Add GET /api/documents/inbound/capability (document provider order + vendor-blind
egress class), repoint the vault capability hook to it, and show a per-egress
notice in the detail sheet before a document read leaves the machine. Cover the
document provider order, the egress consent gate, and the notice with tests.
i18n x6 + OpenAPI regenerated.
Co-Authored-By: Marc-André Bombeck
Add a document-class provider resolution that demotes external/train-by-default
providers below local (local -> BYOK no-train API -> operator key -> codex last)
and a pick-based document-egress consent gate that requires an active receipt
for any external provider. Coach/insights keep the app-wide cost-first order.
Co-Authored-By: Marc-André Bombeck
P2-1-T2/T3/T4/T5/T6.
- assist.ts + [id]/suggest: one provider call proposes a { title, kind,
documentDate } filing draft over the stored original or browser-OCR text.
Runs the full extract gauntlet (module → provider → consent → rate-limit →
budget) but writes nothing — never touches the fact rail or status.
- describe.ts + [id]/summary?mode=summary|text: session-only summary or raw
transcribed text. Descriptive only, no diagnosis; persists only the budget
ledger + audit.
- content-index.ts + [id]/index: normalise → tokenise (de-accent, stopword
drop, whole-word) → HMAC-SHA256 under an HKDF-derived subkey → store the
token array plus AES-256-GCM ciphertext of the text. Vision path transcribes
via the provider (consent + budget); text path indexes posted local-OCR text
(no egress). Gated on the existing AI consent — no new per-user toggle.
- crypto.deriveSubkey: HKDF-SHA256 purpose-bound subkey off the active key.
- list search now unions title/filename ILIKE with a GIN-accelerated
whole-word content-token match; DTO gains hasContentIndex.
- usage gains assistAvailable + contentIndex { enabled, indexedCount,
totalCount }; detail gains hasContentIndex.
Thread the interpretation block into the generic per-metric assessment
path (the ~40 HealthKit metric cards). For a metric the knowledge base
covers, the prompt now carries the band table plus the server-computed
position of the current value, so the assessment names where the value
sits on the guideline scale and judges the trend by that position — the
visceral-fat 2.7 case now reads as "comfortably in the healthy band"
rather than reciting how many times it was logged.
The block is absent for uncovered metrics and for a sex-split band with
no known sex, so those cards fall back to exactly today's personal-relative
behaviour. Band edges carry no strip risk here — the generic assessment
path runs no post-generation numeric grounding gate.
Add two shared cross-surface contracts and compose them into the
assessment cards (base-system) and the comprehensive overview + daily
briefing (insight-generator):
- anti-recitation: a measurement count, a logging cadence, or page
mechanics is not an insight and may only appear when it carries a
consequence — stated as the consequence, not the count.
- interpretation-depth: say what a value MEANS, place it on its guideline
scale where one is attached, judge the trend by band position, and frame
the consequence without diagnosis; where no band exists, interpret
against the person's own baseline and say so honestly.
The cross-surface coverage guard now asserts both contracts reach the
surfaces that enforce them. The briefing's grounding and format machinery
is untouched.
Add a typed, hand-curated registry mapping metrics to guideline
interpretation bands derived strictly from the health knowledge base —
band edges, plain labels, valence, direction-of-good, and a source tag,
each entry citing the KB file + section it came from. A metric the KB
gives no numeric band for gets no entry, so its assessment stays
personal-relative rather than inventing a threshold.
The band position is computed server-side (`classifyBandPosition`, strict
upper-edge semantics + edge-hugging proximity), and `buildInterpretationBlock`
renders the table + placement into the assessment prompt so the model states
where a value sits and what the band means instead of reciting counts.
`interpretationBandEdges` exposes the edges for any numeric grounding gate.
Covers visceral fat, resting heart rate, blood oxygen, respiratory rate,
body temperature, sleep duration, waist (sex-split), waist-to-height,
pulse-wave velocity, and BMI. Tests pin every band edge against the
guideline value so a silent edit breaks a test.
The structured Allergy / FamilyHistoryEntry stores reached FHIR export
only: the doctor-report PDF shipped without the two sections every
clinical intake asks for first, and the Coach knew allergies only from
the free-text answer — a penicillin allergy recorded in the structured
records UI never reached either surface.
Doctor report: two new section toggles (allergies, familyHistory; both
default ON, same stance as labs) across the prefs schema, the export
selection, the aggregator, the PDF renderer, and the export panel. The
tables render stored fields only — substance/category/kind/severity/
reaction/status and relationship/condition/age-at-onset; the reaction
decrypts fail-soft per row and free-text notes are never read, matching
the illness-journal stance. OpenAPI regenerated for the extended
sections shape.
Coach: the about-me context block now appends the structured rows as
clearly labelled entries (authoritative over the free-text answer on
conflict); the projection is a pure composer with unit tests. New
section labels land in all six locales.
The shared formatting contract asks every narrative surface for short
blank-line paragraphs, but two field-level instructions overrode it:
the assessment OUTPUT FORMAT pinned 'flowing prose (2-4 sentences)'
and the briefing paragraph asked for 'a connected 70-160 word read' —
so both consistently arrived as one unbroken block regardless of the
contract. Both now spell out 1-3 short paragraphs separated by a
blank line written as \n\n inside the JSON string.
The shared contract banned every structure beyond 1-3 paragraphs, so
longer replies read as unformatted blocks. It now asks for short
blank-line paragraphs (2-4 on longer replies), permits a plain "- "
list when a reply genuinely enumerates three or more parallel items,
and one bold key takeaway per reply. The Coach ground rule keeps
prose-first; a list is the exception, never the default shape.
The daily-briefing number-grounding gate strips the whole dailyBriefing block
to null when it finds a restated figure it cannot trace to a pre-computed
server value. Its allow-set covered signalsOfDay plus the glucose / labs /
preventive-care / workout / grip / waist / pain aggregates, but omitted the
core vitals blocks — weight, blood pressure, pulse, body fat, mood, sleep,
activity, rate-pressure product, seasonal variation, historical comparison —
even though the briefing prompt is fed all of them.
A verdict-first briefing that restated a blood-pressure average or a resting
pulse (figures every BP-tracking user has) therefore tripped the gate on every
run: one ungrounded number strips the entire briefing while the rest of the
insight refreshes, so the surface reads as "every other text updated, only the
briefing is gone".
Admit every figure the server pre-computes for those blocks into the allow-set.
The anti-fabrication guarantee is unchanged — only server-computed numbers are
ever added, so a genuinely invented value is still flagged and stripped.
The comprehensive briefing aborted its provider call at the 120s budget
on a full-history account — a heavier reasoning model spent longer than
that before the first token, so the chain timed out, fell through to no
fallback, and the briefing rendered its greeting with an empty body.
Raise the comprehensive timeout to 180s so a slow first token no longer
clips a generation that would otherwise complete; the per-user response
timeout still overrides it for operators who need more, and the worker
cap tracks this value.
The insights alert cards computed medication compliance with the naive
`schedules.length × days` denominator, so a fully-adherent weekly
injectable read ~14% and fired a false low-compliance alert (and a paused
medication mis-fired the same way). Route the alert path through the
canonical `calculateCompliance` engine like every other compliance
surface — pause eras, schedule revisions and dose windows honoured.
Document confirm: committing a numeric observation against an existing
biomarker silently replaced the stated unit with the catalog's, with no
conversion — a stated "6.1 mmol/L" against a mg/dL marker committed
"6.1 mg/dL". Reject on a unit mismatch (fail closed) so the user
reconciles on the review screen, the same posture as the unit-required
guard.
Wire grip strength, waist (with waist-to-height) and pain into the
briefing feature set and the grounding allow-set so the clinical-depth
signals reach the daily narrative; the mental-health screeners stay
excluded by construction.
Biomarker detail: fall back to a generic description for catalog-less
markers, and keep the heading + description with a retry on the error
state. Collapse the medication External-API and Lifecycle groups by
default. Re-enable the insights scroll-to-top regression and assert the
shipped main-content reset.
Align the EN haemoglobin/haematocrit display names to British spelling.
Drop the redundant `(userId, deletedAt)` index on inbound documents — a
strict prefix of the new three-column index. Trim the dead
MedicalDisclaimer variants to the one live `dataPosture`. Move the
crisis-resources card off the retired amber tint to a calm
destructive-subtle border. Add a registry invariant asserting every
biomarker signal key resolves to a real catalog slug.
Introduce one canonical "what this is / is not, not a medical device"
statement plus a layered, calm disclaimer set, and apply it on the
surfaces that earn a context note.
- New shared `MedicalDisclaimer` component renders the copy as plain
React text (no markdown renderer, per the XSS posture). Variants:
intendedPurpose, coach, insights, labs, medications, emergency,
dataPosture.
- New `selfDescription` i18n namespace across all six locales: the
intended-purpose line, the per-surface notes, the emergency note, and
an explicit data-posture statement (self-hosted, data stays on your
instance, encrypted at rest, BYOK / local model, no third-party
analytics).
- Apply once per surface, layered not plastered: a calm standing line on
the Coach new-chat hero, one framing line in the Insights layout shell
(covers the briefing and trends), a labs note (ranges vary, only a
clinician interprets), a neutral medications note, and the data-posture
line where the provider key is configured. Mental-wellbeing keeps its
existing crisis copy; records keep their existing note.
- Extend the Coach system prompt (EN + DE) with a self-description ground
rule so the Coach answers "what are you / where does my data go"
honestly and stays non-diagnostic.
App-authored copy was already observational; the few diagnose/treat
tokens are negations, idioms, or framed as the user's own self-reported
statement, so no softening was required.
Rework the proactive Coach nudge so it reads like one human checking in on
another instead of a terse English alert.
Copy: the title is a name-led greeting that rotates day to day ("Guten
Morgen, …" / "Hey …"; the 05:15 frame makes a morning opener safe) and the
body is one calm observation plus a gentle invite — never an imperative.
Drop the verbatim quoting of the user's decrypted self-context sentence:
the focus is referenced only by presence, abstractly, and is no longer
decrypted at all. Remove the 80-char mid-word clamp; each channel composes
to a length budget instead.
Cadence: suppress the nudge when the user already talked to the Coach in
the last ~24 h. With the existing rolling frequency cap (never two in a
window) and the no-trigger-no-nudge rule, the surface stays anti-nag.
AI enrichment (opt-in via the coach prefs blob, default off): when a
provider is healthy the body is composed through the model under hard
guards — a per-user budget gate, a tight per-call timeout, a per-tick
ceiling so the sequential tick can't stall across the user base, and the
outbound safety screen on the output. Any error, timeout, budget refusal
or unsafe reply falls silently back to the deterministic template, which is
always the default and the fail-closed fallback.
Localized greetings, per-trigger bodies and the focus note across all six
locales; the prefs opt-in is surfaced in the notification settings.
The requirePublicHost egress paths now dispatch through undici's own
fetch, so the existing global fetch stubs in the AI, GeoIP, moodlog,
send, and ntfy suites no longer intercept them. Delegate undici's fetch
back to the installed global stub so each suite keeps asserting its
request shape and canned responses without exercising the network.
The daily briefing and the per-metric status assessments ignored the
per-user response-timeout setting: they hardcoded their per-surface
AI_BUDGETS default, so raising the timeout for a slow self-hosted or
local backend had no effect and the surfaces still aborted early. Only
the coach threaded the setting through.
Resolve the effective timeout from User.aiResponseTimeoutSeconds in one
shared helper and thread it onto every non-coach generation call:
- runStatusCompletion is the single chokepoint every status / reference
surface funnels through (per-metric cards, the batched assessment, the
derived assessments, the period narratives), so resolving it there
applies the setting to all of them at once — on both the upstream
call's own timeout and the outer wall-clock cap that previously clipped
a raised value at 60 s.
- The comprehensive briefing generator and the user-initiated regenerate
route thread it onto the main generation, the JSON retry, the grounding
retry, and the paragraph re-roll.
On a failed generation the briefing keeps its last good text — no cache
row is written, so the previous payload stays intact — and records a
dated, append-only failure marker (no migration; reuses the audit-log
JSON pattern). The read path compares the marker against the last
successful generation so it can tell a held briefing from a current one.
A cluster of related correctness gaps in the compliance denominator and the
insight surfaces that read it:
- Paused medications. A doctor-ordered pause flipped `active` off and stamped
`pausedAt`, but resume nulled `pausedAt`, so the interval was irrecoverable
and every paused day became a missed dose after resume — collapsing
adherence across the Health Score, the BP gate, the Coach snapshot and the
report. Persist pause intervals as durable eras (new `MedicationPauseEra`,
migration 0211), close the open era on resume, thread the eras through the
compliance context and drop slots inside a paused interval from the
denominator. Every compliance call site now fetches and passes the eras.
- Health Score eras. The Health-Score fast path was the only compliance
caller that skipped the shared schedule select and schedule revisions, so
after a schedule edit it scored past days against the current schedule
while every other surface used the historical era. Switch it to the shared
select and thread revisions (and pause eras) like the canonical path.
- Insight bucketing timezone. The per-metric status series bucketed every
user in Europe/Berlin, so a near-midnight reading landed in the wrong day
and the relative-day phrasing was off for non-Berlin self-hosters. Thread
the resolved user timezone and key off `userDayKey`.
- Correlation significance. User-facing correlations surfaced from five
paired points with strength read straight off the coefficient. Route them
through the rigorous engine (at least twenty pairs, Student-t p below 0.05)
so small-N coincidences no longer render as confident findings.
- Resting-pulse input. The mood-versus-"resting pulse" correlation fed the
raw PULSE day-mean, which folds in workout HR, under a "Resting pulse"
label. Feed `resolveRestingPulseSeries` (clean RESTING_HEART_RATE, else a
daily low-percentile proxy) so the label matches the input.
Cap and page the unbounded reads that let a multi-year tenant pull its
whole measurement history into memory on hot paths.
- extractFeatures: read newest-first with a 6000-row cap (mirrors the
Coach snapshot) and reverse to ascending, so a 365-400 day window no
longer materialises hundreds of thousands of PULSE / glucose rows per
Coach turn and nightly briefing. Full-history extremes still come from
readAllTimeExtremes, so the cap never relabels an extreme.
- Add a (userId, measuredAt) index on measurements (migration 0210) for
the type-less range scans that extractFeatures and the consolidation
passes drive; the existing (userId, type, measuredAt) cannot serve them
as a prefix.
- runConsolidation: keyset-paginate the per-(user, type) source scan on
(measuredAt, id) instead of one unbounded findMany, removing the giant
single result-set allocation that drove the boot storm. Output is
identical; covered by a paging unit test.
- Stagger the boot-time backfill enqueues with increasing startAfter
offsets so the heavy full-history loads no longer all drain in parallel
onto one tenant at startup. Non-boot callers keep their immediate send.
- Backups: replace load-all-then-stringify with paged narrow-select reads
for the weekly and off-host writers, bounding transient heap. Off-host
still includes soft-deleted rows. Covered by a serialize/restore
round-trip test.
- Drop the per-request auth.bearer.success audit insert; the wide event
already records auth_method and user_id, so the polling client no longer
drives a write + pool checkout on every request. Failure-path audit
stays.
- Coach conversation detail: cap the message include to the newest 200 so
a long thread no longer decrypts every message on each open; response
shape unchanged.
- Admin audit-log: floor the scan at the retention horizon so the listing
and its count stop sweeping open-ended over the substring filters.
Encrypt MedicationSideEffect.notes, MedicationDoseChange.note and
MedicationInventoryItem.notes with AES-256-GCM, closing the last plaintext
PHI columns left after the v1.23 note-encryption rollout. Add the three
*Encrypted Bytes columns (migration 0208), register them in the encrypted-
column rotation set, route every read/write through the note cipher, and add
a transactional fail-closed boot-time backfill for legacy rows. Plaintext
columns drop in a later release once the backfill reports zero remaining rows.
Require a fresh second factor on password change for MFA-enrolled accounts,
matching the existing step-up on MFA-disable and recovery-code regen. Redact
free-text note values from the Zod validation excerpt that can reach a wide
event.
Code-review and design follow-ups before the v1.22 cut. No behaviour change
beyond the items below.
Coach / insights:
- Coach SSE route: a non-provider error raised inside the open producer (e.g. a
DB failure persisting the reply) now records on the wide event and forwards to
GlitchTip instead of masquerading as a provider outage; it still surfaces an
error frame since the stream is open.
- Lab ↔ outcome correlations: hedge faint-tier pairs ("a faint hint, if
anything … too small to lean on") so the narrated confidence never outruns the
effect, matching the daily engine.
- Reminder sweep: mint the plan-review reminder and clear the plan's reviewDate
in one transaction so a half-applied write can't duplicate the reminder.
- Fold the two raw provider-fallback runners onto one shared chain walker.
Design / accessibility:
- Lift sub-44px tap targets to the 44px mobile floor (suggested-action card,
Coach bubble affordances), brighten the resting timestamp clock for non-text
contrast, and route the lab chart height through the shared constant.
- Replace ad-hoc arbitrary type sizes with the scale token and static size-11
icon buttons with the named icon-lg variant.
- Settings: add the missing About card-header descriptions, normalise the
low-stock numeric widths, and relax the notification-status rows from nested
cards to a divided list.
Housekeeping:
- Drop three unused exported types (and the now-orphaned source const), re-stamp
v1.21.5 comments to v1.22, and register the live-judge eval harness as a knip
entry so the script-only module resolves.
Add three per-user AI settings, all configured in the app (no new env vars),
behind one migration (0198):
- Response timeout (seconds). A per-user upstream timeout threaded onto the
provider call. A local/self-hosted backend can take more than the legacy
60 s default to load a model on the first request, which timed the Coach
out before the first token. Defaults to a generous ~180 s when unset;
validated to 10–600 s.
- True token streaming for the local OpenAI-compatible client. The client now
POSTs with stream:true and parses the SSE delta chunks, emitting tokens as
they arrive, with a per-idle-gap timeout instead of a whole-call ceiling and
a graceful fallback to the buffered path when a server cannot stream. The
Coach route runs the provider call inside the response stream behind a
keepalive heartbeat so a reverse proxy no longer idle-drops the connection
during model load or generation. Grounding, refusal, and sentinel-stripping
still run on the complete reply before any token reaches the client.
- Dedicated document-scan (Lab-OCR) provider. Off by default — the OCR path
uses the main provider chain. When enabled, lab-report processing resolves a
separate provider/model/key, with a hint when the chosen model is not
vision-capable. The new key column is AES-256-GCM at rest and registered in
the encrypted-columns registry and the rotation script.
i18n: EN + DE translations for the new Settings → AI strings (other locales
carry interim copies to keep locale parity green).
Add the medication-class to target-vital map and the adherence-to-vital
storyline so the Coach can surface, association-only, when an adherence dip
lines up with a drift in the vital that medication targets. Never causal,
never a dose-change suggestion; dose questions defer to the prescriber.
Add an in-repo CUSUM / binary-segmentation changepoint detector over DAY
rollups with a high firing bar (persistence plus within-segment-spread
magnitude) so the Coach can name a dated, sustained level shift instead of a
vague "lately", and stays silent when no break clears the bar.
Surface which source recovery is read from when two sources materially
diverge, with the honest reason, instead of silently averaging.
Lay the n-of-1 experiment substrate: a nullable encrypted outcome column on
coach_plans (migration 0197), the review_due / reviewed lifecycle statuses
wired into the encrypted-column registry and rotation script, and a daily
worker that reads the experiment result back in the user's own before/after
numbers — association-only, honest-null, no cheerlead on an adverse trend,
and routing a worsened vital to the clinician. The user-visible read-back
stays behind COACH_EXPERIMENT_VERDICT until its review cases clear.
Extend the outbound dose screen to catch experiment-shaped dose changes
(halve / skip / stop a medication to test it). Add evaluation cases and unit
tests for each new surface.
Feed the user's own usual blood-pressure range — the median plus or minus
MAD band, computed from the readings already in the snapshot — into the Coach
context with a clause to quote those bounds verbatim, so a band is never
invented from window means. Thread the stored first name into the prompt and
surface it sparingly on roughly one turn in three, never as a rote greeting.
Add the context-layer rule that names the relevant context before a
suggestion, a single-chart-token allowance that works on every provider, and
a per-turn opener hint.
Wire the already-shipped suggested-action SSE frame to render: the streaming
hook collects it and the message thread shows the action card, mirroring the
reminder-suggestion path. Extend the eval golden set with band-verbatim,
chart-token, and no-overclaim cases, and pin a red-team case proving the
chart-token syntax cannot smuggle a dose past the outbound screen.
Replace the fixed three-beat assessment skeleton with a menu of moves the
model orders freely, and reframe the daily briefing from per-signal
enumeration to a verdict-first read that weaves the few salient signals into
one associative story. Score cards now name both the helped and the hurt
contributors, translate the band into a forward read for the day, and close
on one grounded nudge; the deterministic fallback gains the same band-meaning
sentence so provider-less and demo accounts read as well.
Add a hash-seeded opener-archetype rotation (deterministic, testable) so
consecutive cards and days no longer open the same way, a day-rotated sampling
seed for the period narrative in place of the fixed reference seed, and a
shared paragraph formatting contract. Add a score-number grounding gate and
extend the briefing grounding allow-set to the glucose, labs, preventive-care
and workout aggregate blocks so real figures can be cited without opening the
gate. Gate the daily score warm on an input fingerprint so a quiet day skips
the model call.
Add unit coverage for the closed when-grammar and `---REMEMBER---` capture
(propose-free active write, recall-only note, malformed drop, per-user cap),
the closed action-card allowlist (off-list and bad-interval rejection, no
auto-apply), the daily sweep (due flip, plan-review mint, undecryptable
skip), and the routes (owner-scoped no-IDOR 404, field-by-field build, the
off-allowlist 422, and the module gate).
Capture the casual "remember this / remind me about X" utterance the Coach
used to drop. The model emits a provider-neutral `---REMEMBER---` sentinel
that the chat route strips inline on EVERY turn — so a short conversation,
where the request usually happens, no longer waits for the long-chat memory
worker. The note is the user's own explicit request, so it is stored
silently as active; a validated closed grammar (ISO date, relative +Nd/+Nw,
or a context cue) sets the trigger, and a malformed block is dropped.
Surface the stored reminders two ways. A snapshot recall sub-block injects
near, overdue, and already-surfaced reminders so the Coach references them
unprompted next turn. A daily sweep flips overdue reminders to due and
mints a one-off review reminder from a plan whose review date has passed,
finally activating the plan review-date column. In-app surfacing only here;
push delivery follows later.
Also generalise the propose-then-confirm action card: a closed
`---SUGGEST-ACTION---` allowlist (a preventive-care item or a reminder note)
rides an additive SSE frame. Nothing is created until the user confirms.
Thread the rolling early-detection pass and the labs -> outcome pass through the
correlations route and the coach correlations reader, both of which already
build the same daily series. The route response gains an optional `emerging`
block (recent-window pairs not yet established, flagged provisional) and a
`labCorrelations` block; the coach reader gains matching `emerging` and
`labDrivers` fields. Regenerate the OpenAPI contract for the new response shape.
The comprehensive insight that feeds both the daily briefing and the
insights trend narrative is a single reasoning-heavy generation over the
full feature snapshot. On a large account it runs past the provider
client's fixed 60 s timeout and is aborted mid-stream, so the chain
reports an all-failed outcome, no cached block is written, and the
surface falls back to its static placeholder on every visit. The warm
paths bounded the same generation tighter still (45 s) and counted the
failure with no observable signal, so the blank state was permanent and
silent.
Make the per-request upstream timeout configurable
(CompletionParams.timeoutMs, default 60 s for every existing caller) and
raise it to 120 s for the comprehensive surface only, across all three
generation paths: the on-demand regenerate, the on-demand warm, and the
nightly pre-generation. Raise the warm-pass budget to 130 s so the
provider's own timeout governs a genuine completion rather than clipping
it early. Annotate the comprehensive failure in both the generator and
the nightly loop so a generation that keeps failing is queryable instead
of leaving the surface empty without a trace.
Drop the unused extractCoachClaims helper (+ its CoachClaim types) and the
unused CoachPlanStatus type — the deterministic grader uses the boolean claim
helpers directly, and the plan status enum is consumed via the const array.
Update the dashboard-snapshot suite for the locale-keyed cell: stub the i18n
context and add getQueriesData to the fake query client so the locale-aware
preload probe is exercised.
A Codex backend 400 on the richer Coach tool-call request (and the daily
briefing) propagated out of the provider chain runner unwrapped: a 4xx other
than 401/403/429 was not classified as a hard provider failure, so it bypassed
AllProvidersFailedError and rethrew from the chat route as an HTTP 500. Plain
status-card completions, which never send a tools array, were unaffected.
Harden the failure path so a provider error can never surface as a 500:
- Treat any 4xx a provider wire returns as a hard failure so it cascades and is
wrapped in AllProvidersFailedError. InsightSchemaError (the strict-JSON 422
surface) stays the only non-hard class, so that diagnostic is preserved.
- The chat route now also classifies a tagged provider error that reaches its
catch unwrapped and emits the graceful coach.provider.* SSE frame instead of
rethrowing. Only an untagged error keeps the 500 + error-reporter path.
Make the Codex 400 diagnosable: fold the redacted upstream response body into
the thrown error message and the chain wide-event, so the field the backend
rejects is visible in logs (it was previously dropped).
Fix the Codex request shape behind the tool-call 400:
- Send the required strict flag on each Responses-API function tool.
- Replay prior assistant turns as output_text, not input_text, per the
Responses API multi-turn contract — the violation only surfaced on round two
of the tool loop, which is the path that broke.
Codex-only; the OpenAI, Anthropic, local and mock providers are untouched.
A deterministic golden-set + red-team floor that gates every Coach change
offline, plus an opt-in live-judge layer that scores warmth and groundedness
with a pinned model only when COACH_EVAL_API_KEY is set.
- 42 golden cases across grounding, cross-metric, data-honesty, provider-parity
and own-baseline taxonomies, graded by behaviour-anchored binary criteria.
- Structured-claim grounding helpers extend the numeric verifier from numbers
to threshold / own-baseline / confident-verdict claims; the numeric API is
unchanged so the existing grounding tests stay green.
- Red-team battery asserts the existing inbound refusal, outbound dose screen,
acute red-flag escalation and learn-link guards fire.
- Nightly + manual workflow runs the deterministic floor always; the live judge
step runs only when the secret is present and is non-blocking.
Add a CoachPlan model: durable, user-scoped if-then implementation plans
the Coach proposes and the user confirms. The background memory-refresh
pass extracts plans the user agreed to and writes them as "proposed"; only
the user-facing PATCH on /api/coach/plans/[id] activates one. The free-text
cue / action / target are encrypted at rest with the shared Bytes codec;
metric, status and dates stay plain so the injection picker ranks without a
per-row decrypt. Active plans fold into the Coach snapshot's memory block
and survive the prompt-budget drop alongside facts. List, confirm and
soft-delete routes are Coach-gated and narrow the owner from the session.
Key the dashboard-snapshot query cell by locale so a locale switch reads
freshly localised tile copy on its own cell instead of serving the prior
locale's prose until the slot goes stale. The zero-arg key stays
byte-identical, so every prefix-based invalidation keeps reaching the
locale-keyed cells; the offline-banner and preloader match on the prefix.
Migration 0194 (additive CREATE TABLE). Encryption-rotation registry and
script cover the three new Bytes columns.
Add a deterministic narrative layer the Coach narrates rather than invents:
- Briefing callbacks + forward-looks: the daily briefing recalls the prior
period and points ahead, anchored to a figure, one callback per surface.
- Tension Verdict: when the readiness composite's contributors disagree
(good sleep but a rising resting pulse), surface a brief honest internal
read on the health-score card. A real clinical red-flag is never reconciled
away into a calm verdict.
- Streaks + return-to-baseline: a pure detector over a DAY-bucket series
closes a worry once a metric returns inside the personal MAD band after a
genuine out-of-band run; a gap day breaks a streak.
- Coach prompt: add develop-discrepancy, roll-with-resistance, and an
anti-persuasion clause; ratchet PROMPT_VERSION.
The tension flag and streak event are computed server-side; the card lines
render resolved fields and the model narrates them.
The post-hoc numeric verifier only graded the tool path, so local /
no-tools providers got no grounding net. Grade the no-tools prose
against the snapshot the model was shown — its structured sections,
which already carry the correlations-snapshot block — and flag any
cited number that traces to no snapshot leaf. Non-blocking: the count
is annotated under coach.prose.number_unverified and the unverified
token soft-stripped, never a stream-killer. Only runs when the full
figures were delivered this turn; a cheap follow-up has no fresh set
to grade and no-ops. Tool-path behaviour and tolerance unchanged.
The correlation discovery matrix gained two non-MeasurementType channels
(medication compliance, symptom severity), but the Coach get_correlations
reader still filtered its measurement query by hand and never built the new
channels. The channel keys leaked into the Postgres enum cast and threw for any
user with medication or illness data.
Extract the route's two channel-series fetchers into a shared module
(correlation-channel-series.ts) so both consumers build the channels identically,
then wire the reader to derive its type filter from discoveryMeasurementTypes and
fold the fetched compliance + symptom series into the discovery input. The reader
now humanises the channel keys to "medication adherence" / "symptom severity".
The labs block already surfaces value + unit + range + date, and the Coach
prompt instructs it to quote them verbatim, so labs are cited end-to-end — no
gap. Pin that with a tool-level guard asserting value and unit survive
get_labs, so a future change to the labs pass-through can't quietly strip the
number the Coach needs to cite.
The Coach may point at a published /learn guide, but only a slug in the
catalog. That was a prompt instruction with no enforcement — a model could
hallucinate /learn/lower-your-cortisol and ship a dead link. Add a deterministic
post-filter on the assembled reply that keeps a published slug verbatim and
strips any reference whose slug is not in the catalog, tidying the prose around
the removed link. Makes the catalog's impossible-by-construction claim an
enforced guarantee.
The illness-recovery tool and the illness context block carried only labels,
lifecycle, and dates plus the recovery composite — never the recovery-gap, the
metric that drove it, the nadir, the pre-onset deviations, or the red flags the
illness card renders. So a recovery question got the composite, not the gap the
user sees, and a sustained-fever or low-SpO2 escalation was invisible
in-conversation.
Run the existing episode-correlation read-layer (the same engine the card and
the red-flag notifier use) for the most relevant episode — active first, else
the most-recently-resolved — and attach the computed scores to get_illness_recovery,
coverage-gated so a thin signal yields nothing rather than a fabricated number.
Document the illnessScores shape in the Coach prompt so it restates those
numbers verbatim and escalates a red flag rather than reassuring.
Each retrieval tool re-scoped the snapshot to a single source, so its cache key
never matched the inventory's full-source build — a four-tool turn paid four
extra full snapshot builds on top of the inventory's. Thread the inventory's
probe scope to every tool so a per-tool read lands the one cache entry the
inventory already primed, collapsing the turn to a single build (a window
override still gets its own correct build). Raise the loop's round cap from two
to three so a sequential cross-metric why-chain isn't starved; the per-round
token budget is unchanged and the absolute ceiling tracks one above so the
final round still forces prose.
