Moving the navigation into the shared helper took both sign-in hand-offs
out of what the headerless-transport guard walks, which only reads the
component and page trees. The helper now only clears the offline copies
with the capped wait, and each page navigates itself. The device
registration test checks the end of the no-device rest inside the test
that already pairs an iPhone instead of a copy of it.
The register refusal codes, the end of the no-device rest, the
no-recipient annotation and the sign-in shell comment had each grown a
function the size ratchet holds. The codes become constants, the rest
ends through a helper beside the one that starts it, and the
annotation moves into that helper. The settings Web Push card dropped
under the limit when its subscription post moved out, so its entry goes.
The sign-in page and the profile take-over page each cleared the offline
copies with the same capped wait before leaving for the identity
provider, with the limit written down twice. leaveForSingleSignOn does
it once for both.
A send with no browser subscribed, no iPhone paired or a transport the
operator has not set up no longer counts as a failure, but nothing else
replaced it. The mood reminder and the daily briefing retry every
minute until delivered, so such a channel wrote a skipped row each
minute, for good; the dispatch still counted as an attempt, so the
watchdog and status alerts took it as announced; and the status screen
said "active" with no hint why nothing arrived.
The channel now rests for an hour after such a send, with the reason on
its status, and the send does not count as an attempt. Subscribing a
browser or pairing an iPhone ends the rest at once.
Sign-out removes the push row first, through the raw fetch helper, which
has no default timeout. On a connection that is up but carries nothing,
the sign-out waited for the browser's own timeout before the logout
request even started. The removal now gives up after five seconds.
The dose tag comment still explained the clear push sent for a logged
dose, the session-end wipe comment still named the worker's API cache as
a live layer, and the worker's header and foreign-origin note did not
match what the worker stores and when it resumes.
The Web Push test answered a refused push with `ok: false` and the
sender's `reason`, but the Test button reads `meta.errorCode`, so it
only said "Test failed". The answer now carries `upstream_rejected`, or
`connection_failed` when no push service answered.
The test push used the administrator's test title, which put the brand
first in several languages; it has its own title now.
Once the sign-in and registration pages read codes instead of the
server's prose, a taken username or address said only "Registration
failed", an expired invitation said "You don't have permission for
this", and a passkey this server does not know, a passkey sign-in that
took too long, an unreadable passkey answer, a single-sign-on-only
server and an already signed-in visitor all said "Login failed".
The registration route now names its refusals (auth.register.taken for
both 409s, auth.register.invite_invalid for a refused invitation), and
those codes, the three passkey refusals, oidc_only and
auth.already_authenticated get a sentence in every language.
The server set the Home Screen badge only through the clear push sent
for each logged dose, and that push is gone. A device whose last clear
left a count (one forgotten dose, an update in the middle of the day)
kept that number on the icon for good, since iOS never clears a web
app's badge by itself.
The worker now clears the badge once on activate. The badge handling in
the push handler, the unused outstanding-dose counter and the test of a
payload no server sends are removed, and the self-hosting guide no
longer promises the clear push or the badge.
The sign-in refusals carry only their status's generic code, so once the
sign-in page read codes, a wrong password, a wrong second-factor code
and a refused security key all said "Your session has ended", and a
lockout said "too many requests just now".
localizedSignInError reads a generic 401 as the step's own failure and a
generic 429 as "Too many attempts. Wait a few minutes and try again.";
a code the step names itself keeps its sentence. The password, passkey,
code and security-key steps use it. The rate-limit sentence moves from
auth.claim to auth.signInRateLimited, since the claim page and the
sign-in page now share it, and the Korean one takes the polite form the
rest of the catalogue uses.
Only the menu's sign-out removed this browser's push subscription. The
sign-out buttons on the registration page and the profile take-over page,
which exist so someone else can use the browser next, left it behind, and
the next person kept seeing the last account's reminders. All three now
go through one signOutRequest.
The browser's own subscription is now removed only once the server has
confirmed the sign-out; the server row still goes first, while the
session can name it. A refused sign-out stores the row again, so the
person who is still signed in keeps push. The logout test pins the new
order and the refusal.
The helper file is renamed web-push-subscription.ts, since it now also
saves a subscription for the settings card.
The Playwright project list still named the deleted onboarding AI
offer spec. The CoachPlan comments said the review writes
`outcomeEncrypted` and surfaces it behind a flag; nothing writes or
reads it now, and the column is marked internal. Comments that named
`resolveProvider` as the resolver now name the chain, and the function
itself says it is test-only.
The fixtures added with these fixes switched the type checker off with
`as never`, which the realism ratchet refuses to see grow. They now carry
their types, three webhook mocks needed no cast at all, and the lowered
counts, with the targets route's one fewer version stamp, are written
back.
The child-band change left the classification dividing by the record's
height where it had checked the adult-band height, which the type checker
refused. Both now read the same value.
Since a codeless refusal reads by its status, three answers lost their
meaning. A failed ChatGPT sign-in poll said "Error saving" instead of
asking to start the sign-in again; it now falls back to the sign-in's
own failure sentence. A dose-history import still running when the
card stopped watching said "Nothing was written" although the job may
still write; that case now has its own sentence. Setting up an
authenticator while one is active answered "setup failed"; the route
now sends `auth.mfa.already_active`, which says to disable the active
one first, in every locale.
The plan review rule told the Coach to read how a plan went from its
progress line, but a progress line exists only for an active plan, and
the review sets the plan to reviewed the morning its reminder surfaces.
For the plan being discussed there was nothing to read, so the model
could make a result up. The rule now reads a result only from a
progress line the snapshot carries, and otherwise asks how it went.
The notification categories' contract description, which lands in the
published API document, and a handful of new code comments and test
titles used em dashes; the house style keeps them out of new prose.
The form read every 409 as "a metric with this name already exists".
Now that it keeps one idempotency key across taps, a second tap while
the first create is still running also answers 409, and the form told
the person their name was taken. The create and rename routes send
`customMetric.name_taken` with the clash, and the form matches the code.
The "taken <name>" keyword looked up the first medication of that name
and only then asked whether its course had ended, so an old ended
course beside the current one of the same name answered "not found".
The lookup now reads every match and takes the first that runs.
A reminder button judged the course at the moment of the tap, so the
last evening dose of a course, confirmed after midnight, was refused
although the web still takes it. When the message names its slot, the
course is now judged at that slot. The handler file grows by six lines;
its size baseline is raised to match.
Logging the next dose satisfies a matching booster reminder with the
dose date, a calendar date stored at UTC midnight. The reminder engine
reads that instant in the person's zone, where west of UTC it is the
evening before, so the next booster came due a day early. The first
booster's anchor was already counted from the start of the dose day in
the person's zone; the satisfy path now hands the engine the same
instant.
The weight band and the BMI chart read `/api/insights/targets` inside a
shared record, and the route asked for a manage grant, so a read or
write delegate got a 403 and saw no band and no BMI. The targets are
the bands the dashboard snapshot already paints on a whole-record read
grant, computed with no provider, so the read now asks for the same
level. A section-scoped grant is still refused.
Now that the BMI chart draws inside a managed record, a child's normal
BMI sat in the red under-17 band, and the weight page painted the adult
WHO band (about 26 to 35 kg for 118 cm), so a healthy 21 kg read as
underweight. The release's own weight-band test pinned that band.
No band here knows a child's age yet. For a record under 18, or a
managed record with no date of birth, the targets read names no WHO
weight range, no BMI range and no category; the record's own weight
target still shows, and the values and the chart stay. The BMI chart
and the weight band draw only what the server names. The weight-band
test now uses an adult height and gains the child case.
The targets builder and the vitals builder grow by two lines each for
the flag; their size baselines are raised to match.
A create form kept its key until a save succeeded. The server stores a
refusal under the key for a day, so after a 422 or 409 the corrected
retry ran as a different body without any protection, and a timeout on
it could store the entry twice; an unchanged retry after the cause was
gone replayed the stale refusal until the form was closed.
Any answer now settles the key. Only a timeout, a dropped connection, or
the 409 that says the first request under the key is still running
(`X-Idempotent-Replay: false`, now `ApiError.inFlight`) keeps it, so
those retries still replay the first answer.
Three refusals answered without an error code, so the web printed the
status's generic sentence. A connection test with nothing configured
read "Some of the entries were not accepted", and a mistyped password
while adding a passkey read "Your session has ended. Sign in again",
which sent people to sign out for nothing.
An empty chain now resolves to no provider, which the test route
answers with `ai.provider.none`. The passkey registration and the
step-up mint answer a failed or stale proof with `auth.reproof.failed`.
`ai.provider.none`, `ai.budget.exceeded` and `auth.reproof.too_weak`
get sentences in every locale.
A person on the OpenAI-compatible gateway lost every AI surface once
their chain was saved at all: the fallback to the picked gateway asked
whether the chain was the default object itself, so a chain reset to
the default, reordered, or written by the gateway form's reasoning
setting (which appended a disabled entry to hold the value) no longer
counted, and the chain resolved empty. Rows written that way before
this release broke on deploy.
The fallback now reads the chain's content: the picked gateway serves
when nothing else resolves, unless the chain holds the gateway switched
off. A disabled gateway entry that carries a reasoning effort is the
form's holder, not a switch; the chain editor drops the effort of a
gateway it switches off, so a switch-off stays one. Saving the gateway
form with the gateway picked writes its entry switched on. Resolution,
the presence probe, the account payload and the iOS availability flag
all read the same rule.
The iOS app sends the date of birth as the instant its date picker
returned, and the profile writers stored that instant as sent. Since the
report cover and the share page read the date part of the stored value,
an entry made just after midnight in Berlin printed the day before, and a
Californian entry made in the evening printed the day after.
Both profile writers now store the calendar day the value names in the
owner's zone, at noon UTC like other date-only values, so an iOS client
reading the instant back in its own zone keeps the date. Older rows keep
their instant: dateOfBirthKey reads a date-only anchor (UTC midnight or
noon) by its date and any other instant in the owner's zone. The report
collector, the targets age, and the web forms that send the date back
read it through that helper.
The end-to-end setup still seeded the shared-provider offer's account
and captured its session, although the journey that used it is gone.
Three comments still pointed at what was removed: the tour launcher
cited the deleted personal-record badge as its model, the personal
record worker said the badge still renders, and the morning briefing
push said a concerning reading goes out as an anomaly notification,
which nothing sends.
The session read copies the language cookie into the account row when
the two differ, so on a demo every visitor's language still landed on
the one account all visitors share, although the locale route already
left the row alone there. Under DEMO_MODE the session read now answers
in the visitor's language and writes nothing.
The route's unit test mocked the database and the handler wholesale;
both paths are now proved against Postgres and read back from the row.
The visit, vaccination, allergy, address-book and workout forms read a
refusal's code through the catalogue, but nineteen of the codes their
routes send had no sentence: a stale doctor or check-up, a planned visit
that would close a check-up, a dose without a name, a booster that
cannot be planned, an invalid body, a workout batch that is too large.
Each of them read as the form's generic "could not be saved", which
hides what the person can change. They now have their own sentence in
all ten languages.
About forty refusal messages still printed the error string the server
sent: English in every locale, and for a dropped connection the
browser's own TypeError text. Mood tags, targets and thresholds, family
history, medication side effects, phases, API tokens and dose imports,
the Coach model picker, the AI provider, fallback chain, timeout and
Codex cards, the notification channel cards, the avatar, passkeys,
security keys, the authenticator card, the re-proof dialog and the
admin pages now resolve the refusal's code through localizedApiError
and fall back to the surface's own sentence.
Two refusals carried something a person needs and had no code: a
private AI base URL (the AI_PRIVATE_ORIGINS hint) and removing the last
passkey. Both now send a code with a sentence of their own. A throttled
re-proof says so in the reader's language. Raw fetch sites throw the
same ApiError as the verb helpers through the exported throwApiError,
which replaces the admin's getApiErrorMessage; the authenticator card's
copy of describeStepUp is gone, and describeStepUp no longer passes a
non-step-up refusal's prose through (its test pinned that).
The admin password reset keeps the server's verdict on a 422: the
server words the strength and breach verdict in the account's language.
Someone whose doctor agreed a target of 140/90 entered it on /targets;
the chart and the dashboard tile followed it, but the BP and pulse status
texts, the general status, the comprehensive overview, the insight-card
alert and `extractFeatures` (and through it the comprehensive text, the
daily briefing and the Coach) still graded every reading against the age
default. The text under the chart could contradict the chart.
All six places now resolve the band the way the charts do since #1261:
`resolveEffectiveBpTargets` for blood pressure, the PULSE override for
resting pulse, the default only where no target is set. The resolved band
rides in each snapshot's targets block, which is part of the input hash,
so the next generation follows a changed target. The Health Score keeps
its clinical targets on purpose.
The Web Push test went to the newest subscription only. With a phone
and a laptop subscribed, Test pressed on the laptop buzzed the phone
and the laptop looked broken. The text was English in every locale,
and the push was dialled through the library's own transport, past the
pinned safeFetch that real sends use.
The route now calls sendViaWebPush, so the test reaches every
subscription, reaps endpoints the push service reports gone, and goes
out through the same dial as a reminder. Title and body come from the
account's locale. The response keeps ok and latencyMs and gains reason;
sent and perEndpoint are gone, nothing read them.
The nightly plan review read a before/after window of the plan's metric
and stored an English verdict in `outcomeEncrypted` for every plan. No
screen and no route returned it; only the Coach snapshot read it, behind
the `COACH_EXPERIMENT_VERDICT` flag that is off by default. The worker now
only moves an active plan whose review fired to `reviewed`, which is what
takes it out of the standing group. The flag, its snapshot block, the
three eval cases for that block and the compose and env entries go with
it. The Coach prompt keeps its experiment rules and now ties the
read-back to the plan's progress line instead of the removed block.
Existing stored outcomes stay in backups untouched.
The plans page and the Coach reminders card printed the metric as a raw
key in capitals (WEIGHT, BLOOD_PRESSURE) in every language. Both now show
the measurement's name from the catalogue, through the plan metric
resolver moved to `plan-metric.ts`; a key HealthLog does not read shows
as written. The unused `coach.plans.proposalTitle` string goes.
Signing out ended the session and wiped the caches but left the Web
Push subscription in place, in the browser and on the server. The
browser kept showing the signed-out account's medication reminders,
also after someone else signed in on it, and the next account's
Settings showed push as already on.
Sign-out now runs the same steps as "Disable push" before the logout
call: DELETE the server row while the session is still valid, then
unsubscribe in the browser. Both callers share one helper. A failure
there does not block signing out. With the dispatcher no longer
counting "no subscription" as a failed send, signing out does not
start the five-failure countdown on the account's channel.
Each taken or skipped dose sent a `clear` push to every subscription of
the account. The service worker closed the pending reminder and showed
nothing. A subscription made with userVisibleOnly may not receive
pushes that show no notification: Safari revokes it after a few, the
next reminder meets a 410 and the channel is switched off, with nothing
in the app saying why.
The clear push, its three call sites and the badge count only it fed
are removed. A reminder already on screen now stays until it is tapped,
dismissed or replaced by the next reminder for the same dose. The
sender's badge field went too: nothing else ever set it. The parity
snapshots lose the two recorded effects.
The dashboard snapshot carries the daily briefing and its recall line,
and the persister wrote the snapshot to IndexedDB for up to a day,
although its own comment said narrative text never reaches the disk.
The persister now nulls both fields before the write. The web
dashboard reads the briefing from the Today digest, which is never
persisted, so the restored snapshot paints exactly what it did before.
A channel switched off because the last browser's endpoint answered 410
(app removed from the Home Screen, site data cleared) stayed off when push
was turned on again: the subscription was stored and the test push
arrived, but no reminder ever did. Subscribing now re-enables a disabled
channel through reEnableChannel, the same reset the status card's button
runs.
Turning push off on the last browser deleted the subscription and left
the channel on, and each later reminder was booked as a failed send
(web_push_no_subscriptions) until the channel gave up after five. The
dispatcher now leaves channel state alone when a sender had nobody to
send to; APNs without a paired iPhone and email without an address or
SMTP had the same problem and are covered by the same rule.
The resolver tests queue their fixture rows through one helper instead of
an `as never` per row, the no-provider status case also proves the
single-provider resolver is never asked, and the size and comment
baselines take the lower numbers the change left behind.
The session wipe deleted every HealthLog cache, the static one too,
although it holds only locale catalogs and icons. Each sign-in and
sign-out threw away the 190 KB catalog and refetched it.
The wipe now deletes the page cache and the API caches older releases
left behind, current and legacy versions, and keeps the static cache.
Function name and signature are unchanged.
On a cold start the IndexedDB snapshot was restored before
/api/auth/me had answered, and the owner check was skipped while the
account was unknown. On a shared browser the previous account's
dashboard could be put into the next account's cache until the refetch
landed.
The restore now waits until the signed-in account is known and runs
again when it changes. A snapshot whose owner is not that account, or
that was written before any account was known, is discarded and wiped.
Because the restore now lands after the page has mounted, it fills
queries that have no data yet instead of skipping every query the cache
already knows; data the page holds is still never overwritten.
Without a connection the strip said "Showing your last synced data",
but one tap on a tab or link traded the readable page for the worker's
bare offline page: signed-in pages are no-store, so the document
navigation Next falls back to has nothing to load.
While the browser reports offline, a tap on a link to another page of
the app now stays on the open page and repeats the hint in a toast. The
strip says what is true: this page stays readable, other pages open once
the connection is back. The README stops promising offline support.
A cold start or a typed URL without a connection still gets the offline
page; the e2e spec keeps that and adds the tap.
The worker buffered every allowlisted API read, the 60-second version
poll included, and wrote it to CacheStorage before the page got it. No
page could ever be opened offline to read those answers back, since
every signed-in document is no-store. What the cache did do: keep one
account's readings on a shared device, answer a refetch that failed
right after a save with the copy from before it, and let a late write
recreate the cache after the sign-out wipe.
API requests and the hashed /_next/static/ chunks now go to the network
untouched; the HTTP cache already holds the immutable chunks. The
cache-name pattern keeps `data`, so caches older releases left behind
are still deleted on activate and by the session wipe.
The TypeScript copy of the allowlist and its test go with it, and the
record-switch e2e spec now checks that nothing under /api/ is kept on
disk before or after the switch.
Eight call sites (Coach turn, status texts, briefing, comprehensive
insights, medication extraction, about-me questions, both lab OCR paths,
the saved-config connection test) answered an empty chain by calling the
single-provider resolver, which ignores the chain's switches. A person who
kept only the local model on and lost its address got the ChatGPT sign-in
or the operator key instead, and every copy tagged that provider as the
operator key, so consent and cost attribution saw the wrong owner.
The rule now lives once in `resolveProviderChain`, and the presence probe
reads the same rule: a saved chain is the person's word, so an empty one
stays empty. Only a chain never saved falls back, to the OpenAI-compatible
gateway picked under Settings, AI, because the default chain has no entry
for it; that entry carries its real type. The callers drop their copies,
and the presence helpers that restated the old fallback go with them.
A person who switched both the Coach and AI analysis off under Settings,
Modules still got invited to set AI up: the Insights hint, its copy in
Settings, AI, the checklist row and the onboarding panel all keyed on
`provider.canConfigure`, which only knows the record kind and the
operator switch.
The flag itself stays as it is, because the AI usage card and the consent
request read it too and still serve document reading, lab photos and
medication extraction. The three invitation readers now also check the
module map through `hasOptedOutOfAi`.
Marc-André Bombeck