The import recognised a deleted result by the identity on its own
tombstone, so anything that removed the row brought the result back
with the next Apple Health export or resend: deleting the whole marker,
which hard-deletes its readings, the nightly tombstone purge, and a
portable backup, which leaves deleted readings out.
The identities now move into an import ledger (LabResultImportKey,
migration 0501) in the same transaction as the marker delete and the
purge, and the import checks it beside the stored rows. The backup
carries the ledger, a portable file also the identities of deleted
readings still waiting as tombstones; the restore writes it back
without the identities a restored reading holds. The data wipe clears
it, account deletion cascades.
Both the Health Records import and the CSV import checked the unit
before writing and refused a numeric reading against a marker whose
unit was still empty, although the write itself lets the first number
choose it. An empty marker unit now fits any unit in unitFitsMarker;
the write's conversion stays the authority.
The per-type reads descend measurements_live_covering_idx once per type and
stop at their cap, but the planner chose that descent only when its row
estimates favoured it. The type is a lateral parameter, so one generic
estimate covered every type; an account whose rows arrived after the last
ANALYZE was expected to hold about one row; and statistics taken while the
table held one type made the type condition look free. In those states the
planner walked the type-less (user_id, measured_at) index through every row
of the account once per type, or read every row of each type into a sort.
On a 9,300-row account that was 33,900 row visits for 6,300 rows returned,
1.66 million for the span probe and 845,000 for the dashboard's latest
reading per type.
One shared fragment now holds the descent: the account and the cap reach
the planner through scalar subqueries, and the type is matched as a closed
range it is also ordered on, so only an index leading (user_id, type,
measured_at) can return the rows without a sort. readNewestRowsPerType,
readLiveMeasurementSpan, the dashboard's readLatestEver and the latest-instant
probe of readLatestCanonical use it, and readLatestReadingsPerType reads
through readNewestRowsPerType.
The plan guard moves to its own file, which builds four statistics states on
purpose (none, predating the account over many types and over one type, and
analysed with other accounts) with autovacuum off, and holds every read to
its bound in each. The previous guard depended on statistics another file
had left behind, so it failed in the full suite and passed alone.
Refs #1398
Integration coverage for the nutrient batch under each credential (the
source row, the 403s, the module gate, re-post, backfill, coexistence
with the phone's row and the bridge bucket), kcal and g ingest with kJ
skipped, the day fold on the overview and in the correlation read, and
composition refusing energy, macros and alcohol. Route-level checks run
against Postgres rather than a stubbed handler. The energy card reports
a failed save through the shared error toast.
Nutrient targets travel in the account-settings slice of the full
backup and are admitted code by code on restore, so an unknown code or
an out-of-range value is dropped and named while the rest come back. A
data wipe clears them. The targets route joins the OpenAPI registry,
and the regenerated contract carries the wider nutrient enums, the new
sources and the target and authority fields. Refs #1402.
The export import through the real worker (LOINC matching, unit
conversion, panels, skips, a hand-typed duplicate, re-upload and a
deleted result), and the batch route (resend, an export-written result
recognised, token scope, size and shape refusals, idempotent replay).
Both the disaster-recovery and the portable file now carry
lab_results.external_id, and a restore writes it back, so a restored
account still recognises a Health Records result when it arrives again
and a deleted one stays deleted. Older files restore unchanged.
The bulk measurement reads behind the insight features, the Coach snapshot
timelines, the mood cross-metric pairing, the correlation windows and the
weight and blood pressure status cards read newest-first across all their
types under a single row cap. On an account with a wearable heart rate
stream the cap filled within hours, and every weight, blood pressure or
other sparse reading fell out of the read, so blocks reported no data for
metrics with years of it.
All of them now go through one read that descends the live partial index
once per type, each type stopping at its own cap. The feature extraction
and the Coach snapshot cap each type at 6000, so a block built from one
type's read equals the block the full read builds (worst case 72,000 rows
for the twelve feature types). The mood, correlation and status reads keep
their old totals and share them fairly: each type keeps at least its share,
a sparse one all of its window.
The feature context's data span now comes from index probes over every
type in the window rather than from the capped rows, and its measurement
count is the readings the blocks were built from.
Refs #1398
A marker minted from a text result (OCR, CSV import, inbound document)
stores an empty unit: none has been chosen yet, and its first numeric
reading sets it. The editor read that empty unit as a unit still to be
typed and refused to save, ticking "No unit" sent "1", and the update
route refused any unit change on a marker with readings. Every save of
such a marker failed, a plain rename included, because the form always
sent the unit.
- The form opens an empty unit with "No unit" ticked and sends the unit
only when it changed; the empty unit and the unity count as the same
"no unit" there, so a rename leaves the stored unit as it is.
- PUT /api/biomarkers/{id} locks the unit only on a real change while
numeric readings exist (deleted ones included). Another spelling of
the stored unit is no change, and text results carry no number the
unit could misstate, so a text-only marker may take any unit.
- The empty unit stays distinct from "1": it still adopts the unit of
its first number, which "1" must not. No stored value changes.
Refs #1393
A night on which every account was skipped, for instance because every
provider was paused for background work, wrote all_failed false. The
repeated-failure alert read that as a successful run: it ended the
streak and announced a recovery while the provider was still down, and
nights then alternated between failed and paused so the streak never
reached the alert again.
The run now also records run_attempted. A completed run with
run_attempted 0 is neither a failure nor a recovery: the alert and
readConsecutiveRunFailures step over it.
Refs #1384
The nightly briefing pass settled an account as failed whenever its
generation failed, including when every provider of its chain was
benched for background work and no request was sent, and when the
provider refused the account's own credential. A key revoked at its
issuer stays saved and is refused every night, so a single-account
instance read as all-failed night after night and the repeated-failure
alert reported a broken job.
- A chain that sent nothing settles as skipped (provider_benched), with
no 45-minute retry.
- A refused credential of the account's own settles as auth_failed,
which the run reports as partial but does not fail over; a refusal of
the operator's shared key still fails the run.
- The generator's failed outcome names the provider that failed first,
whether its credential was refused and whether anything was sent.
- The run's facts carry failure_cause and failure_provider.
The admin Background jobs card listed only pg-boss failures, while the
alert also counts runs that completed with every account failed, so an
operator sent to the card read that nothing had failed. The card now
lists those runs, and the card and the alert describe them in words
with the cause and provider instead of quoting the run's JSON.
The warm tally counted a status card that fell back to its deterministic
text as a warmed assessment; it now counts model-written notes only.
Refs #1384
Directions such as "30 to 60 minutes before food" had nowhere to go: a
medication carried no free text of its own, only the notes on a dose
change or a course. The medication now has one note, edited on the last
step of the wizard and shown under the name on the detail page, above
the dose actions on both card variants, under the picked medication in
the log-dose sheet, and per row in the take-all-due list. One
MedicationNote component renders it everywhere, in foreground text.
The note is sealed at rest in Medication.noteEncrypted (migration
0498_medication_note, lock-bounded), with no plaintext column. Create
and update take an optional `note`: trimmed, at most 500 characters, an
empty string or null clears it, absent keeps it. The wizard sends it on
an edit only when it changed, so an unrelated save never clears it. The
list and detail reads strip the ciphertext and answer `note`.
Both ends of the data path carry it: the backup (decrypted in a portable
file, ciphertext in a disaster-recovery file), the restore, the data
export, the key-rotation registry and script, and the MCP medication
schedule, resources and fetch, fenced as user text. The doctor report
and the FHIR export leave it out: it is the person's own reminder, and
the report has no selection for personal prose.
The size baselines rise by the one-line wiring of the field in the
medication page, detail tabs, cards, log-dose sheet, wizard step and
list read; the wizard payload grows by the field's own encode and
hydrate.
A script or a home-automation rule holding a measurement token can now
post a value to POST /api/custom-metrics/{id}/entries for a custom metric
its owner already created. The route names MEASUREMENTS_WRITE_SCOPE on
the POST only; the value list, the metric list and every create, edit,
delete and restore leg still name no scope and refuse the token 403.
The token stays on its owner's record: requireAuth refuses any account
selector, and an id that is unknown, deleted or another account's answers
the same 404. The entry has no source column, so the audit row names the
token that wrote it.
The bearer-scope guard now freezes the scope's reach at three files, with
the custom-metric file naming it once and only on the POST. Integration
cases cover the admitted write, idempotent retry, the foreign and deleted
ids, five other narrow scopes, every other custom-metric leg, and a shared
record. The OpenAPI descriptions and the token card text in every locale
say what the token can do and where a metric's id is found.
Refs #1377
The Fitbit guide still called the integration experimental and spoke of
the 30 October 2026 turndown as ahead. Three new files carried em
dashes in a doc comment and two test titles.
The case with a year of earlier takes proved the ledger stays the same,
which the old read of every intake satisfied as well, so a revert of
the bounded read passed every test. It now also counts the rows the
route read and expects only the window's.
Several test titles and comments said whose readings a fixture came
from, with a host name beside them: a blood pressure average, a month
of paired readings with dates, a cached insight about weight, cycle
history and dose timing. They now describe the fixture itself (a
profile averaging 134/87, a ten-pair month, a long history); the values
and what each test proves are unchanged.
Subscribing re-enabled a disabled channel but left the failure counter
and cooldown on one that was still enabled. Failures counted before the
new subscription (no browser left, a push service outage) kept every
reminder skipped until the cooldown ran out, and one more transient
failure switched the channel off. Subscribing now resets them too.
A browser also kept the row of an account whose session ended without a
sign-out, so the next person on it saw that account's reminders. An
endpoint is one browser: subscribing now removes other accounts' rows
for it, and the settings card, which posts its subscription on open,
claims the browser for whoever uses it now.
Before this release, a send with no browser subscription counted as a
failure, and five of them, or one 410, switched the WEB_PUSH channel off.
Subscribing again stored the browser but left the channel off, so the
card read "enabled", a test push arrived, and no reminder was ever sent.
Subscribing now re-enables the channel, but nothing prompts the people
already caught to subscribe again.
The migration re-enables every WEB_PUSH channel the dispatcher switched
off (a disabled_reason is set; a channel switched off by hand has none)
whose account still has a subscription, and clears the counter, the
cooldown and the last failure the same way reEnableChannel does.
The tab strip on every Insights page read the whole overview (22 KB),
a workouts page, the ECG list and the analytics summaries to learn which
pills to show. GET /api/insights/presence answers just that, for the
record on screen: the reading count per type after the record's module
switches, and whether it has mood entries in the last 90 days, a tracked
scheduled medication, workouts and ECG recordings. A flag whose page a
switched-off module hides is false and its rows are not read. The strip
now reads this and the saved layout; each page keeps loading its own data.
The cases for the bounded read asserted the query's where clause on a
mocked client and cast their fixtures past the type checker. Against
Postgres they now show what the read promises: a rolling medication last
taken before the window shows the dose that fell due inside it, and a year
of earlier takes answers the same ledger as a short history.
The route test that pinned the cached answer replaced the database client
and the engines, so it saw the calls the route makes and not what a reader
gets back. The integration test drives the shipped route: a reading written
past the API leaves the second open on the stored answer, while a
measurement written through the API and a module switch both show it.
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.
An unreadable provider key is now reported once: the empty-chain fallback no longer opens the same key a second time. Two files shrank, so the size baseline steps down.
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.
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 notification matrix listed "Unusual measurement", "Low adherence",
"Withings sync failed" and "Personal record" with a switch per channel,
three of them on by default, yet nothing dispatches any of them: a reading
far outside its range or a week of missed doses produced no message.
eventTypesForRole now leaves them out for every role, next to the
admin-only filter, and the preferences read no longer serves rows stored
for them. Their labels in all locales and their tap targets are deleted.
The types stay in EVENT_TYPES, so the preference write and the stored rows
keep parsing; no migration.
The label and tap-target tests walked EVENT_TYPES and so pinned the four
in place; they now walk the event types that are sent.
The report cover, the PDF author field, the FHIR Patient and the share
page printed `fullName ?? username`, and the collector never read the
display name. A managed record has no way to get a full name, so its
generated `managed-<hex>` login landed on every clinical document made
for it; an own account printed the login instead of its display name.
The collector now resolves the name once (`patientName()`: full name,
display name, login, never a managed record's generated handle) and the
five readers print that one field. `patient.username` and
`patient.fullName` leave the report payload; nothing else read them.
Every browser test iframe is named "vitest-iframe" and the Playwright
provider looks the frame up by that name. Once a page runs a second
file, the first file's iframe can still be found until its detach event
arrives, so the second file's first click lands in a frame that is going
away and fails with "Frame was detached". With fifteen browser files and
at most twelve pages, three pages now run two files each, and a full
local run failed this way in four of five tries. The setup now waits
until the frame found by name is its own.
The revoke audit is now asserted in the integration race test where a
rotation commits while the turn-off waits, instead of a route test that
mocked the handler and the client. The other new tests type their
fixtures rather than casting them away.
A reading entered in mmol/L is stored as whole mg/dL, and the index
divided the unrounded conversion, so a 5.2 / 0.3 pair printed beside a
17.4 rather than 17.3. Round the glucose to the decimal it is shown with
first and divide that. The tiles also printed a two-decimal ketone
reading to one decimal; show it as recorded.
A rolling reminder ("every N days") is anchored on its last reading, but
the reminder tick rolled any cycle of a week or less on after a delivered
reminder, re-anchoring it on the send. A missed every-two-days blood
pressure reading therefore moved two days ahead each evening, read "in 2
days" on the card, and never showed as today or overdue.
A rolling reminder now stays due after its reminder, whatever its
interval, until a reading lands or it is done, skipped or snoozed. While
it stays open the repeat follows its interval, capped at a week, still in
its notify hour and under the per-day claim. Calendar rules that recur
within a week (a twice-daily course) keep rolling on; appointments are
unchanged.
holdsOpenAfterReminder stays the one definition for the tick and the
edit route. The daily digest asks the narrower question underneath it,
hasLongCycle, so a short rhythm still waits for its time of day on the
start page and is not pinned under the rail cap.
Refs #1369
DELETE revoked the one feed it had read, so a rotation committing while it waited left the successor live behind a 200. It now revokes every live feed of the record in two passes. A rotation that loses to a turn-off answers 404 cycle.feed.none instead of 409; PATCH documents its 409. The in-repo feed guide describes the three forecast styles.
A domain with no new rows kept its last row's stamp, so once that was older than the tombstone horizon every pull answered cursorExpired and the client resynced from scratch on every pull. A domain whose page runs dry now takes serverNow minus the overlap on every page where its watermark has not passed that line, and on the last page every row domain does, raising a quiet domain and holding back one that delivered rows in the stretch. A domain drained early in a long walk is covered as well. The overlap grows from 60 to 120 seconds, above the 60-second timeout of the longest transactions that write served rows.
Ketones had to live in a custom metric, with the glucose-ketone index typed
in by hand beside it, where it drifted from the readings it came from.
- Type: BLOOD_KETONES (beta-hydroxybutyrate, mmol/L, migration 0496),
plausible from 0 to 10: the blood ketone meters read 0.0 to 8.0 and print
HI above it, and a venous value off a lab sheet fits under 10. Zero is a
real reading. No band, no direction, no personal record, off the Coach
snapshot and the correlations; readable over MCP and owned by the glucose
module like glucose.
- Entry: choosing blood ketones in the measurement form also takes the
glucose from the same fingerstick, in the account's glucose unit. Both
rows go through the array arm of POST /api/measurements with one
timestamp, so they land in one transaction or not at all.
- Index: glucose (mg/dL / 18.0182, the tree's one factor) over ketones,
computed at read time and never stored. Each ketone reading above zero
pairs with the glucose reading at the same instant, else the nearest one
within 15 minutes either side; a tie goes to the same source, then to the
earlier reading. Nothing pairs across a wider gap.
- Surfaces: GET /api/measurements/series takes kind=ketones and
kind=glucoseKetoneIndex (each point names the paired readings), and the
day carries glucoseKetoneIndex when both reading types are visible. The
new Ketones page charts the readings with the index beneath; the day view
shows the index beside the value tiles; the doctor report, its PDF and the
clinician share print it in the glucose group when both leaves are chosen,
since beside the ketones it gives the glucose back.
- FHIR: ketone readings export under LOINC 104816-4 (Beta hydroxybutyrate
[Moles/volume] in Blood). The index is left out: it has no LOINC term,
and both readings it is built from are in the bundle.
- The measurement form's action row moved into its own component to keep
the form inside its size budget.
OpenAPI regenerated; all ten locales; the import unit table lists the type.
Rotation revoked the feed it had read without checking it was still live, so a turn-off in between came back under the new link, and it copied options read before a concurrent change. The revoke is guarded and the options are read after it; PATCH writes only to the live feed and follows a rotation to its successor.
A page's watermark is its last row's stamp, and a long import stamps its rows before it commits them: a quick write committing in between carried the watermark past them, and the keyset never returned them. A caught-up page now hands out watermarks at most a minute before serverNow in every row domain, as start=now does, so the next pull reads that minute again. The mirror domain keeps its watermark.
The labs list layout gains categoryOrder, a bounded allow-list of
category keys (200 entries, 64 characters each, deduplicated, malformed
values dropped on read). GET /api/labs/overview reads the caller's own
layout and answers categories already in that order, so every client
renders them as given. Listed categories come first; keys the data does
not have are ignored; a category the order does not name follows in
the default order (catalog panels, then the person's own by label, then
other). Other left at the end of the order stays last, behind any
category added later; moved anywhere else, it stays there.
Settings > Arrangement > Labs shows a category order card while
categories are on. It lists the overview's categories under their
localised labels, sends the whole order after every move, moves the
cached overview at once and refetches it. The card says how the default
order works. The OpenAPI document carries the new field.
A forecast period was one all-day event over every day it might fall on,
so a period of a few days showed up as a week-long estimate next to the
logged ones. Each feed now carries a forecast style:
- EXACT: the calculated period only, from the most likely start for the
predicted number of bleeding days.
- MARGIN: that event, plus separate "possible" events for the days it may
start earlier and the days it may still run on, each only where the
band reaches that side. Discreet titles use an open dot for them.
- BAND: the single range, unchanged.
Migration 0495 adds the column with BAND as its default, so every live
feed keeps the events it serves. A feed turned on from now on starts on
MARGIN, set by the mint route. PATCH changes it, rotate copies it, the
status and the audit diff carry it, and the settings card offers it with
a one-line explanation of the chosen style. The horizon and its stop rule
are the same in every style, and the period event keeps its UID across
styles so a calendar updates it in place.
Four fixed rounds could all finish one rotation before the other started on
a fast runner, so the guard that the race really ran failed with eight 201s.
Rounds now continue until a 409 shows up, with the mint bucket emptied
between them, and the guard still fails if no pair ever collides.
A link subscribed on a phone, a tablet and a computer at a five-minute
refresh came close to the 30 fetches an hour a link was allowed, and a
household behind one address shared 120. The per-link bucket is now 120
an hour and the per-address one 600. The token is a random secret checked
by its HMAC, so guessing links is not what these limits stop.