On a host whose CPU cannot run the page renderer (x64 without AVX2, arm64
musl without LSE/RCpc), every AI read of a PDF for a provider that only
takes images failed with "Scanning PDFs needs a Claude vision provider",
which says nothing about the cause. Most medical PDFs are produced
digitally and carry every word in their text layer, which pdf.js reads
without rendering.
`prepareVisionInput` now answers a text read (`mode: "text"`) when the
renderer is unavailable on this host, the provider cannot take the PDF
whole and the PDF has a usable text layer. Summary, suggest and the
background summary send that text through their existing text mode;
"Read with AI" and the index backfill write it as the local text index it
is, with no provider call; the automatic index leaves it to its local path.
`planVisionReads` does the same for the extraction and the lab scan: text
reads by page group, with anchored provenance, and the pages without a text
layer reported as `extractionIncomplete` on a document and as `unreadPages`
on a lab scan. A supplement label PDF is read from its text layer too.
A scan on such a host now answers `documents.inbound.pdfRendererUnavailable`
and `labs.ocr.pdfRendererUnavailable` (422): this server's processor cannot
run the page renderer, so a scanned PDF can only be read by a provider that
takes PDF files directly, or as photos. The `pdfNeedsAnthropic` codes are no
longer emitted, since with a working renderer every provider reads a PDF;
the recorded index outcome keeps its `pdf-needs-anthropic` value with
corrected copy.
The extraction path no longer folds a render failure of one document into
the capability error: like `prepareVisionInput` since #776 it reports
"The PDF pages couldn't be rendered for scanning." (`extractFailed`), and
the lab scan answers `labs.ocr.extractFailed`.
Hosts that render, and providers that take PDFs whole, are unchanged.
Refs #1311
A re-check or reservation that failed itself stopped the scan as
"unavailable", and the review screen said AI had been turned off. It now
stops as stoppedBy error with its own message in every locale, and the
failure is reported as one: an error signal and the worker error
reporter. A scan that failed that way before any row answers 500
aiRuns.failed.
The overspend booked before a further read is now net of the cached
share, as the settlement books it. A reservation that names no known
owner stops the scan instead of taking the user plan's larger cap.
A page's text layer is cut at the last whole line; only a page whose
first line alone runs past the cap keeps part of it, cut at a space or
hard, rather than collapsing to the bare cut mark, and an endless line is
no longer read past the cap. The unused per-call text limit is gone. The
run result's description names every outcome a stopped scan can have.
The text layer went to the model as the authority for digits. It is not
one: a scanned PDF often carries an invisible OCR layer with its own
mistakes, a subset font can map to the wrong characters, and a crafted
PDF can carry invisible text that differs from what the reviewer sees.
The prompt now calls it a hint only, makes the image authoritative, forbids
adding a value the image does not show, and asks for a low value
confidence where the two disagree. The layer is fenced with the document
text markers, which it cannot close, and so is text mode's OCR'd text.
Each page's layer is capped while it is collected, and a read shares its
text budget out per page, cut at a line and marked, so a long first page
no longer pushes the later ones out.
A further read now first books what the reads so far cost beyond their
reservation, then its own ceiling, both against the cap of the owner the
route judged the first read by, so the counter the next read is judged
against never lags. The wire is re-checked before every read, the first
included. A re-check or reservation that itself fails stops the scan with
the pages read rather than failing it, and a scan stopped before any row
answers with the stop's reason (budget, or the wire's own refusal) instead
of an unreadable page.
The route reserved one read's budget and checked the wire once, and a
20-page report then ran up to 40 calls for many minutes on that single
admission. Every read after the first now reserves its own share of the
day's budget and asks the wire again (aiEgressRefusal) before it is sent;
when either refuses, the scan stops there, keeps the pages already read
and names the rest with stoppedBy budget or unavailable. The run settles
against everything it reserved, with the provider's reported usage net of
its cached share, and a scan that fails after several calls is charged
for them instead of refunded in full.
A rendered PDF page now goes out with its own text layer, read by pdfjs
beside the image, so a digitally produced report's digits, decimal
separators and units come from its characters rather than from a 1600 px
picture. Splitting the PDF into single-page documents would need a PDF
writer the project does not ship; the text layer needs none and helps
every provider. A scanned page has no text layer and reads as before.
Codex's cut reply keeps the text it streamed (on a non-enumerable
property, so no log or event serialises it), and its closed rows are
salvaged like every other provider's. The merge folds a reading only into
one an earlier page listed, never two rows of the same page, and the
printed range is part of the key. OCR'd text the deadline cannot split
small enough for one read is refused up front as labs.ocr.tooLarge.
A scan sent the whole report to one call under an 8000-token ceiling and
validated the reply against a 100-row envelope. A long multi-panel report
lost everything either way: the reply stopped mid-row and failed to parse,
which the person saw as "try a clearer photo", or it held more than 100
rows and the schema refused all of them. One malformed row refused the lot
as well. Reasoning models made it worse, because the call sent no
reasoning setting and OpenAI, Codex and Claude's 5 generation reason by
default against the same output allowance.
A PDF of more than one page is now rendered and read a page per call on
every provider (up to 20 pages), as many reads as the run's deadline
allows, and the pages are merged in order with a reading printed twice
listed once. A photo, a one-page PDF on a provider that reads PDFs, and
OCR'd text up to 4000 characters stay one read; longer text is split at
line boundaries. Each read carries an explicit reasoning off, so a model
that can stop reasoning does and one that cannot gets its minimum budget
on top of the answer's. A reply cut at the output limit is detected from
the finish reason (Codex's incomplete stream included), keeps the rows
that closed before the cut and is named in the result instead of retried;
when no read produced a row the run fails as labs.ocr.outputLimit. A
provider failure after the first page keeps the pages already read.
Rows are validated one by one, the row cap is 400 for the scan and the
commit, and a report past it keeps its first 400 and says how many it
had. The run settles what the provider reported it spent rather than the
fixed reservation, so a page-by-page read still counts against the day's
budget in full.
POST /api/biomarkers/{id}/merge folds markers into one in a transaction,
converting readings where the registry knows the units and refusing the
whole merge where it does not; a dry run answers the same result and
writes nothing. Aliases appear on the marker and are edited through the
marker's PUT. Scan and document review rows name the matched marker or
offer up to three, and a chosen marker is remembered as an alias. The
doctor report, FHIR and MCP group readings by marker. Aliases and the
merge ledger travel through backup and restore.
Every new endpoint of the contract exists, wrapped in apiHandler, with its
authentication, module gate and request validation, and answers
501 not_implemented until its behaviour lands. The OpenAPI document carries
the new operations and schemas and the additive response fields on the
changed routes. The four server feature flags are published as false.
contract-stub-inventory.test.ts registers the stubbed routes; each one
leaves the register with the change that fills it. The route, sharing,
session, security, module-gate and outcome inventories name the new routes
deliberately, and the error-code catalogue lists the codes the routes emit
today.
The matcher stored the registry entry's key, so LDL and HDL markers were
recorded as total-cholesterol. The key now names what the marker is (the
catalog slug, or a registry-only key for an analyte with no seed), and the
registry resolves ldl and hdl to the cholesterol entry for the conversion as
before. A key already written in the old form is rewritten by the same fill
step that brings a stale key in line with its name, the first time a reading
needs it. The OpenAPI descriptions say what the key is and that it follows
the marker's name and unit.
A converted reading's bounds move with its value, but the printed range text
stays in the unit the report used, and every reader preferred that text: the
list showed "Report range 3.9 to 5.5" beside 100.9 mg/dL, and the doctor
report and the FHIR referenceRange.text did the same. The text is still
stored verbatim. The range resolver now takes the reading's sourceUnit (as a
required field, so no reader can forget to select it) and, for a converted
reading, leaves the window's text empty so every surface states the converted
bounds in the marker's unit, and hands the printed text over separately with
its unit. The lab DTO carries it as printedReferenceText, and the note beside
a reading shows it as "as printed in mmol/L: ...".
Slice 3 of phase 3 of #1095. The scan review no longer treats every reading in
another unit as unwritable. The server tells it which analyte the matched marker
is (markerAnalyteKey on each extracted row); a row whose unit differs but can be
converted starts selected and carries a badge saying what Save will store and
what the report stated ("Saved as 100.9 mg/dL, converted from 5.6 mmol/L"). A
row that cannot be converted keeps the existing blocked badge and error. Save
still sends what the report printed; the server converts, as in slice 2.
The review's duplicate flag also recognises a re-scan of a reading that was
stored converted, by what the report printed, the same rule the commit applies.
A converted reading carries a quiet "converted from 5.6 mmol/L" on the lab list
(row and tile) and the history list (phone and table), so a number nobody typed
can always be traced back to what the report printed.
Strings in all seven locales.
Refs #1095
Slice 2 of phase 3 of #1095. A reading that stated a unit other than its
marker's was refused. When the marker's analyte is known and the registry has the
pair, the three write paths now convert it instead: POST /api/labs, the OCR
commit and the document commit.
The converted value is stored at full precision in the marker's unit, with what
the report printed in the existing source_value and source_unit columns. The
report's own reference window is in the printed unit, so it is converted with the
value; otherwise the verdict would compare a converted value with an unconverted
window. A pair the registry does not have is still refused as before (the 422
now says why), as is a marker whose analyte is unknown.
The marker's key is filled the first time a reading needs it. The same scan
committed again is still a duplicate: a converted row holds the number in the
marker's unit, so the duplicate check also matches on what the report printed.
Editing the value of a converted reading clears the conversion record, because
the number is then the person's own.
The lab result DTO and the OpenAPI schemas (labs and the OCR commit) carry
converted, sourceValue and sourceUnit.
Refs #1095
"Read with AI" and the lab scan used to run their model calls inside the
request. At the default response time a read of three calls outlasted the
100 s a proxy in front of the server commonly allows, and the person saw a
cut connection for a read the server was still doing.
Both now run as a background run (`DocumentAiRun`, migration 0366):
- `POST /api/documents/inbound/{id}/index` with `Prefer: respond-async`
answers 202 with a run id once every quick refusal has passed. Without
the header it answers exactly as before, which the iPhone app relies on.
A second request while a read of the same document is queued or running
attaches to that run.
- `POST /api/labs/ocr/extract` always answers 202. Only the web client
calls it.
- `GET /api/ai-runs/{id}` serves the run: the same body the synchronous
route answers with, or the failure code and status it would have given.
Scoped to the caller; another account's run is a 404.
The worker (`document-ai-run`, a long queue with a one-hour expiry, one
job per run under `lockedPass`) claims the run, asks the capability again
for the record and re-checks the wire for the provider it picks, sets the
run's deadline from the person's AI response time and refuses a read that
cannot fit the job's expiry. Sync route and worker run one body per kind,
so the two paths cannot drift. Budget is reserved at enqueue and settled
once by whoever ends the run.
Input and result are sealed with labelled AES-256-GCM, registered for key
rotation; the input is dropped when the run ends. A five-minute reaper
fails a run no worker took within fifteen minutes (and hands its
reservation back), fails a run past its deadline, and deletes a finished
run after an hour. Runs are outside backups and inside the data wipe.
The web client polls through the query-key factory and says, calmly, that
the read continues in the background, and when no worker has picked it up.
A numeric lab reading was stored with its value as received and the marker's
unit stamped on it, so 5.6 mmol/L against a marker kept in mg/dL became
5.6 mg/dL and read as far below range. Both the manual/API write and the
scan commit did this; nothing compared the two units.
- POST /api/labs returns 422 labs.unit.mismatch, with the marker's unit and
the reading's unit in meta, and writes nothing. The unit is compared only
when one was sent: on the biomarkerId path it may be omitted, and the
marker's unit is then the stated one. A qualitative reading has no unit.
- The scan commit skips such a row with reason unit_mismatch, so the outcome
reads partial or failed instead of a clean save, and the toast names the
analytes concerned rather than saying "duplicates".
- The extract response carries the matched marker's unit. The review screen
shows a calm badge on a row in another unit, starts it unticked like a
duplicate, and treats it as a blocking error on the unit field if it is
ticked anyway.
- A small lab-specific normaliser forgives spelling only: spacing, the micro
prefix, the case of the litre, and a unit printed in capitals for a short
list where capitals cannot mean another unit. Case is otherwise significant,
so mIU/L and MIU/L, or g/L and G/L, are never merged.
The document-commit path already refuses a mismatch and is left as it is.
Readings stored wrongly before this cannot be identified, because the
original unit was never kept; they are not touched.
Adds EN and DE strings; the other locales are left to the maintainers. The
OpenAPI spec and the error-code catalogue carry the new code.
Fixes#1094
A PDF longer than the pages a scan renders was read from its first ten
pages, and the cut was recorded only in the wide event. The person checking
the extracted rows had no way to know that values on later pages were never
read.
The rasterizer now returns the PDF's page count, the extract response
carries pageCoverage { read, total } when the cap applied, and the review
screen says how many of the pages were read.
The document, labs-scan, medication-extraction, fenced-chat, attachment,
connection-test and auto-read operations document the capability they
answer under, the refusal envelope and codes, and the consent rule. The
two capability probes publish the capability beside the provider answer.
A self-hoster scanning a stack of lab reports ran into a fixed six scans
an hour with no way to raise it. LABS_OCR_LIMIT_PER_HOUR now sets the
per-user ceiling (default 6, clamped to 1-1000, falling back to the
default when unset or not a number), the same shape as
DOCUMENT_AI_LIMIT_PER_HOUR. It is on the compose whitelist, in the
production example, the env manifest and the AI provider guide.
The route also charged the slot before it had looked at the request, so
a malformed body, an unsupported or oversized file, a PDF that could not
be rendered or a scan the daily budget refused still used up the hour.
Those failures hand the slot back now; a scan that reached the provider
keeps it. The 429 carries meta.retryAt. Refs #985
`.meta()` clones the schema in Zod 4 and registers the clone, so a bare
`schema.meta({ id, description })` statement registers nothing at all.
Forty-one of them sat across nine route modules. The ids they named —
`MeasurementType`, `MeasurementSource`, `LoginPasswordRequest`,
`CreateMeasurementRequest`, `CreateBatchWorkoutRequest` and thirty more —
reached `components.schemas` zero times, the descriptions written for them
were published nowhere, and every use site inlined an anonymous copy. A client
generated from the contract got duplicated nameless types where one named
model belongs, which is the part that matters, because the iOS client is why
the contract is published.
Forty of them now bind their clone, and every use site points at it. Where the
schema comes from `src/lib/validations/*` the import is aliased and the clone
takes the name the use sites already spell, so nothing mutates the runtime
validation schema. The seven that cross module boundaries are exported from
`./shared` and the eight consumers import them from there; importing the raw
schema from the validation module instead is exactly what inlined them.
Thirty-four named components appear where inline copies stood. Fully
dereferenced, the document differs by eighty-six added descriptions and
nothing else: no operation's request or response shape moves. `CoachPrefs`
was already force-registered in the index and finally carries its description.
Five ids legitimately stay out of `components.schemas` —
`ListMeasurementsQuery`, `ListAllergiesQuery`, `ListEncountersQuery`,
`SuggestEncounterQuery` and `ListPractitionersQuery`. The emitter expands a
`requestParams.query` object into inline `parameters`, so a query id has
nowhere to land; they stay in the assigning form for consistency.
Three findings the wiring turned up:
`CreateShareLinkRequest` and `HealthRecordExportRequest` were annotated in the
Coach module, which imported the two schemas for no other reason and never
referenced them. Both operations that send these bodies live in the
health-record module, so the annotations move there. Both descriptions
described a `sections` field: the export body has no such field and rejects it
with a 422, and the share-link body freezes scope through `selection` plus
`documentIds` / `documentOnly`. Publishing either sentence verbatim would have
published a lie, so both are corrected to the shape the routes actually take.
`CoachPlansListQuery` — the forty-first — named a component that nothing
referenced and that a query object could never reach. The operation publishes
its two parameters by hand, because the schema carries a cross-field
refinement the expansion cannot express, and every clause of the annotation is
already published on the operation and on both parameters. It is deleted
rather than left registering nothing.
`withIdempotency` has made writes safe to retry since v1.5.x. It was described
in route comments and in no schema, so an app generated against
`docs/api/openapi.yaml` could not learn that a retry was safe. The iOS side
asked for it; the mechanism does not change, only what it says about itself.
34 operations now declare `Idempotency-Key` and the `X-Idempotent-Replay`
header, plus the 409 that refuses a duplicate still in flight. The description
leads with the one silent failure in the mechanism: a key that misses
`^[A-Za-z0-9_\-:.]{8,128}$` is IGNORED rather than refused, so the write
proceeds unprotected and looks exactly like a request that carried no key.
## What the work turned up
Three things the contract had wrong or missing, none of them the thing being
built:
- `/api/insights/chat` describes `withIdempotency()` in its header comment and
applies it nowhere. The Coach chat is not retry-safe. Its prose says it is.
- `/api/measurements` exports GET, POST, PUT and DELETE and wraps only POST.
Publishing the header per PATH would have promised a safe retry on two
methods that do not honour one — the dangerous direction, so the guard works
per OPERATION.
- `/api/insights/feedback` and `/api/admin/backups/{id}/restore` are not in the
document at all. Both are named in the guard's exemption list with a reason,
not filtered out silently.
## The guard
`idempotency-contract-publication.test.ts` reads the expected set out of the
route tree on every run — nothing is hand-listed, so it cannot fall behind the
code. It fails in both directions: a new wrapped route that forgets to publish,
and a published claim no handler honours.
A fourth assertion checks the exemption list itself is still true. It earned
its place immediately: the first draft exempted the whole `/api/medications`
family on the strength of a grep for path literals that came back empty. The
paths are not written as literals there; the emitted document carries 19 of
them. The test went red, the claim was wrong, the list is two entries now.
Measure the artefact, not the source that builds it.
A lab panel files itself against the visit it came from, and never has to. The
OCR review and the manual form both offer the suggestion through the one shared
field, anchored on the draw date rather than on now, so backdating a reading
moves the offer with it. The commit writes one link PER RESULT ROW, which is
why the join is m:n: a marker re-run on a different day belongs to its own
visit. Best-effort by design — the readings are the write that matters and must
not be lost to a link that could not be made.
The condition side gets a Visits about this card with an action that pre-links
the episode. From the episode side only, and that is a decision: an episode
spans weeks and a visit is a point, so matching the other way would be wrong
often enough to teach a person to ignore every suggestion in the product.
A due checkup gains a second action beside erledigt. Filing the visit as done
satisfies the checkup through the same forward-only, idempotent primitive the
checkbox uses, so doing both is one satisfaction and no error. The two actions
answer different questions — one says it happened, the other says where.
An appointment now appears where a person looks, and NEITHER surface relaxes
the reminder-table exclusion that keeps it out of the preventive-care lists.
Both read the visit table: a quiet line under the dashboard checkup block for
anything inside two weeks, and one rail item for today or tomorrow. A diff in
either that touches origin is the wrong fix.
A visit write now also refreshes the reminder reads. Closing a checkup moves
its next-due date on three surfaces that would otherwise keep showing it due
until their own poll.
OpenAPI regenerated in the same commit.
A lab runs its own method on its own device, so the window a physician
evaluates a value against is the one printed next to it on that report,
not a catalog band averaged over every lab. Reading a report pulled the
value and dropped that range: the extraction asked for it, the commit
handed it to the biomarker resolver, and the resolver only uses it when
it mints a NEW marker. For every analyte the user already had — the
normal case after a first panel — the printed range was read and then
thrown away.
Add `sourceReferenceLow` / `sourceReferenceHigh` / `sourceReferenceText`
to `LabResult` and carry the range from the report to the reading. The
text is stored whether or not numbers can be read out of it, because
"bis 5,0" and "negativ" state a window too; a range that could not be
parsed is kept verbatim, never replaced with invented bounds.
`resolveEffectiveReferenceRange` is the one place the precedence lives:
the source window wins for that single reading, and the catalog stays
the net for readings without one. A printed string with no derivable
bound does not displace the catalog. Every surface that judges a
reading routes through it — the API verdict, the doctor report, the
insights blocks and per-marker assessment, the coach snapshot, the MCP
read and the FHIR `referenceRange` — so a value cannot read in range on
one surface and out on another.
The API keeps `referenceLow` / `referenceHigh` as the window in force,
so a client that reads only those two renders the right range and the
right verdict with no change of its own; the catalog band, the source
window, its origin and a divergence flag ride alongside for clients
that want to show both.
Read side: the chart draws the source window as a second, visually
distinct band with the reading it came from named beneath it, the
reading rows name the printed range and the catalog band where the two
disagree, and the doctor-report PDF prints the report's own string with
a footnote naming the saved range. Entry side: one optional field on
the lab form, folded away until asked for, and the printed range is
editable on the scan review screen.
The range parser covers the notations German and English reports
actually print — two-sided, one-sided, comma or dot decimals, thousands
separators, trailing units — reads units without ever converting them,
and refuses to guess on anything ambiguous.
Six provider syncs returned a bare count, which narrowed their failures away:
a Nightscout window in which every SGV row was refused by a constraint
returned the same 0 as a quiet night, and an Oura or Polar run whose
collections all failed returned 0 too. The routes then answered 200 with a
count and the settings card drew a green tick over it.
Every sync now reports what it imported AND whether any part of it did not
land, and the route resolves the outcome server-side so the web card and the
native client cannot disagree. Google Health keeps its 502 for a run that
failed and wrote nothing, but a run that failed after writing some resources
is now a partial rather than a discarded result. The labs OCR commit carries
the same resolved outcome.
Lab-report OCR now rasterizes a PDF to page images for any non-Anthropic
vision provider, exactly as the document vault already does — Anthropic keeps
its native PDF block, every other vision provider (codex/OAuth, OpenAI, local)
reads the rendered pages. The capability probe reports `pdfSupported` whenever
the picked provider reads PDFs natively or the server-side rasterizer is
available, so the "scan a PDF" affordance shows for a codex provider too.
Drop the dedicated document-scan (Lab-OCR) provider: labs OCR always uses the
main configured provider chain. Removes the settings card, the `aiOcr*`
resolution branch, the GET/PATCH handler fields, the OpenAPI schema fields, the
i18n keys across all locales, and the five backing `User` columns (migration
0234). The on-device tesseract fallback is untouched.
Dashboard hero: stack the overdue-dose action directly beneath the verdict
text (tight gap, left-aligned) on every viewport instead of pushing it to the
far container edge on desktop.
Drop the unreachable text-only-model reason from the OCR capability
enum and its DTO type. Replace the literal NUL delimiter in the Fitbit
dedup map keys with a named constant so text tooling stops treating the
file as binary (behaviour unchanged). Correct the Coach tool-executor
docblock, which overstated cross-tool snapshot-cache sharing.
Add an opt-in in-browser OCR path so a user whose provider cannot read
images (ChatGPT-OAuth/Codex, a text-only model) can still scan a lab
report. The capability probe gains a `mode` field: `vision` when the
provider reads the image directly, `text` when local OCR is enabled and
any provider is configured, or null. The extract route grows a text
branch that accepts the in-browser-OCR'd text as JSON and structures it
through the existing schema + review/commit flow; the raw image never
reaches the server in text mode. New `labsLocalOcrEnabled` user column
(migration 0184, default off) gates it, with a dedicated PATCH endpoint.
Add the OcrCapability, OcrExtract, and OcrCommit operations under
/api/labs/ocr/* and regenerate the YAML. Document the multipart upload, the
untrusted-extraction contract, and the source OCR commit value. Mark the three
routes exempt in the module-route gate inventory under the labs data-layer
reasoning.