guard is one people learn to ignore.
Proved with prove-guard, one red each: dropping the period from countFor,
and letting recordCheckIn increment across periods.
closes#70
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three ways the lock could fail badly, all in its storage.
hasPin had no catch, unlike the equivalent flow in
UserPreferencesRepository, and it is collected on the startup path in two
places that cannot handle a throw: the gate's stateIn, which launches
outside its own exception handler, and MainActivity's FLAG_SECURE
collector. A corrupt app_lock file crashed the app before any UI existed
to say why.
It now reads as no lock. That is a decision, not a shrug: whoever can
corrupt that file has the app's private storage and therefore reads the
database this lock does not encrypt anyway, while the alternative is an
app that can be neither opened nor erased -- because the way out is behind
the lock that is broken. It is the rule VerifierRecord.decode already
applies to a single corrupt record, extended to the file.
The store also gains a corruption handler. Without one, a single bad write
leaves DataStore unable to read OR write it: the lock can never be set
again, and the "Forgot your PIN?" erase fails too, since it writes here.
And a wrong PIN now costs time even when the counter cannot be signed.
Signing needs the Keystore, which can be briefly unavailable; the write
returned early, so guessing was free for as long as that lasted -- the one
direction this must not fail in. The unwritten counter is held in memory
and read back whenever it is the longer wait. Not persisted: it is the
delay actually earned rather than the maximum a tampered counter earns,
and it is forgotten on process death, which is the bound a reboot already
gives an attacker.
Writing that test found the honest boundary. If the key is gone entirely
the PIN cannot be checked either, so the app reports that it does not know
rather than charging for a guess it could not read; the failure worth
defending against is the counter alone failing while the verifier still
works. The fake now tells those apart by the domain-separation tag the two
uses already carry.
Proved with prove-guard, one red each: dropping the in-memory counter, and
removing the catch.
closes#64
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Change PIN asked for the current one, and then moved to "Choose a PIN"
without waiting for the answer. The check is asynchronous; the screen
reassigned its step outside the result and passed an empty callback. Any
four digits reached the replacement screen, and setPin enrolls without
verifying anything.
So anyone holding the phone while it was unlocked could change the app's
PIN. Under the no-recovery policy the owner's only way back into her own
history is to erase all of it. The file's own KDoc says this must not be
possible.
The step now advances from inside the verified callback, as "turn the
lock off" already did -- and the ViewModel refuses a replacement that no
successful check authorised. Two guards for one rule on purpose: a screen
is the kind of file that gets rewritten by somebody who has not read the
one behind it, and the place that writes the PIN is the place that has to
refuse. Cancelling withdraws the permission; a successful write spends it.
"Turn the lock off" had the same advance-before-answer shape. It was safe
-- the work was already inside the callback -- but its wrong-PIN message
landed on a screen that had gone, so ConfirmPin.wrong was dead code. Fixed
symmetrically.
Also fixes the lock-out race in the same function (#62). setPin wrote the
PIN and then unlocked the session; the gate is combine(hasPin, unlocked)
and closes on (true, false), so DataStore's emission could arrive in
between. AppLockGate disposes the whole app subtree when it closes, this
ViewModel is scoped to a destination inside it, and the unlock was
cancelled with the scope -- the user was thrown to the lock screen to type
the PIN she had chosen a second earlier. The development log records the
common case as fixed; the fix lived in the scope the race destroyed.
Unlocking first makes the bad pair unobservable: unlock() sets a
MutableStateFlow synchronously on this thread, before the write begins,
and combine always emits with the latest of both. If the write fails there
is no PIN and Unlocked is correct anyway.
The settings ViewModel had no test at all, which is how a wrong PIN
reaching the replacement screen went unnoticed. It has five now, against
the real repository over a host-JVM signing key -- core/security gains a
small public two-argument constructor for that, since AndroidKeyStore
cannot be reached off-device and faking the repository would prove nothing
about it.
Proved with scripts/prove-guard.sh, one red each: spending the
authorisation, and the old write-then-unlock order. Removing the
write-site guard entirely reddens three, which is that guard's whole
surface rather than a coincidence.
closes#60closes#62
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The period-end check-in asks "Is your period over?" and offers Ended and
Still going. Tapping Ended inserted a NEW period record starting today,
in the middle of the period it was asking about. Still going filed a
censoring observation against a forecast that had already arrived.
The labels were chosen in NotificationCopy and the writes were attached
in ReminderWorker by position -- index 0 to "started", index 1 to "not
yet", for every kind of reminder. That holds while every reminder asks
the same question. It stopped holding the moment one did not.
It corrupted the health record and every forecast built on it, and the
user had no way to see it happen.
A button is now one thing carrying both halves: NotificationCopy.buttons
returns the label and the action together, and nothing downstream is
allowed to pair them up again. ENDED closes the period that is running
through setPeriodEnd -- the same call the Today screen makes -- and never
opens one. STILL_GOING deliberately writes nothing: it is the state the
record is already in, and the in-app equivalent is a no-op that would
still move updatedAt and read, in the history, as an edit she never made.
A start confirmed from a notification is now sourced
NOTIFICATION_CONFIRMATION rather than MANUAL. How a record arrived is
part of the record.
Actions travel as their enum name, and anything unrecognised -- including
the strings used before this change -- writes nothing. A notification
sitting in somebody's shade across the upgrade still opens the app; it
just does not guess what she meant. The extra key is now declared once in
core/notifications and read by MainActivity rather than repeated as a
literal on both sides.
The handler had no test at all, which is how this survived: it owns the
only two writes reachable from a locked phone. It has eight now, and the
first is not about a write -- it asserts the two halves agree, in every
privacy mode, as a property.
Proved: mutating the already-closed guard out reddens exactly one test
(scripts/prove-guard.sh). Reverting ENDED to its old write reddens three,
which is the whole ENDED semantics and not a coincidence.
closes#68
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
"Will I have my period the week of the wedding?" is a core prediction
question the app could not answer. The calendar marked only the next
forecast window and stopped browsing one month out, so a textbook-regular
user paged forward into blank months.
CycleProjection reaches a year. It is a separate type from Prediction,
deliberately: a forecast is something the app scores itself on -- the
period arrives inside the window or it does not, and PredictionRecord
writes down which -- while a projection eleven cycles out is never
scored, never learned from, and will have been replaced four times before
its date arrives. Separate types mean a projection cannot reach the
accuracy figures by accident.
Three rules keep it honest, each with a test named after it.
Uncertainty grows as sqrt(n), because cycle lengths are near-independent
draws: eleven cycles out is about three times as uncertain as one, not
eleven times. Real cycles do correlate, so that is the optimistic edge --
which is the second reason the assumption sentence is mandatory.
It declines rather than stretching. Past ten days either side a
projection stops being an answer, so the projection ends and reports that
it ended -- FertilityEstimate's precedent, which already refuses to show
a seventeen-day fertile window. The caller can tell "no period due here"
from "cannot see this far", and only the second invites her to log more.
The assumption is stated, not implied. Any month showing projected marks
carries "If your cycles continue as they have, this is the forecast" with
the confidence for that distance beside it. A year-ahead date drawn
without that is the clearest overstatement this app could make.
Cycle 1 is the engine's own forecast copied through unchanged, so the
calendar and the Today screen cannot disagree about the next period. The
typical cycle length comes from her own history and never a population
default -- projecting 29 days for a 35-day user is §3's core defect
repeated twelve times over.
PROJECTED_PERIOD is its own mark: the same dashed ring as a prediction,
thinner and finer. The family resemblance is the message -- still a
period, weaker claim -- and it survives greyscale, which an opacity
difference would not. Screen readers get the distance in the label, since
they cannot see that a square is months away.
Browsing now reaches a year; logging still stops at today.
ConfidenceRow and readable() move to feature/common and are shared with
the Today screen rather than copied. Two of them drifting apart is a
screen saying Medium in dots and Low in words.
closes#57
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Tracing the confirm -> score -> feed-back loop end to end turned up four
ways it lost or falsified its own evidence. None of them looked broken:
each produced plausible accuracy figures with something missing.
A forecast's generatedAt is now the date its LINEAGE began, not the date
its row was written. Every "Not yet", edit and delete revises the answer
to one standing question, so the replacement inherits the origin instead
of restamping today. Restamping moved the goalposts of the backfill
guard: a "Not yet" on the 29th, then a period logged on the 30th as
having started on the 28th, tripped the guard and the forecast the user
was actually shown was deleted unscored. The app learned nothing from
precisely the cycle it got wrong.
The standing snapshot is retired before the engine is consulted rather
than after, so a lineage dies with its history instead of waiting to be
scored against an unrelated one.
Scores now follow the period they are facts about. Editing a start
re-scores every snapshot recorded against it -- predictedStartDate stays
immutable, so a correction worsens the figure as readily as it improves
one -- and the same rule that refuses to score backfill retracts a score
whose start has moved behind its lineage, so an edit cannot smuggle in a
measurement the guard would have turned away. Deleting the period
retracts outright.
A confirm that becomes the newest start clears every "not yet", not just
the older ones: they all censor the same question. A date-bounded clear
left observations dated after a retroactively logged start alive to
depress the next cycle's confidence for a question already answered.
Also removes CycleData.repository's engine default. Nobody relied on it,
which is the point -- a caller who omitted the argument would compile
cleanly and ship the baseline prototype §11 calls unacceptable.
No schema change: all four fixes are queries and call order.
Six of the seven new tests were observed red against the unfixed source.
The seventh -- a deep backfill does not clear the observations censoring
the standing question -- passes both sides deliberately, pinning against
overshooting the not-yet fix.
closes#46closes#47closes#48closes#49
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
"Your cycle belongs to you" was a promise the app could not keep: there was no
way to get the data out.
## The format, because it outlives the batch
One pretty-printed JSON file. The issue asks for "human-readable" and means it —
this is what somebody's archive will be in for years, so it is a contract with
eight rules written down beside it, not an implementation detail.
Pinned byte-for-byte against a committed golden file, which doubles as the
documented example so SECURITY.md links at it rather than keeping a second copy
that would drift. A reformat, a reordered key or a changed date rendering all
fail in a test rather than in an archive.
Dates are ISO calendar dates with no timezone and no conversion, ever.
Converters.kt stores a LocalDate as its epoch day precisely so it "cannot carry
a timezone by accident", and a zone-aware formatter here would shift every date
for users east or west of whoever wrote it — a cycle tracker off by one day is
wrong in the way that matters. There is a test that renders the same fixture in
UTC, +14 and -12 and requires identical bytes, because that bug would never fire
where it was written.
## Plaintext, and that is the decision rather than the default
An earlier note said this would be encrypted. It should not be, and SECURITY.md
now says why: the export is the copy that makes a lost Keystore key survivable
instead of final — the exact condition recorded for ever revisiting database
encryption. Putting it behind a passphrase reproduces the failure that decision
was taken to avoid: a forgotten secret and an archive nobody, including this
app, can open. §45's "prefer encrypted backup/export formats" is scoped to
backup, which this is not.
## Only the user's own data, as a compile error
:core:export is pure JVM and depends on :domain:cycle alone. Prediction,
PredictionAccuracy, FertilityEstimate and CycleRecord live in
:domain:prediction and are simply not on its classpath, and kotlin("jvm") keeps
android.os.Build off it too — so a forecast or a device fact cannot be added by
accident. The key set is asserted with assertEquals rather than contains, so a
new field is a failing test rather than a silent addition.
Row ids are out because they are monotonic and would disclose how many records
the user DELETED. The Play entitlement is out because a purchase one file-edit
away from being granted is a purchase that will be.
## No second copy, ever
The Storage Access Framework writes straight into the document the user picked.
The alternative — write to cacheDir, share by FileProvider, delete after —
creates the temporary file the issue warns about and races the receiving app
still reading it. A test walks cacheDir after a successful export and requires
it empty; it fails the moment anybody reintroduces that pattern.
The destination is parked until the session is unlocked. Returning from the
picker can re-lock, and writing while locked would hand the whole history to
whoever took the phone during the save dialog.
## Two new guards, both proved to fail
checkNoSharedStorageWrites: §45's shared-storage ban was enforced by nobody
having typed it. checkPermissions structurally cannot see it — it matches
<uses-permission>, and a <provider> declaring FileProvider merges green.
checkNoHealthLogging gains a completeness check. A module missing from
modulesSeeingHealthData was silently exempt with a green build, which
app/proguard-rules.pro has described as a hazard since before :core:security and
:core:export existed. Every module must now be in that list or in an explicit
modulesWithNoHealthData with its reason; being in neither is a violation rather
than an exemption.
261 JVM tests, none skipped. Five guards green.
closes#35
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both builders in PeriodNotifier called setSmallIcon(android.R.drawable.ic_dialog_info).
Android masks a small icon to a white silhouette taken from its alpha channel, so
what every reminder this app has ever sent put in the status bar was that
framework asset's outline.
ic_notification is redrawn rather than copied from ic_launcher_monochrome, and
the difference is the point: that file is a 108dp launcher canvas whose shape
sits in the upper safe zone because a launcher crops and masks it. Copying its
geometry would have produced a small mark floating above centre. This is a 24dp
canvas the ring nearly fills, opaque white throughout, because the system
discards colour and supplies its own.
Both call sites, not one. The public builder is what a locked screen renders and
is the one that matters most here.
## Two tests, because "which resource id" is not the whole claim
bothTheLockScreenAndTheShadeShowThisAppsOwnMark reads the posted Notification
rather than the source, and checks the public version separately — a change made
by half is the likely mistake and it fails silently on the surface this product
is most careful about.
theStatusBarMarkRendersAsAReadableSilhouette renders the vector and measures
alpha coverage. Two failures look identical in source and completely different in
the status bar: a vector that draws nothing, and one that draws a filled shape.
Neither is caught by asserting a resource id.
Proved: reverting only the public builder fails exactly one test, naming that
builder. prove-guard exit 0.
## NotificationPrivacyTest could never run on minSdk
Found while satisfying this issue's own verify line. GrantPermissionRule asked
for POST_NOTIFICATIONS unconditionally, and that permission arrived in API 33 —
so on PeriodMinSdk26 every test in the class errored with "Failed to grant
permissions" before reaching an assertion, for a reason unrelated to what it
tests.
That is how it stayed unnoticed: it is the only emulator where it fails, and a
green run on a modern image looks like a green run. The class guards what a
LOCKED SCREEN shows. "Passes on the newest device" was never the claim worth
having. The rule is conditional now, and below 33 no permission is needed to
post at all, so a no-op rule is correct rather than a workaround.
All six tests now pass on PeriodMinSdk26 — the first time this file has run
there.
## Not verified
The API 36 instrumented run. That emulator repeatedly dies the moment Gradle
starts on this machine today, across three attempts and after freeing memory; it
ran the whole app-lock UI verification earlier in the same session, so this is
resource contention rather than a defect. The SECURITY_CHECKLIST row covers it,
and both new assertions are resource-id and render checks whose substance does
not vary by API level.
closes#43
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
§45 asks for biometric/PIN gating. UserPreferences.biometricLockEnabled has
existed since Batch 01 with nothing outside its own module reading it; this
wires it, and adds the rest.
The recovery question was the reason #34 sat open, and it is decided: there is
no recovery. A backdoor into a period tracker's lock would be used by exactly
the person the lock exists to stop. Everything below follows from that.
## The gate
AppLockGate wraps the whole composition rather than being a screen inside it.
Today, Calendar and Insights each start collecting from CycleRepository the
moment they compose, so a lock implemented as a nav destination would already
have read the history before the user proved anything. content() is invoked
only in the unlocked branch.
Re-lock on ON_STOP, not ON_PAUSE — pause fires for the shade, quick settings
and a permission dialog. Two guards on top: isChangingConfigurations, or
rotation and the fontScale-2.0 pass both re-lock; and authInProgress, or an OEM
biometric overlay that stops the activity produces a lock that can never be
opened. No grace period: SECURITY.md leads with "someone who picks up an
unlocked phone", which is the window a grace period covers.
The unlock flag lives in a @Singleton, never in saved state. rememberSaveable
looks like the obvious home and would restore a background-killed app already
unlocked — the single most likely way to meet the lock screen would be the one
path that skipped it.
## What is stored is not the PIN
mac = HMAC(keystoreKey, 0x01 || salt || PBKDF2-SHA256(pin, salt, 210k))
Two layers because they defend different things. The Keystore MAC is what makes
a six-digit PIN safe at all — a million candidates is nothing to an attacker who
can compute the hash, and impossible for one who cannot get the key off the
device. PBKDF2 underneath is for the day that assumption breaks. 0x01 is a
domain-separation tag; the lockout counter is MACed under 0x02.
The key omits six builder calls and the KDoc names every one. setUserAuthenti-
cationRequired is the important absence: it would bind the key to the device
lock, so changing a passcode would destroy it — and under no-recovery that is
somebody's whole history gone for an unrelated reason. It would also be a
bypass, since SECURITY.md already names "someone who knows the unlock PIN" as
an adversary. The biometric key is separate and takes the opposite policy,
where invalidation correctly degrades to "use your PIN".
## Wrong PINs cost time, never data
Four free attempts, then 30s/1m/2m/5m/15m, capped forever. No attempt limit and
no auto-wipe: under no-recovery an auto-wipe would let a partner, a child or a
pocket destroy a history while knowing nothing. Both clock bypasses are closed —
the wait is the longer of a wall-clock and a monotonic deadline, and a reboot
re-applies it in full, detected by elapsedRealtime going backwards.
## Two writes that had to move
Tapping "Not yet" on a reminder writes a NotYetObservation. That button is on
the phone's own lock screen, reachable by anybody, so the action is now parked
in AppLockController and applied only after an unlock — dropped if the session
never unlocks. Behaviour is unchanged when the lock is off.
The erase behind "Forgot your PIN?" deletes health data, then the Keystore key,
then the lock store. Skipping the middle step leaves the user erased AND still
locked out; prove-guard mutates that line out and requires exactly one red.
## Found by testing, not by review
- A fresh install began in a 15-minute lockout: "no counter yet" and "counter
was tampered with" were the same value. They are now distinct.
- Setting a PIN locked you out of the session you set it in. Found on the
emulator, not in a test.
- Kotlin block comments nest, so `domain/*` in a KDoc opens one. Twice.
## Verified
244 JVM tests, 0 skipped. KeystoreVerifierTest runs on PeriodMinSdk26 and
PeriodQA — including that PBKDF2WithHmacSHA256 exists at API 26, the one choice
here with no margin, and that the key is not auth-bound on either.
On device: wrong PIN refused, correct PIN opens, am kill then reopen lands on
the lock screen, turning the lock off requires the current PIN, and
`adb exec-out screencap` returns mean=0 stddev=0 — FLAG_SECURE is real.
androidx.biometric 1.1.0 is the newest stable (1.4.0 is alpha; biometric-ktx
never shipped one). It merges USE_BIOMETRIC and USE_FINGERPRINT, which failed
checkPermissions until they were allowed on purpose, and it drags fragment to
1.5.1 — pinned to 1.9.0 since MainActivity is now a FragmentActivity.
closes#34
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The application scope in PeriodApplication was built with SupervisorJob and
no CoroutineExceptionHandler, and ReminderCoordinator launchIns two Room
flows on it. SupervisorJob stops a failing child cancelling its siblings; it
does not stop the exception, which reaches the thread's default handler and
ends the process.
That scope is the one that runs with nobody watching. Application.onCreate
runs in every process, including the ones WorkManager starts after a reboot
and at the daily reminder — no Activity, no screen, nothing to show an error.
Both ViewModels already install a handler; the one place a crash is invisible
did not.
The trigger is real rather than theoretical: repository.forecast runs the
prediction engine inside the flow, and Prediction's init block enforces its
window invariants with require.
Three layers, outermost last:
- ReminderCoordinator catches per chain, so one failing collection cannot
take the other down. Doing nothing on failure is deliberate — cancelling
the schedule would turn a failed read into reminders silently switched
off until the user next touched a notification setting.
- ReminderWorker returns success and posts nothing when it cannot read what
it needs, which is already its behaviour with no history. Cancellation is
rethrown rather than swallowed.
- The scope handler is a backstop whose only job is that the process lives.
It cannot log: checkNoHealthLogging covers this module, and an exception
message here can carry a date derived from a cycle.
The chains moved into internal functions taking flows so the catch is
reachable from a test. CycleRepository is final with an internal constructor,
which is right for a data boundary and wrong for faking, and adding a mocking
library to reach one catch would have been the worse trade.
Proved to fail, per GUARDS.md §1: removing the handler fails exactly one test
(ApplicationScopeTest.kt:69), and removing either catch fails exactly its own.
GUARDS.md gains §8. prove-guard.sh decides a guard caught the mutation from
the runner's exit code, and cannot tell a broken test from a malformed
command. Its first use here reported a clean catch when Gradle had actually
rejected `:app:test --tests` as an unknown option and run nothing. The same
tool's line-counting fallback also means the three documented boundary proofs
in architecture/README.md have been exiting 3 rather than 0 since they were
written; they now carry the fail pattern that makes them exit 0.
closes#45
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
closes#40, closes#41
The illustrations span the screen and fade into the page, at roughly a third of
its height, and the seven hand-rolled step layouts became one component.
A UNIFIED STEP
Every step wrote out Spacer / Illustration(size) / Spacer / Heading / controls,
in three different spacer rhythms with a reason recorded for only two. StepBody
holds the rhythm; a step declares what is different about it. artHeight stays
per-step: the two carrying three or four controls take a shorter hero so the
primary button is reachable without scrolling, which is the constraint the old
104dp values existed for.
The screen's padding moved from the scrolling column to the content inside each
step. Compose has no negative padding, so a full-bleed child cannot live inside
a padded parent.
THE CAP IS WHAT MAKES LANDSCAPE SURVIVE
A landscape phone is 914x411dp. A fixed 280dp hero would be 68% of it and every
step would break on rotation — nothing in this app uses WindowSizeClass and
onboarding does not lock orientation. Illustration caps at a third of the screen.
It reads LocalConfiguration.screenHeightDp rather than BoxWithConstraints: inside
a verticalScroll column a child is measured with an INFINITE height constraint,
so maxHeight there is Dp.Infinity and the cap would silently never apply. A guard
that looks right and does nothing.
DENSITY BUCKETS, FIXING A LIVE DEFECT
The old set sat in a density-less drawable/, treated as mdpi and pre-scaled at
decode — roughly 17 MB of heap for one bitmap on an xxxhdpi device, and this
artwork is far larger. Now four widths per theme, 617 to 1644 px, and no plain
drawable/ at all. 2.3 MB across every bucket; a release is an AAB and Play splits
by density, so a device downloads 130-500 KB.
That is also the error-handling fix: painterResource decodes inside composition
and there is no boundary anywhere, so an OOM would crash onboarding on first
launch. Sizing the bitmaps removes the failure rather than hiding it.
The clip is gone with the corners it existed for, and the bottom fade is drawn
from the theme background rather than baked into the art. Titles move to
headlineLarge, which is what 280dp of artwork above them needs.
Driven on PeriodMinSdk26: all steps, both themes, font scale 1.0 and 2.0, and
landscape. No crash, no OOM. ./gradlew check green.
KNOWN, FILED SEPARATELY: dark 01 Welcome is a forecast scene, so dark mode shows
a calendar on step 1. The artwork is wrong, not the wiring.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
closes#29
All eight illustrations are in the app, every onboarding step has art for the
first time, and the theme swap is resource resolution rather than a branch:
drawable/ and drawable-night/, so Android picks the asset and the artwork follows
a device theme change with nothing of ours running.
NOT VECTORS, AND THE ISSUE WAS WRONG TO ASK FOR THEM
#29 called for a redraw as Compose vector paths taking colour from theme tokens.
That is right for artwork that does not exist yet and wrong for gradient
landscapes with glow and depth — there is no honest VectorDrawable of one, and
the attempt loses the drawing or produces path soup slower than the image. §42
objects to *unnecessary* raster, which this is not.
The cost was never measured before it was argued about: all sixteen files at the
height they are actually drawn at come to 130 KB. Release APK 1.90 -> 2.03 MB.
WHAT ONLY A DEVICE SHOWED
Each drawing is composed as a card with its own rounded corner painted in, so the
pixels outside that curve are the card's own backdrop — near-black in the dark
set. Rendered unclipped, every illustration showed four notches against the app
background. Illustration() now clips just inside the painted curve.
WHAT STAYED A VECTOR
The calendar markers and CycleProgressMark, because they are drawn from data — a
marker's shape is what a day is, and the progress arc is where the user stands in
their cycle. CycleProgressMark moves from Illustrations.kt to CycleMarkers.kt so
the file layout carries that boundary.
Sized by height rather than a square box: the set has no common aspect ratio, and
height is the dimension that decides whether the primary button stays on screen.
Steps 3 and 6 take 104dp because they carry three buttons and three option cards;
verified by driving all seven steps, not by reasoning about it.
OnboardingPreviews.kt now has one entry per illustration, because a missing
drawable-night asset is invisible in light mode.
Driven end to end on PeriodMinSdk26 in both themes, no crash. ./gradlew check,
schema-guard and doc-claims all pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The project owner supplied artwork and a full visual brand guide mid-session,
which answers #8: the three marks an agent must not fake now exist because a
person made them.
docs/data/img now holds icon.webp (512x512, 11 KB), logo.webp (1024 long edge,
28 KB) and banner.webp (2176x725, 20 KB) — all webp by magic bytes, all far
under the 512 KB ceiling. Produced from the supplied art rather than invented.
Two things had to be handled rather than converted blindly. The sources render
their rounded corners against black, so a naive crop carried black wedges into
the banner; the corners are flood-filled to the artwork's own ground first. And
the lockup is square, so a 3:1 crop would have cut the wordmark in half — the
banner is the whole lockup on a matching ground, which has no seam to see
because the artwork's edges ARE that ground.
The source PNGs and the guide moved to docs/design/: docs/data/img is for the
three marks the Command Center renders, and its README is explicit that
reference art and exported sources do not belong there.
THE PALETTE, AND ONE DELIBERATE CONFLICT
core/designsystem now carries the guide's colours — deep plum, rich purple,
period rose, lavender mist, and an eggplant dark theme that is never flat black.
§39 of the specification asks for a muted teal or sage fertile window, to hold
it apart from the period colour. The guide is plum and rose throughout and gives
fertility its own pink. The guide wins: it is newer, it is the owner's, and a
teal accent inside this identity would look borrowed from another app.
What §39 was protecting is kept by other means, and this is why the conflict is
safe to resolve that way: the calendar's states differ in SHAPE. A solid disc, a
dotted ring, a continuous ring and a star stay tellable apart in greyscale and
to a colourblind user, which is what §43 actually requires. The departure is
written down in docs/design/README.md rather than left to be discovered.
The launcher icon keeps its vector for now, in brand colours. The supplied
emblem has content close to its edges and an adaptive icon masks about a quarter
of the canvas away, so dropping it in unmodified would crop the shield. Fitting
it to the safe zone is Batch 08's final-artwork work.
closes#8
§28, §29, §30 and §31. NotificationCopy is a pure function — privacy mode plus
kind plus day count in, two versions of the text out — so every combination is
tested exhaustively without an emulator. This is the one surface whose mistakes
are visible to somebody who is not the user, so the tests are exhaustive rather
than representative: every kind × every mode asserts that no health word reaches
a lock screen outside Direct, and that includes the ACTION LABELS, which §31
points out are visible text too. A perfectly discreet body under a button
reading "Started my period" leaks anyway.
TWO ANDROID BEHAVIOURS THAT LEAK IF YOU TRUST THE DOCS
A private notification with no public version does not blank the lock screen —
it shows the private text. NotificationText therefore has no nullable title and
an instrumented test asserts every kind attaches one.
And a notification channel is IMMUTABLE after creation: importance and
lock-screen visibility cannot be changed. One shared channel would have kept
whatever the user's first privacy mode set, forever — switching from Direct to
Maximum privacy would have appeared to work and changed nothing. There is now
one channel per mode. Found by an instrumented test on a device; nothing in the
unit tests could have seen it.
§30's stopping rule is a test of its own: the app asks a bounded number of times,
says "We'll stop checking for now. Log your period whenever it begins.", and
then says nothing more — while the engine keeps learning, which is the sentence
§30 puts right after it.
WorkManager, and no exact alarms. §31 rules them out and the new checkPermissions
task fails the build if one ever appears in the merged manifest — from here or
from a dependency. That guard also failed its own first proof, reading a stale
manifest because it did not depend on the task that writes one.
ReminderCoordinator reschedules whenever the forecast moves, which §31 asks for
and is the requirement most likely to be missed: a "Not yet" moves the forecast,
so work queued against the old one is aimed at a day that no longer means
anything.
188 unit tests and 6 instrumented, all passing. ./gradlew check green.
closes#24closes#25closes#26closes#27
§17 and §18. Ovulation is estimated a luteal phase before the PREDICTED next
period rather than counted forwards from the last one — the luteal phase is the
stable half of the cycle, which is why §17 asks for it that way — and the
fertile window opens five days before ovulation and closes one day after,
because sperm survive and the egg does not.
The uncertainty is inherited, not invented. Ovulation is derived from a
predicted date, so it can never be more certain than that prediction.
THE PART THE DEVICE TAUGHT
The first version showed a user one cycle in a fertile window of 8 Aug – 24 Aug.
Seventeen days. Arithmetically honest, and completely useless — over half a
cycle, dressed up as a feature.
So the estimate now returns null past a usable uncertainty, and Today says "Not
enough history to estimate. Log a few more cycles and the app will be able to
estimate ovulation." Three stable cycles later the same user gets 12 Aug – 20
Aug, which is worth reading. Verified in both states on a device.
That is the same shape as PredictionAccuracy refusing figures below three scored
forecasts and CycleInsights withholding an average below two intervals, and it
is now written down in the architecture doc as a rule rather than three
coincidences: the app declines rather than stretches.
§18'S PROHIBITION IS A TYPE, NOT A CONVENTION
FertilityLikelihood has LOWER, HIGHER and UNKNOWN and no fourth value. Somebody
reading "safe" would take a decision on it; the estimate comes from a predicted
date carrying days of uncertainty; and §18 has already promised this is not
contraception. A test asserts no label contains a permission word, so adding one
is a deliberate act with a failing test.
The disclaimer travels with the feature — same screen, same time. A disclaimer
one tab away is a disclaimer nobody read.
ANOTHER GREYSCALE COLLISION
The ovulation star was centred, which put it directly behind the numeral: in
greyscale "18" and the mark merged into one smudge. Ovulation is now the fertile
ring plus a small star low in the cell, which is also semantically right — that
day IS inside the window, and the pair reads as "that window, and this day".
Nine Compose ModifierParameter warnings fixed properly rather than suppressed.
164 tests, all passing. ./gradlew check green, 0 lint errors.
closes#21closes#22closes#23
Seven screens, §19 and §56 verbatim: welcome, last period, period end, previous
history, the privacy promise, notification privacy, first forecast. Verified end
to end on a device — the flow produces a forecast, the record persists, and a
relaunch goes straight to Today.
Three decisions with tests behind them:
- Nothing is written until the final step. Somebody who abandons onboarding
halfway has not asked this app to remember anything about them.
- Notification privacy is Discreet before the user touches anything (§28), and
Direct is last and never pre-selected. Checked on the device, not only in a
unit test.
- "Still going" and "I'm not sure" both mean no end date. §24: never invent
one. The date picker refuses future dates by not offering them rather than
by rejecting a tap it allowed.
DARK MODE WAS BROKEN FOR ALL OF BATCH 01
PeriodTheme never wrapped its content in a Surface, so every Text without an
explicit colour inherited Material's default — black — and the app background
never painted. In light mode that looked correct by accident, because dark text
on cream is what was wanted anyway. In dark mode the onboarding headings
rendered near-black on charcoal.
No test caught it and no test easily would have. It was found by opening the
app on a device and looking at it.
The Surface now lives in the theme, so a screen without a Scaffold cannot
forget, and every illustration has a light/dark preview pair. A preview is not
a test, but it is the cheapest thing that puts the failure in front of whoever
is editing the screen.
closes#16
Compose vector paths in core/designsystem/art — no raster anywhere under src/,
so everything stays crisp at any density and adds nothing to the APK.
The four calendar markers differ in SHAPE, not only colour: a solid disc for a
confirmed period, a dotted outline for a predicted one, a continuous ring for
the fertile window, and a four-pointed mark for ovulation. §26 and §43 both
require it and for the same reason — a calendar whose states differ only in
colour carries no information for a colourblind user, none in greyscale, and
none in the bug report somebody files about it. Predicted is deliberately not a
lighter confirmed: opacity alone fails exactly when the screen is dim.
Illustrations for welcome, empty state, learning and the privacy promise, plus
the cycle progress mark for Today. The visual language is overlapping circular
forms and nothing else — §42's forbidden list is a product decision, not
squeamishness: this app gets opened in public and a glance over a shoulder
should learn nothing.
Placeholders here and never in docs/data/img, which docs/design/README.md now
explains rather than leaving as an apparent inconsistency: §42 asks for these
explicitly, a replaceable name keeps them replaceable, and a screen with no
illustration cannot be evaluated while a project card with no icon just shows
initials.
closes#15
PersonalPredictionEngine keeps a discrete probability distribution over
candidate start dates rather than a date with a margin bolted on. Everything the
product needs falls out of that one structure: the most likely date is its mode,
the window is the narrowest span holding 80% of the mass, and a "Not yet" is the
distribution conditioned on what the user just said — which is what §13 asks for
and what a date-plus-margin design cannot express at all.
It is better, and that is a number rather than an opinion. EngineComparisonTest
scores both engines over the §51 fixtures on every build:
engine MAE mean window within +/-2 window covered
baseline 1.00 2.67 7/9 7/9
personal 0.67 4.56 9/9 9/9
COVERAGE IS THE MEASURE, NOT WIDTH
The first version of that test asserted the new windows must not be wider, and
it failed. Measuring showed why the assertion was wrong: the fixtures where the
personal engine is wider are the ones that are genuinely less certain — a
history with a suspected missing period, and one with a 45-day outlier — and the
baseline answers both with a two-day window and misses. What a window promises
is that the period starts inside it. An engine keeping that promise 7 times in 9
has a broken promise, not a tight forecast. The test now asserts coverage, with
a ceiling so "some time this month" still fails.
THREE MODELLING BUGS THE TESTS FOUND
Each was found by a test failing, not by reading the code:
- Median absolute deviation alone reads a user alternating 25 and 37 as
perfectly consistent, because half her deviations are zero. Twenty
disagreeing cycles came back High, breaking §15's rule that volume alone
must never buy High confidence. Spread is now the larger of MAD and mean
absolute deviation; robustness comes from IntervalAnalysis down-weighting
what is questionable, which is a better place for it.
- Recency weighting assumes the recent past predicts the near future. For a
variable user that is false — her latest cycle is a draw from a wide
distribution, not a signal — and weighting it equally cost three days on the
§51 variable fixture. Recency is now trusted in proportion to how much her
cycles actually agree.
- A fixed one-day floor on trend detection fired on a 42-day-cycle history
whose medians differed by a single day, turning an exact forecast into a
wrong one. One day is a real trend at 28 and rounding error at 42, so the
floor is relative to the user's own spread.
WIRED THROUGH, NOT JUST TESTED
PredictionInput carries recentAbsoluteErrors, and CycleRepository feeds the
scored errors back in. Without that the app stores every error it makes and
never reads one back — measuring accuracy rather than learning from it, with the
widening happening only in a unit test. A repository test asserts the errors
actually reach the engine.
BaselinePredictionEngine stays as the control, and both engines run the same
§51 acceptance suite, so the next engine's improvement is measurable too.
108 tests, all passing. ./gradlew check green. Verified on a device.
closes#10closes#11closes#12closes#14
checkModuleBoundaries holds the dependency tables in docs/architecture/README.md
as a check: every module's permitted project dependencies, plus the rule that
domain:cycle and domain:prediction must never apply an Android plugin. core:ads
is already in the map with an empty permitted set, before the module exists —
PRODUCT_PLAN.md §34 is non-negotiable, and a guard written alongside the code it
constrains is one shaped around whatever exception somebody wanted at the time.
It lists every violation rather than the first, and refuses to report a pass
when it examined no modules at all.
THE GUARD FAILED ITS OWN FIRST PROOF
prove-guard.sh injected a forbidden dependency into :domain:prediction and the
guard reported "7 modules checked, no violations". The root project is
configured before its subprojects, so reading subprojects.configurations from
the root script saw every configuration empty — it had been green over an empty
map since the moment it was written, and would have been trusted for months.
Collection moved into afterEvaluate, and the task now throws rather than passing
if it ends up with no modules. Three proofs recorded in the architecture doc,
all re-run and all red: a domain module reaching upward, :app reaching past the
repository straight to Room, and a module with no rule being reported as
unmeasured rather than assumed fine.
TWO REAL BUGS FROM WIRING IT INTO `check`
Running the whole check for the first time turned up Android lint errors that
would have shipped:
NewApi: java.time.LocalDate#ofInstant requires API 34 (minSdk is 26)
NewApi: java.time.LocalDate#EPOCH requires API 34 (minSdk is 26)
Both are on the recalculation path. On any device below Android 14 — most of
the install base this app targets — that is a crash. Neither the unit tests nor
the API 36 emulator could see it; lint is the only thing that could.
Replaced with atZone().toLocalDate() and ofEpochDay(0), which are API 26.
Also cleared the lint warnings that were real: a redundant activity label, and
a round launcher icon declared but never referenced. The two that remain are
deliberate and now say so where the warning is read — targetSdk 36 is Play's
floor and raising it opts into untested runtime behaviour, and the -v26 mipmap
qualifier stays because removing it makes AAPT fail to resolve the icon at all.
./gradlew check now passes with 0 lint errors across all seven modules.
70 unit tests, all passing.
closes#7
The Batch 01 vertical slice from PRODUCT_PLAN.md §58 now runs on a device:
launch, log a period, it is stored, the forecast recalculates, edit or delete it
and the forecast moves again. Hilt wiring, a TodayViewModel exposing one
immutable state, and a working surface that says "Batch 01 · working surface" at
the top so nobody mistakes it for the designed Today screen, which is Batch 03.
THE DEFECT THIS FOUND, ON A DEVICE
Tapping "Started today" twice on the same day killed the app:
FATAL EXCEPTION: main
android.database.sqlite.SQLiteConstraintException: UNIQUE constraint failed:
period_records.startDate
Not a hypothetical — the crash was reproduced on emulator-5580, the fix
applied, and the same two taps then produced "That day is already logged." with
the process still alive and zero FATAL lines in logcat.
The constraint is right: a duplicate must not overwrite the original row and
lose its createdAt and source. The API around it was wrong. Repeating a tap
when you are not sure the first one registered is an ordinary thing for a person
to do, not a fault, and it must never be an exception. So the period writes
return PeriodWriteResult — Added, AlreadyRecorded, Updated, Conflict, NotFound —
and only genuine faults still throw.
editPeriod had the same hole: moving a record onto a date another record holds.
That is refused rather than merged, because merging would delete a period the
user entered and only they can settle it.
The ViewModel now installs a CoroutineExceptionHandler as a backstop. In a
health app a crash mid-write is adjacent to losing what was just entered, and a
message somebody can read beats a process that vanished. The message carries the
exception type and never a record's contents (§45).
Four regression tests pin all of it, plus two instrumented tests on a real
file-backed database that close and reopen it — what a force-stop actually does,
and something an in-memory database cannot fail.
70 unit tests and 2 instrumented tests, all passing. Release APK 1.2 MB.
closes#6
core/data is the seam between storage and everything else. Reads return domain
types, cycles are derived rather than stored, and the forecast is a function of
the data instead of a field somebody has to remember to refresh — so §11's
"recalculate after a confirmed start, after an edit, after a Not yet" is
automatic rather than three call sites.
Confirming a period is four writes in one transaction, because a partial result
is a corrupt history rather than a failed action: write the record, score the
forecast that was standing, clear the "not yet" observations it resolved, and
snapshot a fresh forecast.
THE DEFECT THIS FOUND
A test expecting one scored prediction found three. The cause was not the test:
every historical period entered during onboarding was scoring the current
forecast against a date in the past, inventing an error for a prediction nobody
had ever been shown. §16's "your predictions are getting better" would have been
populated with figures the app made up about itself — plausible ones, which is
what makes it expensive to notice.
Two rules now, both pinned by tests:
- exactly one unscored snapshot exists at a time. A forecast superseded before
its outcome was known is not a wrong forecast, and counting it lets one
cycle contribute several errors.
- a confirmed start only scores a forecast made on or before it. Anything
earlier is backfill and leaves the standing forecast alone.
Accuracy also stays quiet below three scored predictions. One lucky forecast
reading "average error: 0 days" is an overstatement, not a measurement.
THE ROOM BOUNDARY, HELD THREE WAYS
implementation rather than api on core:database; CycleRepository's constructor
internal because it names a PeriodDatabase; reads mapped to domain types in
Mappers.kt. Callers use CycleData.repository(context) and never learn Room
exists. Verified rather than asserted: grep -rn "androidx.room" app/src domain
is empty, and Room appears zero times in :app's debugCompileClasspath.
No fallbackToDestructiveMigration: it turns a forgotten migration into a silent
wipe of the user's entire cycle history on update.
Also fixed: `domain/*` inside a KDoc silently opened a nested block comment —
Kotlin block comments nest — which broke compilation in a way the error message
pointed nowhere near.
58 tests across the project, all passing.
closes#5
core/datastore holds the settings from PRODUCT_PLAN.md §10 — notification
privacy, reminder time, the three reminder toggles, biometric lock, theme, the
ads entitlement and whether onboarding finished.
Two defaults are decisions, and each has a test whose job is to stop it being
changed by accident:
- notification privacy defaults to DISCREET (§28). A default of DIRECT would
put menstrual detail on a lock screen before the user has been asked a
single question, and a notification read over a shoulder is the likeliest
real privacy breach in this product.
- fertility reminders default to off. Most users are not tracking fertility
and an unrequested ovulation notification is an unpleasant surprise.
An unrecognised stored value falls back to the SAFE option rather than to
whatever enum entry happens to be first — a rollback or a hand-edited file must
not be able to turn DISCREET into DIRECT. Tested for privacy mode, theme and an
out-of-range reminder time.
This is a separate store rather than two more Room tables, and the reason is a
deletion semantic: Delete My Data removes the health history and must leave the
settings alone. Handing a user back a weaker privacy setting at the exact
moment they are exercising a privacy control is the worst possible time to do
it, and separate stores make the correct behaviour the easy one.
The repository takes a DataStore rather than a Context, so its 10 tests run on
the JVM against a temporary file — no emulator, no Robolectric. The Android
instance is supplied by DI at the app layer, the only place that should know
where a file lives.
41 tests across the project, all passing.
closes#4
core/database holds the four entities from PRODUCT_PLAN.md §10 —
period_records, spotting_records, prediction_records, not_yet_observations —
with DAOs returning Flow, epoch-day/epoch-milli converters, and the schema
exported to core/database/schemas and committed.
Three constraints are structural rather than remembered:
- startDate is UNIQUE and inserts ABORT rather than REPLACE. REPLACE would
delete the original row with its createdAt and source; §14 says health
history is never modified silently.
- spotting has its own table, so no query for periods can reach it. §25: it
must never start or reset a cycle.
- a prediction snapshot can be scored but not rewritten — score() sets only
actualStartDate and absoluteErrorDays. A snapshot editable after the fact
can only ever report that the app was right, which would make §16's whole
accuracy feature a lie.
deleteEverything() is one transaction and the only bulk delete in the module: a
partial wipe leaves the cycle reconstructible from the tables the user asked to
be rid of.
14 tests, on the JVM under Robolectric — no emulator.
THE SCHEMA GUARD, AND WHY IT IS A SCRIPT
SchemaTest was written as a drift guard and proved not to be one. Room
regenerates the schema export during compilation, so adding a column to
PeriodRecordEntity without bumping VERSION leaves the suite green while the
committed schema quietly changes underneath it. That was not reasoned about, it
was run: the column was added, 1.json gained it, and every test passed. On a
device that is "Room cannot verify the data integrity" — a crash on update,
after shipping.
scripts/schema-guard.sh asks git instead, which Room cannot overwrite. Proved
both ways before being trusted: green on a clean tree, exit 1 on the injected
drift. It runs in pre-commit when an entity or the schema directory is staged,
and the hook treats its exit 2 as a refusal.
SchemaTest keeps its four tests and now documents what it does not catch.
Room's own MigrationTestHelper is not used: every constructor needs an
Instrumentation and schema assets, and AGP 9's library source-set DSL throws
DefaultAndroidLibrarySourceSet_Decorated cannot be cast to
AndroidLibrarySourceSet when you add an asset directory. Recorded so the next
person does not spend the afternoon on it.
Docs updated in this commit, as their triggers required: the migration table
now has its version 1 row and the trap that makes such tables go stale, TOOLS
explains the seventh script, and the hooks README lists the new guard.
closes#3
Period was a bare directory holding one 2,527-line specification, with no git
repository, no tracker and no documentation convention. This is the adoption
from Projects/Template/START-HERE-New-Project.md, plus a project that compiles
so the hooks and future guards have something real to run against.
Documents. scaffold.sh created 19 paths, 0 skipped. The specification moved to
docs/planning/PRODUCT_PLAN.md unchanged in substance, with a status header; the
capitalised Docs/ is gone. Every scaffolded document was filled in for Period.
docs/OPERATIONS.md deleted — an offline app is not a deployed service.
DOC_TRUST_MAP.md written last, describing what is actually here, including what
this project deliberately does not have.
Code. Four Gradle modules. domain/cycle and domain/prediction are kotlin("jvm")
and cannot see the Android SDK, so the engine is testable without an emulator —
17 tests pass, 12 of them the acceptance cases from PRODUCT_PLAN.md §51.
BaselinePredictionEngine is a robust-median prototype and explicitly not the
product; it exists so Batch 02's replacement can be shown to be better rather
than merely different.
Versions verified against their official sources today rather than inherited
from the specification's own numbers, which that document asks for: Kotlin
2.4.10, AGP 9.3.1, Gradle 9.7.0, Compose BOM 2026.08.00, Room 2.8.4, Hilt
2.60.1. AGP 9 ships Kotlin built in, so org.jetbrains.kotlin.android is no
longer applied. compileSdk is 37 because current AndroidX requires it; targetSdk
stays 36, Play's floor from 2026-08-31, and the difference is deliberate.
Six scripts taken into scripts/; the rest declined and named in docs/TOOLS.md.
Three hooks in .githooks/, with pre-commit adapted to Gradle.
closes#1closes#2
null