The auto-read vault specs mocked only the document capability probe, leaving
the usage endpoint (assist + content-index flags) unmocked — so the read
action rendered but stayed disabled and the click timed out. Mock the same
three endpoints the working document-AI e2e does (usage, labs OCR capability,
document capability) so the action is enabled.
One opt-in in AI settings reads every uploaded document with the configured
provider — no per-document tap. Off by default; the vault stays local-first
otherwise. Scanned PDFs and photos are read too (a PDF's pages are rendered to
images so an image-only provider can read them). The background indexer now
follows the same opt-in, so nothing leaves the machine automatically unless it
is enabled.
- ai-providers.md: an off-by-default 'Automatic document reading' section
(what ON sends where, the subscription-training / no-BAA honesty, local =
no egress, regulated users → BYOK no-train API or local) and a vault
local-first caveat
- e2e: the settings honesty-confirm flow (switch → confirm once → write) and
the vault per-egress-notice / explicit-read-action contract, with the
document capability endpoint mocked on every path
Add a single switch in AI settings — off by default — that turns on ambient AI
reading of uploaded vault documents. Turning it on reveals a once-shown,
vendor-blind honesty confirm (subscription providers may use the content to
improve their models; a self-hosted local model never egresses) that must be
acknowledged before the setting is written. Copy is translated across all six
locales.
Only Anthropic accepts a native PDF block; codex and other image-only-wire
providers dropped it. Add a bounded, best-effort server-side rasterizer that
renders a PDF's pages to JPEG images those providers read as ordinary image
input, so a scanned/image-only PDF is legible on the auto-read path.
- @napi-rs/canvas promoted to a direct dependency; its musl prebuilts + the
pdfjs loader pinned into the standalone trace (next.config.ts)
- rasterize-pdf.ts: lazy pdfjs import (keeps DOMMatrix out of the eager chunk),
first 10 pages, 1600px long edge, JPEG q80, never throws
- prepareVisionInput rasterizes a PDF for a non-Anthropic vision provider and
keeps the native PDF block for Anthropic; the document capability reports
PDFs readable via raster so the UI offers the read
The auto-index-on-upload job resolved over the cost-first app-wide chain
(codex priority 1) and gated with the chain-based consent check, so a
pure-codex chain could silently egress a freshly uploaded document to the
subscription backend on upload. Reconcile it onto resolveDocumentVisionProvider
(local-first, codex last) and govern external egress with the
documentsAutoAiRead opt-in: a local pick always runs (no egress); an external
pick runs only when the toggle is ON, otherwise the job stays strictly local.
This tightens the default and makes the opt-in path deliberate.
Add the documentsAutoAiRead per-user Boolean (migration 0231, default
false) that authorises ambient AI reading of uploaded vault documents.
OFF by default keeps the vault local-first for every self-hoster.
- schema column + migration 0231_documents_auto_ai_read
- documentAutoReadEnabled(userId) read helper
- GET/PATCH /api/auth/me/documents-auto-ai-read toggle route; the ON
write mints an ai_full consent receipt for the audit trail
- the document consent gate short-circuits an external pick when the
toggle is ON; a local pick stays ungated and OFF still requires a
receipt for any external egress
The finding is npm's OWN bundled sigstore (v3.1.0) inside the Node base
image, used for publish provenance at build time only. It is not in our
application dependency tree and is unreachable at runtime — same class as the
already-suppressed picomatch entry. Re-check when the base image ships an npm
with sigstore >= 4.1.1.
The local content-extraction reader imported pdf-parse at module top level.
pdf-parse pulls in pdfjs, whose module evaluation references the browser-only
DOMMatrix global; on the server that reference throws during chunk
instantiation and takes down every route sharing the chunk — the document
list, upload, and search among them, so old and new documents alike fail to
load. Move the parser to a lazy import() inside the extraction function
(surface type kept as a type-only import). Verified under plain Node that the
lazy path still extracts a real PDF text layer, so local indexing is
unaffected.
Three local Playwright screenshot harnesses at the repo root were picked up
by a wide stage during the v1.28.0 assembly. They are never imported by the
app build; remove them and ignore the .audit-*.mjs pattern so a root-level
add can't recatch them.
Marks the document line as one coherent milestone: the vault becomes a
place to understand files, not just store them — read, indexed, searched,
shared, and answered — with a local-first / consented AI posture and a
vendor-blind egress notice held at every step. Ships alongside a full
correctness, resilience, and quality sweep down to the low-priority findings.
fullyParallel spreads a spec's tests across workers and each re-runs the
file's beforeAll; two workers seeding the fixed-id vault corpus at once
raced on the document_condition_links primary key (ON CONFLICT only
suppresses its named arbiter index). Guard both seeders with a Postgres
advisory lock so concurrent workers seed one-at-a-time and every insert
re-runs as a clean idempotent upsert.
Add an e2e suite mocking the capability probe, the chat history GET, and the
streaming chat POST so the panel renders (v1.27.31 lesson): the entry opens an
empty safety-noted surface, a turn streams a reply that settles to one assistant
bubble, a non-indexed document shows the calm hint, and no provider leaves no
chat surface. Clear the streamed tail + optimistic bubble once the persisted
twin lands so the reply never double-renders.
Extract DocumentChatConversation (pure) from the chat panel so the open
conversation surface is statically renderable, and pin: composer + always-on
safety note, plain-text/XSS-safe assistant render (raw HTML escaped), the
streaming tail vs thinking placeholder, mapped error keys, and the container's
indexed vs not-indexed gating. Assert the AI section renders the chat slot only
in the provider-available branch.
Add a scoped 'Chat about this document' panel beneath the Read-with-AI
actions. Opens a conversation grounded only in the document's stored text:
a message list + input streaming the assistant reply token-by-token over
the SSE route, with prior turns loaded via GET.
- Availability: the panel renders in the provider-available branch of the
AI section (a pure chatSlot keeps the section statically renderable), and
gates internally on whether the document is content-indexed. Not indexed
shows a calm read-it-first hint, never an error.
- Plain-text only: assistant prose renders via ProseBlocks / StreamedProse
as React children (no markdown library, no dangerouslySetInnerHTML).
- Streaming reuses the Coach SSE frame parser; the wire frames are a subset.
- A one-line safety note states answers describe the document and are not
medical advice, consistent with the summary panel's not-a-diagnosis.
- Query keys via the central factory; history invalidated on the done frame.
- i18n x6 for every new string.
v1.27.31 routed the document AI section through a new document-scoped
capability probe (/api/documents/inbound/capability) instead of the labs
OCR probe. The AI-flow e2e helper only mocked the old probe, so the section
fell back to unavailable and the assist actions never rendered — the tests
timed out waiting for them. Mock the new endpoint with matching availability;
egress-notice behaviour is covered by the SSR component test.
The merge of the pre-v1.27.29 governance branch left two documents.ai
blocks in each locale bundle: the consolidated Read-with-AI keys and a
separate egressNotice-only block. Duplicate JSON keys drop the first, so
documents.ai.egressNotice never resolved. Fold egressNotice into the
single documents.ai block across all six locales.
The vault timeline is virtualized; on the narrow mobile viewport the seeded
Radiology-note card renders outside the initial virtual window (not in the
DOM), so the card-level attribute read waited out the full test timeout on a
non-existent node. The provenance contract is asserted on every project via
the detail status pill; the card marker (viewport-independent render logic,
also covered by the SSR card test) is asserted on desktop where the card is
reliably in view.
Introduce a shared PageAuthGate (header + tile skeleton) and adopt it on the
top-level pages that flashed a lone centered spinner while auth resolved:
mood, measurements, achievements, illness, mental-wellbeing, checkups, cycle,
and the labs / medications / custom-metrics / illness detail routes. The
auth-resolving frame now reserves the page geometry instead of snapping in
after a spinner.
The cycle/illness/records note DTOs, the biomarker context, and the custom
mood/cycle labels each degraded an undecryptable value to a blank/generic
fallback with no signal — a key-rotation gap read as 'left blank' and the
catch was silent. Keep the fail-soft read posture (one bad row never 500s
the list) but emit a wide-event on each swallowed decrypt so a systemic
key gap surfaces. The visible per-field 'unreadable' marker in the UI is
the deferred remainder.
Port the six medication dialogs (log-intake, intake-edit, dose-history-add,
take-all-due, log-injection-site, intake-import) from the centred Dialog to
ResponsiveSheet with the sticky-footer slot, matching every other module's
add/edit form. Form-backed dialogs associate their footer submit button via a
useId form id so Enter-to-submit and the footer button share one form; the
nested slot-nudge confirm rides the same sheet. Primary actions now sit within
thumb reach on phones.
Marc-André Bombeck