The fixtures added with these fixes switched the type checker off with
`as never`, which the realism ratchet refuses to see grow. They now carry
their types, three webhook mocks needed no cast at all, and the lowered
counts, with the targets route's one fewer version stamp, are written
back.
The plan review rule told the Coach to read how a plan went from its
progress line, but a progress line exists only for an active plan, and
the review sets the plan to reviewed the morning its reminder surfaces.
For the plan being discussed there was nothing to read, so the model
could make a result up. The rule now reads a result only from a
progress line the snapshot carries, and otherwise asks how it went.
The notification categories' contract description, which lands in the
published API document, and a handful of new code comments and test
titles used em dashes; the house style keeps them out of new prose.
Three refusals answered without an error code, so the web printed the
status's generic sentence. A connection test with nothing configured
read "Some of the entries were not accepted", and a mistyped password
while adding a passkey read "Your session has ended. Sign in again",
which sent people to sign out for nothing.
An empty chain now resolves to no provider, which the test route
answers with `ai.provider.none`. The passkey registration and the
step-up mint answer a failed or stale proof with `auth.reproof.failed`.
`ai.provider.none`, `ai.budget.exceeded` and `auth.reproof.too_weak`
get sentences in every locale.
A person on the OpenAI-compatible gateway lost every AI surface once
their chain was saved at all: the fallback to the picked gateway asked
whether the chain was the default object itself, so a chain reset to
the default, reordered, or written by the gateway form's reasoning
setting (which appended a disabled entry to hold the value) no longer
counted, and the chain resolved empty. Rows written that way before
this release broke on deploy.
The fallback now reads the chain's content: the picked gateway serves
when nothing else resolves, unless the chain holds the gateway switched
off. A disabled gateway entry that carries a reasoning effort is the
form's holder, not a switch; the chain editor drops the effort of a
gateway it switches off, so a switch-off stays one. Saving the gateway
form with the gateway picked writes its entry switched on. Resolution,
the presence probe, the account payload and the iOS availability flag
all read the same rule.
The nightly plan review read a before/after window of the plan's metric
and stored an English verdict in `outcomeEncrypted` for every plan. No
screen and no route returned it; only the Coach snapshot read it, behind
the `COACH_EXPERIMENT_VERDICT` flag that is off by default. The worker now
only moves an active plan whose review fired to `reviewed`, which is what
takes it out of the standing group. The flag, its snapshot block, the
three eval cases for that block and the compose and env entries go with
it. The Coach prompt keeps its experiment rules and now ties the
read-back to the plan's progress line instead of the removed block.
Existing stored outcomes stay in backups untouched.
The plans page and the Coach reminders card printed the metric as a raw
key in capitals (WEIGHT, BLOOD_PRESSURE) in every language. Both now show
the measurement's name from the catalogue, through the plan metric
resolver moved to `plan-metric.ts`; a key HealthLog does not read shows
as written. The unused `coach.plans.proposalTitle` string goes.
The resolver tests queue their fixture rows through one helper instead of
an `as never` per row, the no-provider status case also proves the
single-provider resolver is never asked, and the size and comment
baselines take the lower numbers the change left behind.
Eight call sites (Coach turn, status texts, briefing, comprehensive
insights, medication extraction, about-me questions, both lab OCR paths,
the saved-config connection test) answered an empty chain by calling the
single-provider resolver, which ignores the chain's switches. A person who
kept only the local model on and lost its address got the ChatGPT sign-in
or the operator key instead, and every copy tagged that provider as the
operator key, so consent and cost attribution saw the wrong owner.
The rule now lives once in `resolveProviderChain`, and the presence probe
reads the same rule: a saved chain is the person's word, so an empty one
stays empty. Only a chain never saved falls back, to the OpenAI-compatible
gateway picked under Settings, AI, because the default chain has no entry
for it; that entry carries its real type. The callers drop their copies,
and the presence helpers that restated the old fallback go with them.
A person who switched both the Coach and AI analysis off under Settings,
Modules still got invited to set AI up: the Insights hint, its copy in
Settings, AI, the checklist row and the onboarding panel all keyed on
`provider.canConfigure`, which only knows the record kind and the
operator switch.
The flag itself stays as it is, because the AI usage card and the consent
request read it too and still serve document reading, lab photos and
medication extraction. The three invitation readers now also check the
module map through `hasOptedOutOfAi`.
The cycle read returns the projected period starts after the next one,
each with its band, under upcoming; the tool description for MCP clients
and the Coach did not say so.
The calendar read, the Coach and MCP cycle block and the coming calendar
feed now share one loader for the forecast's inputs (cycles, day logs,
overnight wrist temperatures) and one gate (`gatedPrediction`: no
forecast while prediction is off or raw-chart mode is on). The Coach
block reads the calendar's default day-log window, so its next period
is the date the calendar page shows, and resolves the luteal length
through the engine's own clamp.
`projectCycles` extends the engine's next-period forecast a few cycles
ahead. Cycle 1 is the engine's result unchanged; later starts are
anchored on it from the fractional median length, so six cycles out
carry no rounding drift; later bands widen as the SD of a sum of k
lengths with an estimated mean, with the engine's sparsity penalty, and
never narrow. The list stops where a band passes ten days or half a
cycle, holds the next period only while the engine is still learning or
for perimenopause, and is empty once the next period's band has passed
with nothing logged.
`/api/cycle/calendar` publishes it additively as `projection` and
`projectionState`; the Coach block carries it as `upcoming`.
A read whose readings all lie past the lookback limit was reported on its
step as outside_window, and its fetch entry carried no reason, so a client
could not tell an empty window from data the limit keeps out. The step now
carries reason outside_reach and the fetch entry mirrors it. The method
line and the web trail keep the window sentence; the answer names the
limit as before.
Four instruction shapes passed the outbound dose screen in every language:
a negated intake ("Don't take atorvastatin tonight", "Nehmen Sie kein
Atorvastatin"), a pause ("Take a break from enalapril"), an increment with
no target ("Add 4 more units of insulin"), and a measured volume of a liquid
drug form ("Take 10 ml of the syrup").
Each bank gains the shapes, anchored on a medication noun, a drug unit or a
liquid drug form rather than on the verb. A negated intake counts as a hold
only when the clause closes or a time word follows and no protective
qualifier sits in the object ("a double dose", "more than", "half"), so the
standard warnings still pass. `ml` stays out of the drug units: the liquid
shape needs a syrup, solution, drops or medication noun beside it, so
"drink 500 ml of water" passes.
The Russian and Korean banks now know the generic-name stems in their own
script (INN_STEM_CYRILLIC, INN_STEM_HANGUL), and the English object slot no
longer crosses another language's negation, which had blocked the Dutch
warning "Stop niet met ramipril zonder overleg".
End a dose unit at a Cyrillic letter, so «3 месяца» no longer reads as
"3 МЕ". Keep ml out of the drug units as the other banks do, and stand
the target-dose shapes down on a dietary noun (water, sodium, caffeine),
which blocked plain nutrition advice. Cover the stop and inject verbs
(перестаньте, бросить, колите) the English bank already blocks.
The labs block and a single-reading read now build a reading from the same
select and the same shaping, so one reading fetched by its id renders exactly
like the latest reading of a marker.
- baselines: take main's entries, keep the two ru-driven line counts (coach-prose-grounding 1091, insight-generator 1115) and regenerate the function baseline.
- messages/ru.json: translate the five unit-less markers keys the release added (labs.unit.none/noneHint, labs.ocr.errorUnitMismatchNoUnit, labs.ocr.unitDiffersBadgeNoUnit, recordImport.unitlessHint) — parity is 11 159 keys again.
Dose bank
- DRUG_UNIT learns мг / мкг / мл / ЕД / МЕ, so "2,4 мг" anchors a dose again.
- The bank matches verb-first Russian word order too ("принимайте 2 таблетки",
"пропустите дозу метформина", "начните с 5 мг"), and the Cyrillic drug names
people actually see on the box; the deferral words ("спросите врача",
"как назначено") stand those shapes down.
Cyrillic-safe regexes
- \w and \b are ASCII in JavaScript: every stem in the ru bank and in
population-norms/ru.ts is [а-яё]*, and the word edges are lookarounds
(?<![а-яёa-z]), so «высокий риск» and «нормальный диапазон» match again.
- coach-prose-grounding: Russian month names in the genitive, a space as the
thousands separator in parseLocaleNumber and numFrag, and Cyrillic-safe
edges around the written-date pattern.
Registered everywhere a locale has to be
- targetLanguageName() returns "Russian", so the coach, the period narrative
and the document descriptions stop saying "write in English".
- i18n-key-tool.mjs, split-description.mjs, LOINC name normalisation (it
stripped Cyrillic, so no Russian lab name resolved), HOST_TLDS (ru/by/kz/ua),
no-key-fallbacks (32 tables), country-label, the AUDIT-C standard drink.
- docs/api/openapi.yaml regenerated.
Questionnaires
- The PHQ-9 / GAD-7 source comments no longer claim no validated Russian
version exists (Zolotareva 2023; Pfizer's official screeners).
- WHO-5 items are gender-neutral: «чувствовал(а)».
- "body fat" and "fat mass" both read «Жировая масса»; body fat is now
«Процент жировой массы», which is what the two metrics measure.
Tests
- ru in the 13 locale tables, plus mood, onboarding, instruments, clinician
view, status summary, describe screen, citation coverage and LOINC.
- i18n-no-dot-separator: the ru bundle is exempt from the dash rule (a spaced
dash is Russian punctuation, not a fragment join); the middle dot still
fails there.
- Two 800-line baselines raised by the 12 and 5 lines this change adds.
pnpm typecheck, pnpm openapi:check and the full suite (36 316 tests) pass.
Every place a lab unit is printed beside a value now goes through one
helper, so a reading stored in "1" reads "3.6" and not "3.6 1": the
list, history and detail, the chart and its tooltips, the reference bar
and range notes, the trend card, the timeline, the day view, the
clinician page, the doctor-report PDF and the changes card. Where the
unit stands alone (a marker's subtitle, a picker row, the merge preview)
it reads "No unit".
The Coach's lab block, the get_labs history and the visit-prep summary
hand the model an empty unit for such a value, so it never quotes the
"1" either.
Refs #1313
get_labs readings (the latest, earlier values in full, trajectory readings)
carry the fetch ids of what they link to, with linksTruncated; each linked
record is described once per answer in linkedRecords, because one report
usually holds a whole panel and would otherwise repeat on every reading.
fetch lab: and fetch med: list their links like the other record kinds, and
get_medication_schedule rows carry their id and links inline. get_visits and
get_illness_recovery rows carry visit: and condition: ids, so a row's links
are one fetch away. The labs and illness blocks gain an opt-in to carry row
ids; the Coach never asks for them.
Full Russian catalog (11,154 keys, 1:1 with en, identical nesting) and every
locale wiring point the app declares, so ru is a first-class locale instead
of an English fallback:
- messages/ru.json, project.inlang/settings.json, i18n boot script
- i18n config / load-locale, format-locale (ru-RU), date-fns ru
- labs: parse-reference-range RANGE_PROSE for Russian lab sheets
("3,5 - 5,0", "от 3,5 до 5,0", "до 5,6", "не более 5,6")
- AI: citation-coverage normative stems, coach system prompt and insights
reply footer (Russian prompts must override the hard-coded
"Respond in English"), population-norm patterns
- safety: outbound-screen Russian imperative/dose/unit/causality bank
- mental health: PHQ-9, GAD-7, WHO-5 and SCI wording
verify-message-catalogs.mjs passes for all 10 locales.
A call that ran beside the one that learned the lesson read today's
selection instead and failed with the 400. The matcher now also wants the
refusal to name max_tokens, so a refusal of another field that echoes the
request does not switch the token field.
Refs #1304
An OpenAI-compatible gateway entry pointed at an Azure OpenAI v1 endpoint
was treated as a generic gateway and sent max_tokens and temperature, which
a reasoning deployment refuses with a 400. An Azure v1 endpoint now takes
the api.openai.com contract for a reasoning model, and any gateway that
refuses max_tokens for a model is asked again with max_completion_tokens
and remembers it, which covers deployments named something of their own.
Refs #1304
The dialect cache gave a day to every value that says the reasoning
field is refused, and that includes the combined "field refused, reasons
regardless" value, which one answerless reply writes. One reply could
therefore stretch a lesson to a full fresh day, against the rule that a
lesson read from a single reply keeps the hour.
A lesson now says what taught it. A 400 refusal gives a refused field
its day. A reply that upgrades a refused field to the combined value
keeps the refusal's remaining lifetime rather than a new day or the
hour, so the refused request does not come back early either.
Every write to the per-sign-in Codex memory checked the account digest,
but the reads did not. A credential reconnected to another ChatGPT
account therefore tried the old account's working model first, sorted
by its refusals and trusted its model list, until the first write
rebound the entry, and the vision gate kept answering from the old
account in the meantime.
The client reads (working model, refusals, model list) now take the
credential and treat an entry learned under another account as empty.
The vision gate holds no account id, so the connect routes rebind the
entry and the disconnect routes drop it, and the gate reads the entry
as its process last bound it.
A fixed preference list goes stale at every retirement: when GPT-5.5
retired for ChatGPT sign-in on 2026-10-14, every model the shipped list
named was refused and Codex answered 503 until a release changed it. The
working model also sat in one slot for the whole process, so two sign-ins
on different plans overwrote each other's model and the vision gate read
whichever had answered last.
The client now reads the models the account is offered the way the Codex
CLI does (GET /backend-api/codex/models?client_version=..., same sign-in
as the responses call; shape from codex-rs/protocol/src/openai_models.rs)
and walks: CODEX_MODEL first; our preference order kept to what is
offered; the other offered models in the list's priority order, leaving
out hidden, specialist and frontier-tier entries; the rest of our order
last, so the walk never holds fewer models than before. With
CODEX_MODEL_FALLBACK_CHAIN set the order is the operator's and the list is
not read.
The list is a hint, never a gate. The read goes through safeFetch with a
5 s bound and a 4 MiB cap, is shared by concurrent requests, cached per
sign-in for an hour, and a failed read waits ten minutes. Any failure
leaves the static walk, one ai.codex.models_unreadable warning per sign-in
reports it, and each request that builds the walk annotates a pinned
codex.models.* shape.
The working model, refused models and the list are kept per stored
credential (a person's connection, or the operator-shared central one),
bounded by TTL and a 1000-entry LRU, with a digest of the account id so a
credential reconnected to another account starts empty. A model the list
offered that then answers model_not_found goes to the back of that
sign-in's line for six hours. When the list names a model's input
modalities, the Codex vision gate reads them; otherwise the model family
decides, as before.
Refs #1287
The reasoning-field planning and the refusal learning are a self-contained
part of the Codex wire. Moving them out of codex-client.ts leaves room for
the per-sign-in model selection that follows without growing a file the
size ratchet already holds at its limit.
Refs #1287
A gateway that refuses the reasoning field does so by its configuration,
so the lesson now holds for 24 hours instead of one. At the hour every
expiry cost one refused request, a few a day on a gateway that runs the
briefings unattended. Lessons read from a single reply keep the hour.
Refs #1279
The createdAt field and the locale import grew files the size ratchet
holds. The account payload's insurance-number decryption moves into
its own helper, the import joins the zod one, and two comments lose a
wrapped line, so no baseline has to rise.
The Coach chat, the document chat and the fenced chat declared their
locale field as de/en, so a body naming French, Korean, Brazilian
Portuguese or any other shipped language was refused with 422 before
resolveServerLocale, which carries every locale, ever saw it. The same
pair sat on the profile update, so an account could not store any other
language there either, and the stored language is what the cookie-blind
background work (the Coach nudge, notification texts) writes in.
All four now take z.enum(locales), and the published contract for the
metric, insight and biomarker status queries and the profile update
lists every UI locale instead of two.
OpenAI retired the 5.x line from Codex for ChatGPT sign-in in steps: GPT-5.2
and GPT-5.3-Codex on 2026-05-26, GPT-5.4 and GPT-5.4 mini on 2026-08-31,
and GPT-5.5 on 2026-10-14. The backend answers a retired model with 400
model_not_found and "not supported when using Codex with a ChatGPT
account".
Pin that exact refusal as a reason to try the next slug, pin that an
account refused the head of the default list walks on to the next rung
and caches it, and keep the long-retired models off the default list.
gpt-6.1-sol, a rung on the default list, answers reasoning "off"
(effort none) with 400 unsupported_value and lists low as its lowest
effort (seen 2026-10-24). The client must stay on the slug, retry with
low and remember it, not walk on as if the model were retired.
Background jobs ask for medium reasoning at most, and status cards ask for
none. A model that cannot be switched down (GLM behind a vLLM gateway, for
instance) ignores both, spends the whole room thinking and returns only
reasoning, so the daily briefing and the status cards fail as
reasoning_only and the chain pauses. The Coach on High works because it
leaves 12k tokens to think in.
The chain runner now marks every call of a job walk as unattended, in one
place rather than at each call site. The gateway and local clients give
such a call at least the High thinking room once its model has answered
with reasoning alone. The room has to ride on the #1267 repeat with none:
one call may send two requests, and a once-a-day job meets an expired
lesson every morning, so a room that waited for the unswitchable lesson
would never reach the first briefing of the day. It is a ceiling, not a
charge: a model that does stop at none never touches it. Calls a person is
reading, the Coach included, keep the room of their own level, and the
requested effort is unchanged.
The room calculation both clients shared is now one function.
Reported by @Tio61 in #1276.
The capability gate's check for an account pending deletion added a
point read of users to every request that resolves AI capabilities, one
statement over the budget on /api/auth/me, the dashboard snapshot and the
batch ingest. The usage-level read already loads the record's row once
per request; it now carries the deletion lock too.
The translation files merge key by key (no key changed on both sides), the
backups list carries both the database dump and the pre-wipe copy window,
and the backup-restore doc keeps the shorter rewrite plus the dump in the
bucket and a lifecycle rule that covers database/.
The MCP server instructions promise that the person's free text comes back
wrapped in USER_TEXT markers, but get_medication_schedule returned
medication names and doses as plain strings, and get_labs returned analyte,
panel and unit names that can come straight out of an uploaded document.
Text written to read like an instruction reached the connecting client's
model unmarked. The marker scrub also removed only the exact uppercase
spelling, so a lower-case or spaced marker survived inside fenced text.
Both tools now fence those fields, and the scrub removes a marker in any
case and with any spacing inside the brackets.
The Coach snapshot attached its labs block regardless of the module, a
decision from before Labs could be switched off, while the snapshot's own
prelude says a disabled domain never reaches the model. With Labs off, the
Coach prompt, its get_labs tool, the daily briefing's flagged-marker block,
the biomarker status note, the reaction line to a new panel and the MCP
lab_trend_brief prompt all still read lab values for a model.
Each of those reads now asks the labs module first and reads nothing when it
is off: the Coach tool answers module_disabled, the briefing omits the
block, the status note neither writes nor serves text, the reaction line
gets no value, and the MCP prompt refuses like get_labs and the doctor-visit
prompt already did. The data routes stay as they are. A structural guard
holds every lab-table reader under the model-input paths to the gate or to
a listed reason its rows never reach a model.
Deleting an account, a guardian deleting a managed profile, and "Delete
all data" were final the moment they were confirmed: the account row
cascaded into every table and every stored copy, and the off-host objects
were purged in the same breath. Nothing short of rolling the whole
instance back could return one account.
ACCOUNT_DELETION_GRACE_DAYS (14 by default, at most 30, 0 erases at once
as before) now sets a window:
- Account and managed-profile deletion lock the account (migration 0488:
deletion_requested_at, deletion_purge_at, deletion_requested_by_id).
Every session, refresh token, API token, connector, share link and
step-up elevation is revoked in the same transaction, and the record,
its stored copies and its off-host objects are left as they are.
- Sign-in is refused with 403 auth.account.deletion_pending and the
erasure date, after the credential was proved and before a second
factor is asked for. The session and Bearer resolvers and the acting-
account resolver refuse a pending account too, so a credential minted
in the moment before the lock is useless.
- POST /api/admin/users/{id}/undo-deletion (cookie-only, audit-logged)
clears the lock. The admin users list reports the pending state.
- The hourly account-deletion-purge job erases accounts whose window has
ended, exactly as the immediate deletion did, with the off-host purge
request committed alongside. The Guardian invariant is decided again;
a refused erasure stays pending and is reported.
- A wipe first stores the whole record as a PRE_WIPE_<ms> copy through
the writer the weekly copy uses, keeps it out of the wipe, and holds
its off-host purge back with a new not_before on the request. The
ordinary restore brings the record back; restoring that copy withdraws
the held purge. The purge job deletes the copy when its window ends.
Background work skips accounts pending deletion through one shared
predicate (src/lib/account-deletion/active-account.ts): every reminder,
digest, Coach, sync, backfill, rollup, score and backup enumeration
filters with it, and account-deletion-enumeration-guard.test.ts fails on
a query in background work that neither filters nor says why it need
not. Notification delivery and the AI capability loader refuse a pending
account as chokepoints behind it, and the Withings, WHOOP and Telegram
webhooks and the notification quick actions ignore one.
The privacy summary publishes the window. Two pieces of the reminder tick
and the off-host pass moved into modules of their own to keep the files
inside their size ratchet.
The labs block reads the whole record, or up to the lookback limit, not just
recent readings, and a named read also comes back empty when no marker
matches. The header of the labs snapshot and the note at its call site
still described the earlier twelve-month window.
The turn records up to 64 steps while the read path kept only the first
48, so the tail of a long answer's step list vanished on reload. Both ends
now read one MAX_TURN_STEPS constant, and a unit plus an integration test
round-trip a full 64-step answer through the stored provenance.
A status run is one completion, so its ceiling was the completion ceiling
under a second name. The Local client also re-exported a JSON-mode helper
nobody imported from it.
A gateway or local entry can carry its own reasoning setting (#1126), sent
on every call that has no level of its own. An endpoint that refuses
reasoning_effort answered each of those calls with a 400 that was never
learned, because only the call's own level counted as sent: every status
note, every summary and every other level-less call failed, for as long as
the setting stood.
The refusal is now learned whichever level put the field on the wire, by
the same path as any refused field: the call is asked once more without it
inside its two-request budget, later calls send no field for that endpoint
and model, a model seen reasoning unasked keeps its thinking room, and one
ai.provider.reasoning_field_refused line says so per lesson instead of a
failure per call.
Refs #1279
The health ledger records failures per account and provider, and since
background work leaves out a benched provider, an operator who fixed the
instance-wide provider in the admin AI settings still saw every account's
background work sit out the pause the broken setup had earned, up to a day.
Only a person's own provider settings and the Codex sign-ins forgot the
record.
A successful save of the admin AI settings now forgets what the ledger
recorded for admin-openai on every account and closes the status notes'
negative windows, the same reset a person's own save makes for theirs
(clearStatusNegativeWindows without an account covers every record). The
ledger comments that still said a benched row only deprioritises now say
what it does for background work.
Refs #1279
Marc-André Bombeck