A gateway that refuses reasoning_effort outright (LiteLLM in front of
Ollama: "does not support parameters: ['reasoning_effort']") in front of a
model that reasons without answering sent requests until something outside
stopped it. The answerless rule learned to ask with "none", the refusal of
that "none" was learned as "send no field" over it, the next request came
back answerless again, and each lesson undid the other.
Each call of the Local and the OpenAI-compatible gateway clients now carries
one request budget through every repeat (request-hops.ts): at most two
requests, after which the call ends with the failure it has. A structural
guard holds both clients to sending only through the budget and to handing
it to every repeat.
The refusal is now learned as what it is: an endpoint that refuses the
field, in front of a model already seen reasoning unasked, keeps both facts
(FIELD_REFUSED_REASONS_REGARDLESS), so it is never asked with "none" again
and keeps its thinking room. An answerless reply from an endpoint that
refuses the field is final. When the refused "none" was the repeat of an
answerless reply, the call fails as reasoning_only, which is what it ran
into. The refusal handling of both clients lives in wire-refusal.ts.
A status run that hits its bound now aborts the call it gave up on instead
of leaving it running detached.
Refs #1279
A status note whose generation stalled or failed opened a fixed five-minute
negative window, and only the read path honoured it. The ingest
invalidator counted only a stored note as fresh, so every data sync
re-enqueued each note whose last generation had just failed, and the
worker generated it again whatever the window said. With a model that
never answers, every sync bought another round of calls per note.
The window now grows with the generations in a row that ended without a
note (failure_streak, migration 0480): five minutes, doubling per failure,
up to two hours. A written or re-dated note resets it. A failure inside an
open window neither moves nor counts it.
runStatusCompletion waits the window out before any provider work,
whichever path selected the note again (a sync, a warm pass, a queued
job), and the ingest invalidator no longer enqueues a note inside its
window. A derived-score assessment that fails now opens the same window.
The person's own regenerate closes the record's windows first, and saving
the AI provider settings closes them and forgets the provider's health
rows, so a fixed setup does not sit out a pause the old one earned.
The window code moves to status-negative-window.ts; the status parity
golden records that the mood note screened earlier in the same minute is
no longer generated again inside its window.
Refs #1279
The document picker showed a document deleted in HealthLog as "Deleted
in HealthLog" and would not let it be ticked, so a person who deleted an
imported document by mistake met a dead end exactly where they would look
for it again.
Such a row can now be ticked, and the import sends revive: true for it.
A document still in Recently deleted is restored through the shared
restore, nothing is fetched, and the row reads "Restored from Recently
deleted". One the purge already removed is fetched and stored again, and
only once it is stored does the ledger forget its key, so a failed fetch
leaves the deletion standing. Without revive the route answers "deleted"
as before, and the token upload has no such switch: a Paperless workflow
or the import script re-sending a deleted document is still refused,
whether it is in Recently deleted or long purged.
The search names the document a deleted key still points at (documentId,
null once purged), so the row can say "In Recently deleted" rather than
"Deleted in HealthLog". The outcome gains "restored", sent only in answer
to revive. The row logic moves to picker-rows.ts.
Refs #1275
Deleting a document only stamps a tombstone and the purge removes it
after 30 days, but the only way back was the undo toast, gone a few
seconds after the delete. Recently deleted lists the deleted documents
of the record with type, filing date, where an import came from, when
each was deleted and how long it has left, and offers Restore, Restore
all and Delete now.
GET /api/documents/inbound/deleted lists them (owner-only, capped at 200
with the list-cap meta). POST /api/documents/inbound/{id}/purge removes
one deleted document for good through the same function the nightly
purge runs, so its import keys and aliases still move into the ledger
and an importer re-sending it is still answered "deleted"; a live
document answers 409 documents.inbound.notDeleted. The usage read gains
deletedCount, 0 under a grant. The restore route and bulk restore
share one restore, which now also clears ledger entries for the
document's own keys.
The vault header shows a quiet "Recently deleted (n)" link only while
something is in it, and the delete toasts and confirmations say that
deleted documents stay there for 30 days. The purge's queue facts and
grace window move to document-purge-queue.ts so the worker registry
still loads none of the work.
Refs #1275
The health ledger already counted every failed call and paused a provider
that kept failing, with a pause that grows. A paused provider only moved to
the tail of the chain, though, and the tail is still walked: a chain whose
only entry answered every call with reasoning and no answer was called by
every background job regardless.
A walk for background work (surface "job") now leaves out every provider
the ledger has benched until its window lifts; the local floor only once
it is paused, past a brief backoff. One ai.chain.background_paused line
per window says so. A person's own call still reaches the provider at the
tail and reports its own error. The ledger is in Postgres, so the bench
holds across workers and restarts.
The status, briefing and document-index paths refunded a failed call in
full, so background spend on replies without an answer never reached the
day's token ceiling. A failed chain now carries the tokens its hops
reported (tokensSpent) and those paths charge them.
The two clients share one answerless step (askAgainUnreasoned), which
keeps both inside the source size ratchet; its baselines go down.
Refs #1279
The reasoning separation grew three files past their size ratchet. The Local
client's reasoning collector and the new separation helpers move to
local-reasoning.ts, the gateway client's thinking room and reply separation
become two private methods, and the Coach's error-code resolver moves out of
chat-bubble.tsx into coach-error-key.ts, still re-exported from
message-thread.tsx. Code motion apart from comments that said when a branch
was added rather than what it does. The baselines record the shorter files.
Refs #1276
The summaries slice and the comprehensive aggregator took their JS windows
(the DAY-bucket bounds, the rollup slope anchors, lastSeen staleness) from
the process clock and their SQL windows (avg7, avg30, the month before, the
median, the 90-day canonical pick, the cold-path slope evidence) from
NOW(), the database's clock. One response mixed two clocks, so a reading
on a window's edge fell in or out with whichever host answered, and with
the process clock pinned the analytics parity walk changed its avg7 with
the hour it ran at.
Each read now takes one `now` and binds it into its statements
(`request-clock-sql.ts`): `$N::timestamp`, which keeps the driver's
offset-free UTC string as written in any session zone. The canonical-pick
helper takes its window as an instant plus an interval instead of reading
NOW(); the series and calendar readers anchor it on their own `since`, the
BP-in-target read on its 365-day start. The window-mean and live-slope
columns the two aggregates each spelled out now come from one builder.
The streak's local day series read its bounds through `::timestamptz`,
which follows the session zone: outside a UTC session today's reading fell
out of the streak. It reads them as UTC first.
The parity walks for analytics, the dashboard summary and the
comprehensive overview pin a fixed day instead of today: their records are
laid out from the anchor day, so the chronotype's work and free nights and
the weight-by-weekday points moved with the weekday the walk ran on. The
snapshots are unchanged. A new suite pins the clock to four times of day,
far from the database clock, with readings ten minutes either side of
every edge, on both read paths, in a UTC and an Auckland session.
Since the lab staging reads a long report in parts, one auto-index could
make the transcription plus up to twenty reads of two calls each at the
person's response timeout, hours past pg-boss's fifteen-minute default,
which failed the job and retried it beside the reads still running. The
staging now takes the run's deadline as a schedule and plans only the
reads that fit before it. The index job is long, with an expiry above the
calls that run before the plan can say no, a stop on jobDeadline and a
lock per document; the document run hands the same staging its own
deadline, which moves the run's expiry as the reads are planned. The
document summary, one call and its reasoning repeat, gets an expiry above
that worst case.
ocrReadCallMs in the read core is the one per-call timeout both deadlines
divide by. The expiry guard now names every importer of the document and
lab read entry points with the queues that reach it.
The runtime guard's planted cases stripped every file's comments again on
each rescan, which took over a second alone and timed out under a full
parallel run; the stripped text is now memoised.
Refs #1274
A least-squares slope per day over readings minutes apart explodes: three
blood-pressure readings in one sitting, the only ones of the week, fitted
thousands of mmHg per day, and the dashboard tile printed seven days of it
as the 7-day trend. The 30-day fallback had the same flaw.
A window now earns a slope only when its readings fall on at least three
UTC days whose first and last lie at least half the window apart (4 days
for the 7-day window, 15 for the 30-day one). The rule lives once in
window-confidence.ts and gates every path that publishes a windowed trend:
trendSlope (summarize and its callers), the rollup composition and the
live REGR_SLOPE fallback of the summaries slice and the comprehensive
aggregator, and the per-bucket slope of folded week, month and year
buckets. Slope, direction and confidence are present or absent together.
The fit itself stays over individual readings, as the window means do.
The parity snapshots change only where the fixture falls short of the
rule: a weekly body-fat series has two days in its 7-day window, and a
ten-day oxygen-saturation series spans nine days of its 30-day window.
A status run (a metric card, the batched assessment, a derived score, a
workout line, a Coach memory pass) is one model call plus the one repeat
the Local and gateway clients send when a thinking model answered with
reasoning alone, each with the person's response timeout. The run's outer
bound was a single call, so the repeat was cut off, and the open card
stopped polling after 48 s, so a note from a model that needs minutes
never reached the page. The bound is now the per-call timeout times
STATUS_MODEL_CALLS, and the card's poll ceiling follows the published
response timeout (fast for the first 48 s, every 15 s after), only while
the route reports the note in flight; the fallback text stays meanwhile.
The status queue is exclusive, so the longer poll cannot queue a second
generation beside a running one.
The morning refresh is one comprehensive generation of up to five calls
in a row, with no outer bound and pg-boss's fifteen-minute expiry, so a
large setting had pg-boss fail and retry it while it was still running.
It now runs behind the same bound as the warms, and the job carries an
expiry derived from that worst case. The per-metric note, the workout line
and the Coach memory refresh get expiries derived the same way.
queue-runtime.ts gains a bounded class for these jobs: model time read off
model-time.ts, an expiry above it, carried on every send. A new guard
names every caller of runStatusCompletion and generateComprehensiveInsight
with the queues that reach it, and holds each such queue to bounded above
its runs or long under a run lock.
Refs #1274
Extracting lab values from a stored document still sent the whole text
in one call. A multi-page report produced more JSON than one call's
output ceiling, the answer was cut mid-fact, the parse failed twice, and
the read came back as 422 "Couldn't read the document. Try a clearer
copy." The lab scan had already learned to read the same report a page
or a part at a time, but each surface carried its own copy of the read,
so the document path never got it. It also ran under the shared 60 s
timeout, dropped every fact past 120 (or every fact when one did not
match the schema), and charged a flat reservation.
The read now lives once, in lib/documents/ocr-read.ts and ocr-plan.ts:
the per-read limits, the corrective retry, the salvage of a cut
answer's complete rows, the wire re-check and budget reservation before
each read, and the planning of a PDF a page at a time and text in parts
at line boundaries. The lab scan and the document extraction (the
extract route, its background run and the lab auto-staging) both read
through it, and the background run plans its reads on the same deadline
as a lab scan. The per-read budget moves to a shared metered budget.
The document path validates the answer fact by fact, keeps the lab
scan's row cap (400), and says why nothing came back: 422
documents.inbound.outputLimit when the reads stopped at the output
limit, 429 when the budget refused them, extractFailed only for answers
that were not the schema. A run whose reads were cut or not reached
reports `incomplete` on its result.
A guard holds the class: the output ceiling is set only in the core,
both surfaces read through runOcrReads, both runs share the deadline,
and a lab read and a document read go out under the same limits.
Refs #1257
The ChatGPT Codex backend and the OpenAI API serve gpt-6-sol, gpt-6-luna,
gpt-6-astra and gpt-6.1-sol, but the reasoning, Chat Completions and vision
gates only knew the gpt-5 line. A gpt-6 model got no reasoning control and
no real "off", went out on the legacy wire with max_tokens and a
temperature that reasoning models refuse, and was denied image input.
The three gates now read one classifier, openai-model-family.ts, so a new
generation is taught in one place. Every GPT from 5.1 on, the gpt-6 line
included, is the family that takes effort "none"; every GPT from 5 on reads
images. The upstream catalog (codex-rs/models-manager/models.json) lists
image input for every gpt-6 slug and the same reasoning presets as the
gpt-5.6 line; the presets omit "none" for gpt-5.5 too, which accepts it.
The Codex slug ladder moves into codex-slug-chain.ts and gains gpt-6-sol,
gpt-6.1-sol and gpt-6-luna. gpt-5.6-terra stays first as the slug verified
against the backend; the catalog names gpt-6-sol as its upgrade, and a
rotation retires a whole line at once, so the gpt-6 rungs follow before
terra's own siblings. The ladder test now holds every rung to a recognised
reasoning family with a real off and image input. The OpenAI presets offer
gpt-6.1-sol and gpt-6-luna.
The out-of-band warm the Insights page relies on bounded one
comprehensive generation at the account's response timeout plus 30 s.
A generation is up to five model calls in a row: the generation, the
JSON retry, the grounding retry, the outbound-screen repair, and the one
repeat a thinking model gets when it answered with reasoning alone. Each
call honoured the setting, but a local model needing 300 s per answer
under a 600 s setting met the 630 s bound in the middle of its first
corrective retry, and the briefing was dropped.
The bound is now the per-call timeout times COMPREHENSIVE_MODEL_CALLS
plus the headroom, for the nightly pass and the forced warm alike. A
test drives the real warm, generator, chain runner, Local client,
safeFetch and pinned dispatcher against a loopback server that answers
after 300 s; under the old bound it reports a timeout on the third
answer.
The response-timeout guard now names every outer bound around model
work and holds both briefing generators' stages to the call counts the
outer waits multiply by.
Refs #1274
The ChatGPT Codex backend now rejects every slug from gpt-5.5 down, so a
deploy without a CODEX_MODEL_FALLBACK_CHAIN override walked the whole
default ladder into "all configured Codex slugs were rejected". The
default now starts at gpt-5.6-terra, with its luna and sol siblings and
gpt-5.5 as the last rung.
The gpt-6 line stays out of the default: the reasoning, sampling and
vision gates do not recognise that family yet, and a test now holds every
default slug to a family the reasoning gate knows.
Postgres returns rows that tie on an ORDER BY in no particular order,
and not the same one twice. The medication reads ordered by columns two
rows share all the time: every read that hands schedules on included
them with no order at all, medications were listed by createdAt alone
(one import writes many in a millisecond), intakes by scheduledFor (a
twice-daily medication puts every morning dose on one instant), and
"the latest intake" lookups by takenAt. A client read the same account
in a different order from one request to the next, and a latest-intake
lookup picked either of two rows.
Every findMany / findFirst on a medication table, nested includes
included, now ends its order on the id in the direction of its last key,
and schedules are read by window, then id (MEDICATION_SCHEDULE_ORDER, the
order the backup already wrote). medication-read-order-guard.test.ts
holds the tree to it. medication-read-order.test.ts writes tying rows in
the opposite order to the one each read promises, through the medication
list and today's intakes, and fails without the tie-breakers.
The parity suite's own state reader had the same gap: it sorts rows by
content with the ids blanked, and rows that tie then kept whatever order
the unordered read returned, so their <id:N> labels swapped under a
loaded run. Its reads are ordered by id, which for a cuid is the order
the rows were written in.
The JSON import's schema left the route, so its entry in the inline
schema strictness register goes; the schema module says why its objects
are not strict.
A restore never runs beside a backup pass or another large job: the
heavy lane puts its delivery back in the queue for a minute at a time.
The job's own row says queued for the whole wait, and the console said
the restore starts "as soon as the background worker picks it up",
which read as a stuck worker while a backup ran.
The status routes now ask the queue. A delivery the lane put back sits
in created with a start_after ahead of now, and the active row of a
heavy queue is what holds the lane. The job view carries
waiting: "backup" | "busy" | null, and the console says it waits for
the running backup, or for another large job, and since when it has
been queued, in all nine locales. A queue that cannot be asked answers
no wait and is reported at warn.
The job view and its reads move to backup-restore-view.ts, which keeps
the worker module under its size baseline.
The JSON import read six measurement fields and six mood fields while
the full backup and the JSON export carry twice that. A sleep night came
back as stageless rows, a range reading as a single value, an ECG alert
without its verdict, and a mood entry without its note, its level-A
values, its zone, its tag and factor links or its day context.
The import now reads valueMin/valueMax, sleepStage,
rhythmClassification and deviceType on a reading, and note, a1 to a5,
tz, factors, structuredTags and context on a mood entry. Each mood entry
is written with its links and context in one transaction, the way the
bulk mood ingest writes one. A link to a tag the importing account does
not have is dropped and named in droppedTagKeys / droppedFactorKeys, and
the import card says which. The schema moves to
src/lib/import/json/schema.ts so the guards can read it.
json-import-field-coverage.test.ts runs the JSON export and both backup
variants on an account with every column filled in: every field they
write is either read by the import or named with the reason it is not
(row ids, source, deletion records, ciphertext, sync bookkeeping), and a
reason for a field no exporter writes fails as stale. The round trip
imports each file into a fresh account and compares the re-export field
for field. The published contract is held to the runtime schema.
The value chart's axes and labels, as drawn on a 90-day blood pressure
view on a tablet:
- Y ticks were whatever Recharts split the padded domain into, so the
labels read 57, 83, 109. The domain is now widened to a 1/2/5 step
and every tick is handed to the axis, evenly spaced and all drawn.
- X ticks were every n-th point, so their dates drifted with the data.
They now fall on a calendar grid (days, Mondays, two weeks, months,
quarters, years), each on the point nearest the grid line, and the
step is the smallest whose labels fit the measured plot width.
- The "your average" label sat inside the plot over the readings, and a
target zone's label was clipped to its first letter. Both now sit
beside the plot, in a right margin sized to the longest label.
A browser test draws the axes with Recharts and pins the tick values,
the Monday grid, and both labels outside the plot clear of every dot.
Each measurement token row carries a switch, the mint form an opt-in, and
the card a notice when readings exist that no token reaches. Copy in all
nine locales. The endpoint catalogue moves to its own file and the
revoked list out of the token card.
The custom-metric create and update refusals and the medication layout
refusal are filed through auditLog now, so a delegate's refused request
reaches the owner's activity feed, where it read as the generic change.
Each gets its own sentence in every bundle; the refusal sentences live in
their own function. The audit write guard states its floor, and the
schedule-reader guard names the medication create and edit, which answer
through medicationDetail.
Work a request starts and does not wait for runs inside the request's
context, so an annotate({ action }) from it landed on the request's event
whenever it ran before the event was emitted. The measurement tail's
status re-warm logged a hand-entered reading as
insights.status.invalidate.skipped on some requests and as
measurement.create on others.
runDetached starts such work so that its action is dropped for the event
it was detached from; meta, warnings and the auth it reads stay as they
were, and an event the work opens for itself keeps its own action.
fireAndForget now takes a function and starts it detached, and every
trackBackgroundTask handle on a request path is built with runDetached,
which a structural guard holds.
The backfill, the sleep repair, the OAuth-state cleanup and the sync
handlers move from lib/jobs into lib/fitbit/jobs, beside the queue facts
and definitions the registry already spreads. Their tests move with them.
Removing the integration is now the folder, its routes, the registry
spread and the drain's two cases.
The updatedAt backfill guard also walks provider jobs folders, so the
moved backfill stays covered.
Oura, Polar, WHOOP and Strava each keep the transport and endpoints in
`client.ts` and gain `wire-types.ts` (the record shapes) and `map.ts` (the
pure record-to-row mappers, no I/O). Google Health's 1,600-line client
becomes `budget.ts` (slice budget and outcome tally), `request.ts` (error
detail, retry, rate limits), `data-points.ts` (the paginated walker) and a
small `client.ts` for the profile and roll-up reads; its `export * from
"./mappers"` facade is gone and its importers name the module that defines
what they use.
The code moved verbatim; the mapper goldens written before the move and
the provider integration tests pass unchanged.
The section becomes a composer over backups/: the restore dialog with its
preview, the refusal, skip and schedule notices, the restore jobs and
their request, the upload flow, the run and download actions, and the
two row layouts sharing one action pair. The upload total is a pure
function and tested as one. The read-failure guard's 250-character
window had stopped reaching the restore preview's own failure lines; the
preview now sits in a smaller component, the guard sees it, and it is
listed as a deliberate exception with its reason.
Each family form moves into managed-record/ beside a small kit (the
coercion helpers, the form footer and the switch row); the section keeps
the record-scoped read, the save and the family switch.
`integrations/sync-kit/after-write.ts` is the tail every provider write
ends in: the arrivals go out, the touched days are refolded, the read
caches are marked stale back to the earliest touched reading, and the
touched status cards are re-warmed (through `afterMeasurementMutation`,
shared with the interactive routes); Withings also evicts the dashboard
snapshot. A full-history backfill refolds its whole touched range once
through `afterProviderBackfill`.
Withings, WHOOP, Polar, Nightscout, Google Health, Fitbit and Oura drop
their hand-written tails, and no provider module imports the comprehensive
insight generator any more; a guard holds the provider folders to that.
Differences from the copies: the status cards are re-warmed even when the
fold failed, and the caches are marked stale back to the earliest touched
reading instead of wholesale, so older closed blocks survive a sync.
Several provider unit tests mocked an export the old tails never reached
(their failures were caught and only logged); the mocks now name what the
code calls.
Marc-André Bombeck