The writer added in v1.38.6 compared the process's live heap usage against
80 % of V8's heap limit and aborted above it. That reading is not the
backup's footprint; it is the whole process's, garbage included. A
long-running app sits at 400 MB of largely collectable heap, so on a 1 GB
container (a 419 MB budget) the weekly pass aborted the first flush of every
account: four out of four in seven seconds, the smallest of them a demo
record whose entire stored copy is 1.2 MB, each told its record was too large
and its operator to buy memory.
Drop the gauge. What the writer holds is bounded by its shape and needs no
reading to prove: the three unbounded tables are pulled a page at a time,
every other section is released as soon as its JSON is in the sink, and at
most one flush of text is pending. The one copy that does grow with the
record is the stored blob, so the bound moves there — counted in the
ciphertext bytes actually produced, against a fifth of this process's heap
limit, which is what two copies of the answer plus a live set have to fit
inside. The message names both numbers.
A pass that wrote nothing for anybody now fails its job instead of reporting
`ok: true` with `backed: 0`, so the failing-queue panel and the backups page
read the same thing an operator would. A pass where some accounts were
written keeps passing with the failures as counts.
Measured under --max-old-space-size=450 from a heap deliberately dirtied
first: 445 000 measurements peak at 28 MB held by the backup, a one-row
record at under 1 MB.
`DataBackup.data` holds every weekly disaster-recovery snapshot and every
uploaded pack, AES-256-GCM at rest like everything else, under a column
called `data`. The rotation registry was keyed on the `*Encrypted` naming
convention, so the script never read a backup, reported zero rows
remaining, and the runbook read that zero as permission to retire the
previous key. Following the documented procedure destroyed every stored
backup — the copy an operator reaches for precisely when nothing else
works.
Register it, and register the idempotent-replay response cache the same
sweep turned up. The backup column walks in bounded batches (one row is a
whole compressed account) and re-seals the ciphertext without reading the
plaintext, so both stored envelopes come back byte-identical and so will
any later one. The cache is marked disposable: a row it cannot read is
dropped rather than counted as an error, because keeping it only
guarantees the next replay gets a body it cannot parse.
Running the script against a seeded database rather than reading it turned
up two more. It died on the mood day-context note, whose table is keyed on
the entry rather than on an `id`, and abandoned every column after it —
the backups included. And the document index's verbatim text rotated
without ever being named in the summary.
Make the naming assumption stop carrying weight. A new guard traces the
encryption helpers through their wrappers, walks every Prisma write
payload, and fails when a column receives ciphertext without being
registered. It asserts an anchored, non-zero match count so a matcher that
stops matching fails instead of passing. A second check pins each model's
primary key against the schema.
The script now ends by naming the registered columns it did NOT walk and
exiting non-zero when any is missing: a zero that means "nothing left" and
a zero that means "never looked" must not print the same. The runbook's
retire-the-key step was rewritten around that signal, and tells an
operator who already rotated on an older release how to get their backups
back.
Three assertions repeated 18.016 and the product it gave. They were the
reason the two factors could sit in the tree at once looking correct. They
take the exported constant now, and the parser test asserts the round trip
a reader on mmol/L experiences rather than a magic product.
The list renders a reading through the metric/imperial transform registry,
and glucose is not in it, so the list kept saying 95 mg/dL while the
dashboard tile said 5.3 mmol/L for the same reading — correct on both
counts and disagreeing anyway.
The edit sheet is the half that could have done damage. Its label follows
the same resolution as the row, so on mmol/L the number in the input is in
mmol/L and has to come back the other way before the PATCH. Row, edit seed
and inversion move together; converting the display alone would rewrite a
95 mg/dL reading as 5.3 the first time somebody corrected a typo.
An mg/dL account edits in the unit the column already holds and stays on
the identity path, deliberately: the converter rounds, and sending an
mg/dL edit through it would quantise a stored 95.48 the moment somebody
opened the sheet to fix a note. A row whose stored unit is not the type's
canonical one still shows verbatim, as before.
`User.glucoseUnit` has existed since v1.2 and around thirty surfaces read
it — the dashboard tiles, the glucose page, the targets panel, the CSV and
FHIR exports, the doctor report, the Coach snapshot, the low-reading push
copy. Nothing could ever write it. No route, no schema, no field on any
form, so every account resolved to the mg/dL default and the mmol/L half
of the app was unreachable code. Anyone who reads glucose in mmol/L
converted in their head on every screen.
This is the missing end: a `GET`/`PATCH /api/auth/me/glucose-unit` scalar
on the pattern the unit-preference endpoint next door established, and a
dropdown in the profile form beside it. Deliberately its own control
rather than a branch of metric/imperial: metric countries are split on
which unit they read glucose in, so one dropdown would get half of them
wrong.
Storage is untouched and stays untouched. Every reading is written in
mg/dL by every ingest path and the unit only selects a display branch, so
nothing already recorded is reinterpreted. The entry form moved with the
preference, which is the half that could have hurt: it asks in the chosen
unit and inverts through `toCanonicalMgdl` on submit, so a 5.3 typed by an
mmol/L reader is filed as 95 mg/dL rather than as 5.3 — a value the
plausibility band accepts in silence and every surface reads back as a
severe hypo.
The contract gains the endpoint and nothing else; the account payload
already carried the field.
A value imported in mmol/L was multiplied by 18.016; one displayed in
mmol/L was divided by 18.0182. The two sat in different files, both
correct-looking, and a reading that went in as 5.3 did not come back out
as 5.3. Latent while nobody could select mmol/L, live the moment they can.
One exported constant now, on the guideline figure the display path
already used, and the glucose insight page takes it from there rather than
keeping a third copy. The test names the constant instead of repeating the
digits, which is how the two came apart in the first place, and it asserts
the round trip a reader on mmol/L actually experiences.
The provider-health ledger has recorded the HTTP status of every failure
since v1.11.0 and nothing has ever read it. The operator card could say
three of five chains were failing without saying whose problem that was,
which is the one thing the operator needs: a 401 is a credential only they
can replace, a 429 or a 503 is the provider's afternoon and waiting is the
right move.
The fold carries the status of the failure whose instant the row already
shows, not an aggregate over the column. A MAX of two HTTP statuses is
arithmetic on a label, and a status describing a different failure than
the one on screen would send the operator to the wrong place. The column
joins the grouping key instead, and the fold picks the group belonging to
the newest failure.
Blank where there was nothing to record: a connection that never completed
has no status, and a chain that has since recovered had its status cleared
along with the rest of its failure state.
The import status endpoint returns the instant the Health app stamped on
the archive, the published contract requires the field, and the schema
comment said the parser filled it during unpack. It never did: ExportDate
sat on the parser's ignore list, so the column answered null on every job
ever run, for a year.
The parser reads the element now and hands the stamp to the worker on the
next chunk boundary, so the value lands while the run is still going and a
job that later fails still records which export it was working from. An
archive that omits the element, or writes a value the runtime cannot read,
leaves the column null rather than guessing at an instant.
No contract change: the endpoint and the OpenAPI description already
described this behaviour.
The compression that landed last release was real and was measured in the
wrong place. On a development machine with a multi-gigabyte heap the pass
finished; the container the job runs in is capped at 1 GB, which gives V8
a 524 MB limit, and there the same pass died of memory about fifteen
seconds in. Because the job shares the application process, that was not
one failed backup — it was a restart, and every signed-in session on the
host went with it.
Reproduced before touching anything, under `--max-old-space-size=450`
against a seeded account of 445 000 measurements, 30 000 mood entries and
60 000 intake events: `FATAL ERROR: Reached heap limit` at 34 s. The same
fixture doubled to 890 000 measurements now completes inside a 296 MB
limit, holding 16 MB of live heap where the old writer held 168 MB.
The document is written incrementally now. `buildFullBackupPayload` takes
a `deferBulk` mode that declares the three unbounded tables instead of
reading them, and `streamFullBackupJson` walks the payload in key order,
pulls those tables through page by page, and releases every other section
as soon as its JSON reaches the sink. `packBackupBlobStreaming` feeds that
into gzip and an AES-256-GCM stream, so nothing between the database rows
and the stored value ever exists whole. Both writers go through the same
per-row serialisers, and the round-trip suite pins their output byte for
byte rather than trusting that they agree.
The intake and mood reads are ordered totally now. A keyset cursor needs
it, and a partial order also let two exports of one account differ over a
tie, which every multi-dose day has.
The streamed ciphertext is a new shape at rest, because GCM only produces
its tag once the last block is in and a leading tag is what makes the old
layout unstreamable. It carries a marker no older writer can produce, so
the two families are disjoint by construction rather than by sniffing.
All three stored shapes — plain, compressed, streamed — restore, proved
both ways.
The writer also watches its own live heap and stops at 80 % of V8's
limit. An account too large for its host is now one failed backup,
counted in the run's meta, instead of a restart for everybody on it.
The stored column is the ceiling that is left: `data_backups.data` is a
single `text` value, so the finished artifact has to exist as one string,
and that is the only part of the job still growing with the record. The
operator runbook says so, with the number at which it starts to matter.
`User.locale` is the only signal a cookie-blind path has. The nightly
briefing warm, the Coach nudge, the reminder senders and the native aggregate
all resolve from that column; the screen in front of the user resolves from
the `healthlog-locale` cookie first, then the column, then Accept-Language. So
a browser can read the whole app in German while the column says English, and
while it does, nothing on screen looks wrong: the divergence surfaces as one
generated paragraph in the wrong language inside an otherwise correct page.
Keeping the two in step was a fire-and-forget PUT from a mount effect. A 401
on the paint before sign-in, an offline blip, a tab closed mid-flight — each
loses the write, and the next mount repeats the same unchecked attempt rather
than noticing the column is still wrong, so a stale column can stand for the
life of the account.
`getSession` now corrects the column when the request carries a valid locale
cookie the column disagrees with. It is ordered by the request pipeline,
needs nothing from the client, and is self-limiting: once the two agree the
condition is false. Only a cookie counts, and a cookie is only ever written
alongside a committed UI locale, so Accept-Language never stamps a guess onto
a column that generated prose depends on. The mount PUT stays for the
server-set cookie it gets in reply, which Safari's 7-day cap on script-written
cookies makes worth having; it is no longer the mechanism.
The obvious suspect was a first-paint race between two unordered writes. It is
not that: the provider takes the locale the server already resolved and
nothing on the mount path changes it, so the effect fires once with the right
value. `locale-first-paint-single-write` pins both properties so a future
mount-time flip has to face them.
The salutation and the streak label are the two strings this route still
translates server-side, and the native client is its only caller: a Bearer
request carries no language cookie, so `User.locale` is the sole per-user
signal. The coercion here stopped at English when that column was NULL, so an
account that never opened the language picker was greeted in English on an
instance configured for German, while every string beside it was German.
`resolveJobLocale` is where the fallback order lives — stored preference, then
the operator default, then English — and every background writer was moved
onto it earlier. This was the last copy of the old coercion outside that tree.
The structural guard covers `lib/jobs` and `lib/daily` only, so it never saw
this one; the two tests added here pin the behaviour directly instead.
Three pieces of code answered "what hour is it" for the salutation. The
snapshot builder computed a `greetingHour` in the user's zone and put it on
the payload so clients would not have to; no client ever read it. The
insights hero derived one through the shared `hourInTz`. The dashboard header
derived one through a local helper that fell back to the DEVICE clock
whenever the runtime could not resolve the configured zone, which is what
happens to an IANA entry the browser's data does not yet carry. At one
instant, for one account, those three returned 20, 21 and 05.
Keep the client end and drop the server one. The hour is a clock reading, not
a fact about the record, and the snapshot body serves stale for up to an hour
past its fresh TTL, so a stored hour can name a salutation the clock has left
behind. What the server IS authoritative for is the zone, and that already
travels on the same payload and is already what both surfaces read.
Both greetings now share `hourInTz`, so an unresolvable zone falls back to the
same default the rest of the app uses instead of to the device. The header's
private helper is gone; it had no other caller.
Two tests pinned the old shape by asserting the payload carried an hour.
Inverted rather than deleted: the payload must now be shown NOT to carry one,
which is the property that was missing when the field went unread for months.
`greetingHour` leaves the response contract, so a native client decoding it as
required needs the field made optional or dropped.
A `z.object()` without `.strict()` strips an unknown key and parses
clean, so a caller that misspells a field name gets a 200 and loses the
value with nothing to tell it. The worst of them was the Coach
question-dismiss endpoint, where an omitted `question` means "dismiss
all": a typo in that key wiped every pending question and answered as
though one had been dismissed.
Twenty-four inline schemas across twenty-two endpoints are strict now —
the settings toggles, the restore and bulk-delete endpoints, message
feedback, the admin user update, the onboarding acknowledgements, the
provider chain, the chart overlay preferences, the medication extractor
and the two client-side reporters.
The strip stays where it carries weight. Sync ingest, device
registration and the layout PUTs are built out of fields that shipped as
additive and optional, on the promise that a server which does not know
one ignores it; an app running ahead of the install it talks to is
routine here rather than exceptional, and refusing would turn a dropped
field into a refused batch — a year of history lost to one row, a
dashboard that will not save, a phone that quietly stops registering for
push. A read filter discards nothing when a stray query parameter rides
along, the OAuth endpoints are required by RFC 6749 and RFC 7591 to
ignore parameters they do not recognise, and an import file comes from
another version or another tool by definition. Each is registered with
its reason in a new inventory test, so a schema cannot join them without
someone making the call. That test parses the tree rather than grepping
it: fourteen of these schemas are written `z\n .object({…})` and a
`z.object(` sweep never saw them.
`sanitiseZodIssues` renders `unrecognized_keys` instead of passing it
through. Zod leaves `path` empty and puts the names in `issue.keys`, so
the sanitised issue named nothing at all and the caller was told only
that its body had failed. The names now ride a bounded `keys` field and
the message is rebuilt from them, capped at ten keys of 64 characters
and restricted to identifier characters, so a caller-supplied key cannot
echo back verbatim through the error path or into the audit ledger. The
names survive `stripValuesFromMessage` because a field name is
structure, not user-typed content.
The unit suite reported all 1937 files green and then failed the job with
`EnvironmentTeardownError: Closing rpc while "onUserConsoleLog" was
pending`, naming a different scoring test each time.
Vitest replaces globalThis.console with a sink that forwards every write
to the main thread as an onUserConsoleLog RPC call. When a worker
finishes a test file it awaits the calls in flight and then rejects
whatever was started after that point, so a console write from work that
outlived its test lands in the gap and fails the run. The blame goes to
the file being torn down rather than to the test that started the work,
which is why the named file never logged anything itself.
Three paths detach a promise and log on the handle: fireAndForget, the
stale-while-revalidate cache refresh, and the offline query persister.
They now register the handled handle in a set that both suite setups
drain in afterEach, so no test leaves work in flight. The registry is
inert outside test: nothing there awaits it, and holding a reference to
an unsettled promise would retain it for the life of a worker. No
production behaviour changes, so there is no changelog entry.
Evidence. A focused harness that starts failing background tasks and
returns hit the teardown error 10 runs out of 10 before the change and
0 out of 10 after. Instrumenting the real suite showed two detached
console writes per run escaping their test, both from the rollup tail;
after the change, none, and thirteen writes per run now complete inside
the test that started them.
The guard asserts the behaviour, not a count: a registered task must
keep settleBackgroundTasks from resolving ahead of it, and the set must
drain back to empty. It goes red on unregistered work, on a disabled
registry, on an unwrapped call site, on a sweep that matches nothing,
and on a setup that drops the hook.
The column-reader sweep read every `field:` key as a write, so
`where: { ticketHash }` counted as writing `ticketHash`. All twelve columns it
reported were a lookup key or a cron's discovery predicate being filtered on,
and a review tool that has cried wolf twelve times is not consulted on the
thirteenth. It also reported only columns that already had a write, which made
a column with no write at all the one thing it could not see.
Mask each file, bracket-match every Prisma argument, and tag its contents as
write, projection or query so nesting resolves innermost-wins. Follow a payload
assembled before the call: a const handed to `update:`, a spread merged into a
`create:`, a `rows.push({...})` builder filling a typed array. Credit raw SQL
against the `@map` column name. Report a column nothing touches as its own
category, and split the output so a write with no reader reads differently from
a query predicate, a relation's foreign key and an annotated column.
Twelve false alarms become nineteen findings that hold up. Four are worth
acting on and are left for a decision rather than annotated away: the glucose
unit preference and an import job's export date have readers and no writer, and
the provider-health status code and the known-device first-seen stamp have the
opposite.
The matcher underneath is a test now, checked by breaking the fix three ways.
The export was its only user, and the export now reads through a helper that
marks an undecryptable note instead of blanking it. Nothing else wanted the
silent null: the list DTO never carries the note at all, only `hasNote`, and
the single-resource GET has always used the throwing decrypt so a key gap
surfaces there.
Deleted rather than kept behind a dead-code exemption. A fail-soft twin
sitting unused next to a strict one is the exact shape this branch has been
clearing out — the next person writing an export reaches for whatever is
nearest, and the whole point was that the nearest thing should not silently
swallow a row. It had no test of its own, so none was removed; the three
comments that pointed at it are rewritten to say what is true now.
The weekly pass can stop and leave no trace where anyone looks. The
handler's own error never runs when the job is killed rather than thrown;
`recordError()` increments an unnamed counter; and the admin status page's
failing-queue panel looks back 72 hours, which for a weekly queue means a
permanent failure is on screen three days in seven. The backups page
itself showed an ordinary row with an ordinary timestamp, because a copy
made six weeks ago reads exactly like Sunday's until somebody subtracts
the dates.
The listing now carries two facts that fail at different times: the age of
the newest scheduled copy, which never leaves any retention window, and
the terminal state of the last scheduled run, which is the only place the
reason is written down. `readLastQueueRun` takes the newest run rather
than a window, so a weekly queue's outcome stays readable until the next
one replaces it, and it unwraps the value pg-boss stores a thrown error
under, so the page says "job timed out" instead of raw JSON. The
failing-queue reader gets the same unwrap.
Manual uploads are excluded from the cadence judgement by the shape of the
input, so an ad-hoc upload cannot make a dead cron look alive, and a fresh
instance is not called overdue for having no copy yet.
The listing also stops reading every blob into the web process to measure
it; `octet_length` answers that in the database, which on an instance with
large records was its own way of running out of memory.
The weekly pass built the whole account in memory, serialised it whole,
encrypted the string whole, and handed the result to the database whole.
That is four full copies of the record at the same moment — the object
graph, the JSON, the base64 ciphertext at 1.33x the JSON, and the copy the
driver makes of that parameter. Measured against a seeded account of
445 000 readings, the JSON alone is 242 MB and the run dies of heap
exhaustion; under a modest heap it never reaches the insert. The pass was
not slow, it could not finish.
A health record's JSON is extremely repetitive, so gzip takes that 242 MB
to 14 MB and every copy after it shrinks with it. The same run completes
in about twenty seconds with a peak heap of 434 MB and a stored row of
25 MB instead of 322 MB — which also matters because the copy lives inside
the database it would be needed to restore. Both backup jobs now hand the
serialised form back from a frame of its own, so the payload graph is
released before the compress step allocates anything.
The marker sits inside the ciphertext, so the stored column keeps the one
shape `decrypt()` already understands and rows written before this still
restore. The off-host object gets the same treatment through the version
byte its envelope always carried; version 1 objects still read back, which
both the restore drill and the restore script depend on.
The obvious saving — dropping soft-deleted readings, four rows in five on
a real record — is not taken, and the round-trip test now seeds a
tombstone and asserts it makes the journey. A tombstone inside the
retention window is what tells a synced client the row is gone and what
makes a re-sent sample a no-op; a restore without them brings back every
measurement the account ever deleted on the next device sync.
The `data-backup` queue was registered without send options, so every job
it minted took pg-boss's queue defaults: a fifteen-minute expiration and
two redeliveries. A snapshot that serialises every account's whole record
does not fit in fifteen minutes on a large one, so the pass was killed
part-way through, redelivered twice against the same record, and reached
`failed` with "job timed out" about three quarters of an hour after it
started — having never once completed.
Two hours, on the schedule rather than on the queue: `create_queue()` ends
in ON CONFLICT DO NOTHING, so a queue-level default would apply to a fresh
database and be inert on every instance that already has the queue. The
admin run-now route sends with the same options, or an ad-hoc snapshot
would silently keep the default the cron was rescued from.
The default two retries stay. The old redeliveries were pointless because
the first attempt could not succeed either; a transient failure genuinely
benefits from a second go, and the alternative is losing the week.
Derives both sets mechanically — the whitelist from docker-compose.yml, the
read set from the source — and fails when a variable is read without being
forwarded. Written red against the tree before the fix, where it named 35.
A search for process.env. alone is not the read set: four modules resolve the
name through a helper and index process.env dynamically, which hides every
OIDC_* variable and ALLOW_LOCAL_AI_PRIVATE_HOSTS. Each is registered with its
own call-shape matcher and its own non-zero assertion, so a rename fails the
file loudly instead of quietly shrinking the set. The walk and the matchers
both assert a floor, because an empty match set agrees with any allowlist.
Thirteen deliberate exceptions each carry a written reason, and the list is
kept honest from both ends: an entry that stops being read, or that later
appears in the whitelist, fails too.
Six WEB_PUSH_* names and NEXT_PUBLIC_VAPID_PUBLIC_KEY sat behind the three
real VAPID_* variables in the env fallback. None was ever on the compose
environment: whitelist and none appeared in .env.example, .env.production.example
or the env manifest, so under the bundled stack a value set under any of them
could not reach the process. Remove them and the self-hosting note that
advertised them.
`adherenceRatePercent` returns null exactly when nothing was expected, so
callers render a placeholder instead of a misleading 0 or 100. The FHIR
exporter coalesced that null to 0, which would put a clinical claim of zero
adherence into an exported bundle. The guard above it means the branch
cannot be reached today; skipping rather than coalescing keeps it
unreachable if that guard ever moves.
A deleted reading is kept with a `deletedAt` stamp, and the schema says in
as many words that normal reads filter it. Three did not.
The custom-metric correlation channel filtered the metric's own tombstone
but not its entries', so a reading the person deleted kept shaping the
correlation they were shown. The status input hash — which decides whether a
cached AI assessment is still current — filtered its measurement arm but not
its mood aggregate and not its custom-metric count or newest-entry probe. A
deletion left that fingerprint unchanged, so the gate re-dated yesterday's
text and served it as today's over data that was gone.
Pinned in the existing soft-delete guard, whose matchers were checked by
removing each filter again.
A lab note or a biomarker context written under a key this instance no
longer holds came out of a portable export as `null` — byte-identical to a
column that was never written. The restore importer then wrote nothing, so
the text was gone and nothing said it had ever been there. The whole loss
fit inside a fail-soft `catch` written for a list view, where returning null
is the right answer.
Both now export a short marker naming what happened, and the marker survives
a restore. The Coach transcript already did this with a placeholder sentence
and the doctor report with a rendered flag; the marker is now one constant
the transcript shares, and its comment says why an empty string will not do.
The unit test that required `note: null` here pinned the loss. It keeps the
half that matters — one bad row must never abort the export — and asserts
the marker instead. The round trip is proven end to end against a real
restore.
The blood-pressure medication continuity series joined every day that had a
systolic reading against `takenByDay.get(day) ?? 0`, so a day with no intake
record at all entered the scatter and the coefficient as a measured 0 %
continuity. A week away from the app read as a week off the medication.
Only days carrying an intake record are paired now. A logged skip is a real
zero and stays; a day with no record is absent, and absence is not zero.
The timezone-parity fixture in the route test recorded three logged days out
of five and relied on the other two arriving as zeros, which pinned the
defect rather than the timezone question it asks. Its two remaining days are
recorded as skips, which keeps five paired days and keeps every one of them
something the person actually logged.
`calculateCompliance` answers `rate: 100, totalExpected: 0` for a medication
that expects no dose. Arithmetically honest, and a lie on a surface that
reads the number as adherence — it also inflated the averages the AI prompts
are built from. The doctor report has excluded both the PRN arm and the
empty-schedule arm since it was written, with the reason beside it. The
insights envelope, the blood-pressure gate and the medication-compliance
snapshot filtered only PRN, and the empty schedule is reachable:
`POST /api/medications` accepts `schedules ?? []` on a scheduled medication.
Move both arms into one predicate, `expectsDoses`, and route all four
surfaces through it so they cannot drift apart again.
The mood page called the plain `pearsonCorrelation` — strength banded from
|r| alone above a floor of five pairs, no significance test — so five days
of noise sloping the right way arrived at the reader as a "Strong" badge,
with n and r behind an info icon. The guarded twin was written for exactly
that failure and names it in its own comment, and the weight, blood-pressure
and mood-status surfaces had already moved onto it. The mood page had not.
Route the page's five cross-metric channels through
`significantPearsonCorrelation` (n >= 20 AND p < 0.05) and carry the reason
a pairing was refused, so the card can distinguish "still building" from
"twenty days and nothing in them" instead of showing the same blank for
both. New copy in all seven locales.
Add a structural guard that fails when a new caller of the unguarded helper
appears outside a written allowlist. Two callers remain on it, each named as
open work rather than as a blessing: the /insights envelope and the AI
feature payload both still band at n >= 5, and moving them changes payload
shapes their own tests pin.
Dismissing a correlation pattern stores `dismissedEvidenceHash` alongside
the effect size and sample size, and the gate that decides whether to raise
the finding again read the two numbers and never the hash. The hash is the
only part of the snapshot that covers the whole evidence tuple, and a
snapshot that carries it without the two numbers — what a partial restore
leaves behind — could not be released by anything: the gate had no value to
compare and fell through to `false` on every future run.
Move the decision into `shouldReleaseDismissal`, which reads the hash twice
and for two different reasons. Evidence byte-identical to what was
dismissed settles immediately, without evaluating a threshold. Evidence
whose hash no longer matches goes to the existing thresholds when the
numbers are there, and releases the dismissal when they are not, because a
changed hash is still proof the evidence moved.
The thresholds themselves are untouched. A dismissal still has to survive
ordinary sampling noise, which is what they are for; a hash that changes on
every recomputation would otherwise undo a deliberate dismissal on the next
nightly run.
`adminSettingsSchema` validates five `assistant*` booleans and the strict
schema refuses anything else, but the route's `booleanFields` list never
carried them. Sending one alone answered 422 "No valid fields"; sending one
alongside a real field answered 200 and dropped it. The schema comment says
the entry exists so an operator scripting their settings over the single
route can carry the values, so the route is brought up to the contract
rather than the contract cut down to the route.
The flags are echoed in the GET and PUT payloads as raw column values — the
master-over-sub resolution stays with `/settings/assistant-flags`, which
owns that shape and was never affected.
With the switches now enforced at dispatch, a user could still flip a
channel on, see the toggle stick, and get nothing. Refuse the enable
instead, with 403 and a message that names the reason: the ntfy and
Telegram config endpoints on both accepted body shapes, and the Web Push
subscribe endpoint, which is the enable step for that channel because it
creates the channel row.
Telegram is checked before the webhook registration, so a refused enable
makes no outbound call and stores nothing. Disabling a channel keeps
working while the switch is off — a user must always be able to turn a
channel off, whatever the operator has done.
The three refusals are published in the contract; `/api/settings/global-services`
already exposes the switches so a client can hide the control instead of
provoking the refusal.
`AppSettings.telegramGlobal`, `.ntfyGlobal` and `.webPushGlobal` were read
in two places, and neither of them was the one that matters: the setup-card
visibility on Settings -> Integrations and an echo in the preferences
payload. The dispatcher never looked at them. An operator who switched a
channel off instance-wide lost the setup card and kept every existing
delivery on that channel, indefinitely.
Gate each of the three in the dispatcher, in the shape
`isApiGloballyEnabled()` already uses for the twelve API routes. The read
sits above the per-user channel row and below the per-event preference, it
happens at most once per dispatch, and only when a switchable channel is
actually in the cascade. A skip writes a `globally_disabled` row to the
push-attempt ledger so the admin diagnostics name the cause, and does not
count as an attempt, so `dispatched` keeps telling the truth. The account's
own channel row is left enabled: the switch is the operator's, and flipping
it back has to restore delivery without every user re-enabling anything.
APNs, the generic webhook and email carry no instance-wide switch and are
not gated here; they are turned off at their source instead.
The unit test drives the real dispatcher over a stubbed settings row; the
integration test writes the actual singleton and reads the ledger back out
of Postgres, which is the layer that proves the column the admin route
writes is the column the dispatcher reads.
Neither read is a bug a behavioural test can see: both shapes return the
right rows, and only the plan differs. So the guards split by what each
can actually answer.
The unit guard asserts on the SQL the readers build, with comments
stripped so it matches the query text and never the prose describing it:
no DISTINCT ON on this path, a LATERAL driven from an unnested type list
with a LIMIT on the per-type branch, a bound type predicate on the rollup
read sitting alongside the user and granularity filters, and the route
still calling the shared readers rather than growing a second copy. Four
of its five cases fail against the pre-change SQL.
The integration companion runs the shipped readers and the pre-change SQL
side by side over one fixture on the real Postgres and requires identical
rows. The fixture carries the cases the rewrite could have broken: a type
whose only rows are tombstoned, a type with no rows, a type whose newest
row is a tombstone over a surviving one, and coarser rollup granularities
the DAY read must ignore. Dropping the soft-delete filter or shortening
the type list makes both cases fail.
Plan assertions are deliberately not in the integration file: a
testcontainer fixture is too small for the planner to choose the access
paths a real account produces, so an EXPLAIN there would pin nothing.
Two reads behind the metric tiles scanned far more than the rows they
return, because PostgreSQL 16 has no index skip scan: a query that wants
one row per measurement type but never names a type cannot be served by
any index keyed on (user_id, type, ...).
The latest-reading read used DISTINCT ON (type), so it produced every
live row of the account in (type, measured_at DESC) order before keeping
one row per group. On a multi-year account that is a sequential scan plus
a sort that spills to disk once it outgrows work_mem. It is now a LATERAL
join over the type list, one lookup per type against the live-row partial
index from migration 0183, each stopping at the first row. CROSS JOIN
LATERAL is an inner join, so a type with no live reading drops out as
DISTINCT ON dropped it, and ORDER BY reproduces the same enum ordering.
The sparkline read filtered measurement_rollups on user_id and
granularity only, skipping the second column of the table's key, so on a
single-account host neither the primary key nor the user/type/
granularity/bucket index could drive it and every stored granularity got
scanned. It now names the full canonical enum, which filters nothing out
and restores the equality the key needs; the existing index serves it.
Measured on a fixture matching a real account (454k measurements, 79%
tombstoned, 75.5k rollup rows): the first read goes from a seq scan plus
a 3.2 MB external-merge sort at 112 ms to a nested loop at 0.7 ms, and
the second from a seq scan discarding 72k rows at 6.9 ms to a bitmap
index scan at 1.6 ms. Same rows, same order, same types.
Both reads move out of the route into src/lib/dashboard/summary-reads.ts
so the integration suite can exercise the SQL itself rather than a copy
of it, the same seam buildMedsTodayBlock already uses. No schema change:
adding an index on (user_id, granularity, bucket_start) would have landed
in the same range while leaving the existing index unused and adding a
fourth write target to a table the rollup workers upsert into constantly.
The inventory walked a flat, hand-written array of route trees, so its
coverage grew only when somebody remembered to grow it. It had stopped
growing: the coach, environment, mental-health, inbound-document and MCP
trees were all outside it, and so was the doctor-report export route the
header already named as a doctorReport surface. Every route in the first
four turned out to be gated, so nothing leaked — but that was the authors'
care rather than the inventory's, and a green run claimed more than it
covered.
The trees are declared per module key as an exhaustive Record<ModuleKey, …>:
a new registry key does not compile without an entry, one test asserts every
key is declared, another asserts every declared tree exists, and a third
discovers an API directory named after a module key on disk and fails when
it is not declared. An empty array is now a claim with a reason beside it
(glucose and recovery ride the measurement engine), not a silence.
Walking the newly covered trees classifies every route: coach, environment,
mental-health, the document vault and /api/export/health-record all gate
directly. The seven /api/mcp routes are the credential surface behind the
remote endpoint and go on the exempt list with the reason written out — the
endpoint itself lives outside /api and carries the gate, so a credential
minted while the module is off is inert. That refusal is added to the
behavioural list so deleting its proof fails here.
/api/export/health-record joins /api/fhir as a key-pinned tree: both serve
the whole-record aggregate, so doctorReport is the only correct key.
The Withings entrypoints, the Telegram bot webhook and the Coolify deploy
hook limited POST and left the verification verbs unlimited. GET and HEAD
compare the same secret and answer 200 against 401, which is an oracle for
guessing it and a free load channel besides; timingSafeEqual protects the
timing channel, not the status code. The project's rule is that the
per-source limit runs before secret verification, and it was being applied
per handler rather than per file.
Each verification verb now runs the same limit its POST sibling used, on the
same bucket. Telegram and the deploy hook grow the small local helper the
Withings routes already share. HEAD re-shapes the limiter's JSON envelope
into a bodyless 429 and keeps the rate-limit headers.
The new guard walks every webhook route from the filesystem, resolves which
local helpers reach a comparison and which reach the limiter, and fails when
an exported verb reaches a comparison without the limiter in front of it. It
asserts a non-zero match count so an empty match set cannot read as green.
WHOOP already passed: it exports POST only.
The lockstep guard compared only the pins both matchers happened to
recognise, and neither matcher said so when it recognised fewer.
Workspace side: the key had to be quoted. Every entry in the file carries
an `@<selector>` suffix and therefore needs quoting, so the matcher looked
complete — but `mysql2: "^3.22.0"` is a valid pnpm override with no
selector and no quotes, and it parsed to nothing. Unquoting that key and
pointing the Dockerfile at `^1.0.0` left all six tests green.
Dockerfile side: only `npm pkg set 'overrides.X=Y'` was recognised, so
re-quoting the same line as `npm pkg set overrides.X=Y` removed the pin
from the comparison without changing what the image installs.
Both sides now parse line by line: quoted or bare keys and values on the
workspace side, any shell quoting of the same assignment on the Dockerfile
side, plus a raw JSON `"overrides"` block. What neither table can read is
collected and asserted empty, so a future spelling fails loudly instead of
shrinking the compared set in silence.
The operator card offered a free-text base URL with an OpenAI placeholder
and no word about what else it takes, so an Anthropic endpoint looked
unsupported even though the field accepts that host. One line beside the
existing hints, in all seven locales, plus the changelog entry.
Marc-André Bombeck