Both DELETE /api/devices/[id] and DELETE /api/auth/me/devices/[id]
performed the same four-write cascade (refresh-token lookup, refresh
revoke, access-token revoke, device delete) as separate Prisma calls.
A partial failure between the third and fourth write left the device
row alive with all its tokens revoked - the user's iPad showed in the
device list as "still registered" but every request returned 401.
Extract the cascade into a shared revokeDeviceCascade() helper and
wrap the writes in prisma.$transaction so they commit atomically.
Both routes now call the helper; the audit-log via tag and the wide
event action name stay route-local.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three small reuse wins from the W6 simplify review:
- openapi/registry.ts swaps the lazy require() for a static import.
The ESLint disable + as-cast carried no real benefit; the only
callers always loaded the route table anyway.
- Extract useCoachPrefs() into a shared hook so the settings sheet
and message thread share one queryKey + one fallback shape. Removes
the drift between the two existing inline blocks.
- Replace the 9-deep wantsX boolean ladder in coach snapshot with a
single CoachScopeSource -> MeasurementType[] map. Drives both the
WHERE type IN (...) build and the appleHealthBlocks gate, so adding
a new metric is one entry instead of three parallel edits.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The v1.4.22 prose rewrite shipped without a way for users to signal
whether the new tone is landing. v1.4.23 H7 closes that loop with a
light-touch thumbs row + an admin first-week dashboard.
Schema: extend `RecommendationFeedback` with a polymorphic
`target_type` column (default "recommendation" so legacy rows stay
representative) plus a nullable `reason` field. New index on
(target_type, created_at) so the aggregator can scan per-surface
buckets without a full sweep. Migration `0040_recommendation_feedback_target_type`.
Route: POST /api/insights/chat/messages/:id/feedback. Body
{ rating: "helpful" | "unhelpful", reason?: string }. Verifies
ownership (404 on cross-account messages — no existence leak),
decrypts the message body for the dedup key, and writes the row with
target_type="coach" + a metric-source string encoding the user's
active prefs (`coach:tone=warm:verbosity=default`). Per-user 60/h rate
limit. 409 on a P2002 (already_rated) so a double-click is a no-op.
Aggregator: new `buildCoachFeedbackBuckets()` slices Coach rows by
(promptVersion, tone, verbosity). The summary blob the worker writes
to AppSettings.adminAiInsightsFeedbackSummary now carries both
`buckets` (pre-existing) and `coachBuckets` (H7 addition). Targets
the same singleton row so no second worker handler is needed.
Admin view: `/admin/coach-feedback` (new section slug) renders the
Coach buckets as a small DataTable (prompt version × tone × verbosity
× helpful × not-helpful × n × helpful-rate). Helpful-rate column
tinted via Dracula tokens (green ≥80%, yellow ≥50%, orange below).
First useful question this answers: "Is the v1.4.22 prose rewrite
landing well, or did the warm tone overshoot?" If the helpful-rate
for PROMPT_VERSION 4.23.x drops below 50% within the first 100
ratings, v1.4.24 walks the persona back.
UI: `<ThumbsUp>` / `<ThumbsDown>` buttons (lucide icons, no emoji)
under each persisted assistant bubble. Streaming bubbles + refusals
+ error bubbles skip the row so users can't rate before the message
lands on disk. Localised "Thanks for the signal." after submit.
OpenAPI: spec carries the new POST route + the request schema.
Coverage now includes the H4 + H7 routes — three additions over
v1.4.22 baseline.
Tests +9 unit (aggregator Coach-bucket slicing, admin section
render-paths) plus an integration round-trip for the route.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Pearson p-value derives from the Fisher's-Z normal approximation.
At low df (df < 18, n < 20) the normal approx is generous in the
wrong direction — at df=12 a true p≈0.04 reports as p≈0.025, sneaking
past the p<0.05 surfacing gate. The v1.4.22 product-lead memo flagged
this as Code-MED-03 with a deadline of "before v1.5/v1.6 auto-
discovery ships" because the false-positive cost compounds when
correlations get auto-generated.
Two paths considered:
- Rigorous: drop in a 30-LOC incomplete-beta implementation so the
p-value is accurate at every df. Cleanest fix, no test churn.
- Conservative: raise `MIN_PAIRED_N` from 14 to 20 so the borderline
band is excluded entirely. Smaller patch, no maths-library
dependency.
v1.4.23 ships the conservative fix. The rigorous incomplete-beta
swap is queued as a v1.4.24 candidate; if usage data from the
v1.4.16 B5e feedback aggregator later shows users miss the
borderline cards, the rigorous path lands then.
Trade-off: a small set of borderline-significance correlation cards
stop rendering. Acceptable for v1.5 — the prompt-injection guards +
warm Coach prose make it more important to under-report than over-
report patterns.
Tests +1 (`pearson` n=15 → insufficient pin) plus an existing
`correlateBpCompliance` n=13 test renamed to n=15 + the noise-fixture
extended from 15 to 20 entries to clear the new gate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The v1.4.22 B5 audit removed the placeholder cog from the Coach
drawer header because it had no real wiring. v1.4.23 H4 returns
the cog with an actual settings surface backed by a per-user
preferences row.
Schema: nullable JSONB column `users.coach_prefs_json` (migration
0038_coach_prefs). Null = legacy v1.4.22 defaults; non-null carries
`{ tone, verbosity, excludeMetrics, showEvidenceByDefault }` validated
by the Zod schema in `src/lib/validations/coach-prefs.ts`.
Routes: GET + PUT `/api/auth/me/coach-prefs`. PUT validates the body,
persists the canonical defaulted form, and emits an
`auth.me.coach-prefs.put` annotation with the four-knob meta. OpenAPI
registry covers both verbs; the spec regenerates without drift.
Coach prompt builder: `getCoachSystemPrompt(locale, prefs)` prepends a
per-user OVERRIDE paragraph for non-default tone (neutral / concise)
and verbosity (brief / detailed). Concise tone forces brief verbosity
regardless of the explicit selection — concise == short. Default
prefs leave the prompt unchanged so the v1.4.22 train stays in place.
Snapshot builder: reads the prefs row BEFORE any measurement query
and intersects `excludeMetrics` with the resolved scope. The model
never sees data the user opted out of — that's the load-bearing
order: filter first, query second.
UI: settings cog in the drawer header opens a right-edge `<Sheet>`
with three Select pickers (tone, verbosity) and one Switch column
(per-metric exclusions + showEvidenceByDefault). Save calls the PUT
route and invalidates the `["coach-prefs"]` query so the next reply
honours the new defaults without a reload. EN + DE i18n keys for
every label and hint.
Message-thread reads the same `["coach-prefs"]` cache and toggles the
`open` attribute on the evidence `<details>` element when
`showEvidenceByDefault` is true.
Tests +24 unit (coach-prefs schema, system-prompt prefix, snapshot
mock for new prefs row). Integration round-trip seeds a session,
PUTs a non-default prefs shape, asserts the GET reflects the saved
values + the DB row carries the canonical defaulted form.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The W3-of-v1.4.22 sentinel parser silently dropped malformed rows and
fired `coach.keyvalues.parse_failed` only when the entire block was
unusable. Mixed-format drift (one row breaks, the other six parse)
landed without any signal. Surface it.
- `SentinelParseResult.malformedEntries[]` records each skipped row
with a typed reason (`missing_colon` / `value_overflow` /
`label_overflow` / `no_END_marker` / `byte_overflow` /
`schema_invalid`). Block-level failures (truncation, missing close
marker) land here as synthetic entries so the wide-event always
carries a typed cause.
- Per-row label and value caps moved out of the Zod schema into the
parser so the row-level reason can be `label_overflow` /
`value_overflow` rather than the catch-all `schema_invalid`.
- Chat route splits the annotation: `coach.keyvalues.parse_partial`
when at least one row parsed AND at least one row failed, the
existing `coach.keyvalues.parse_failed` stays for full-block
failures. Both annotations carry the per-line reason array.
- Backwards-compatible `parseKeyValueLine` wrapper preserved.
Tests +5: label_overflow, value_overflow, byte_overflow,
mixed-malformed entry assertions, and a tryParseKeyValueLine
direct-API surface check.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-1.4.23, a replay attempt on a single device's stale refresh token
revoked every refresh token the user owned — signing them out of every
device. The legitimate two-device case (iPad + iPhone, one drops into
airplane mode and rotates) hit the same code path. v1.4.23 scopes the
blast radius to rows matching the offending refresh row's deviceId so
only the suspect device is signed out.
Legacy tokens issued before v1.4.23 carry deviceId=null and fall back
to the wider revoke — a safety hatch we keep until the last 60-day
refresh token from the old policy expires.
The native-client routes already forward X-Device-Id on every refresh,
so the legitimate path always carries a deviceId. A missing deviceId
on a replay is itself a suspicious signal, hence the conservative
fall-through.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends the strict insight schema, Coach snapshot pipeline, and
provenance envelope with the seven Apple Health categories that landed
in Wave 2 (HRV, sleep, resting HR, steps, active energy, flights,
distance, VO2 max, body temperature). All fields are additive and
optional — web-only accounts pay zero extra SQL and the prompt's new
GROUND RULE 12 instructs the model to stay silent when the snapshot
doesn't carry the HealthKit blocks.
Ratchets PROMPT_VERSION 4.22.0 → 4.23.0 and refreshes the
sourceMetric / trendAnnotations enums in both EN + DE. Two new test
files pin the prompt body and the snapshot fan-out.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wires the v1.4.23 baseline route table — auth/login, passkey verify,
refresh, measurements GET + POST + batch, devices POST, insights
comprehensive — into the new generator. Schemas reuse the existing
validation files via Zod v4's `.meta()`, so the iOS DTO contract stays
single-source-of-truth.
The freshly emitted spec covers ~880 lines (down from the 5468-line
hand-maintained legacy preserved in `openapi-v1422-legacy.yaml`).
Future routes get added incrementally via `src/lib/openapi/routes.ts`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wires the OpenAPI 3.1 generator scaffolding ahead of the iOS-touched
route registrations. `zod-openapi` (samchungy variant, picked in W1
stream-3) reads Zod v4's native `.meta()` so schemas can stay in their
original validation files; `yaml@2` with `sortMapEntries: true` keeps
the emitted spec stable across re-runs.
`pnpm openapi:generate` overwrites `docs/api/openapi.yaml` from the
registry; `pnpm openapi:check` exits non-zero on drift. The legacy
hand-maintained spec is preserved at `docs/api/openapi-v1422-legacy.yaml`
so the iOS contract reference doesn't disappear during the
incremental migration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
23 unit tests under `senders/__tests__/apns.test.ts` exercise:
* env-var loader — all-or-none guard, inline-PEM `\n` unescape, file
fallback path, cache, APNS_PRODUCTION override
* Provider lazy-init + per-gateway cache (sandbox vs production stay
isolated; APNS_PRODUCTION funnels every send through production)
* permanent-failure detection (`Unregistered`, `BadDeviceToken`,
`DeviceTokenNotForTopic` flip `shouldDisable: true`; transient
reasons stay false)
* `sendViaApns()` dispatcher entry — fan-out, dead-row cleanup,
hard-vs-soft channel-reject classification, HTML stripping
3 integration tests under `tests/integration/apns-dispatch.test.ts`
drive the real dispatcher against a Postgres testcontainer with a
stubbed `@parse/node-apn` Provider:
* Round-trip: notification fans out via APNs and the channel-state
machine records `lastSuccessAt` + clears any cooldown
* Failure path: `Unregistered` deletes the dead Device row,
auto-disables the channel with the captured reason, and writes the
`notification.channel.auto_disabled` audit-log entry
* Cascade: a disabled APNS channel falls through to Telegram on the
next dispatch — the apns send-mock stays untouched
Adds `senders/__mocks__/apns.ts` — a deterministic mock with queued
per-token responses + recorded calls so future test files can swap the
real sender for a fixture-driven double without touching `@parse/node-apn`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wires `sendViaApns()` into `dispatcher.sendToChannel()` as the fourth
channel-type alongside Telegram, ntfy, and Web Push. Per-user channel
selection works exactly as before — a user with `APNS` enabled in
`NotificationChannel` and no preference override receives APNs alerts;
a user who pairs an iPhone but disables APNs in Settings doesn't.
Cascade order is now deterministic: APNs → Telegram → ntfy → Web Push.
The order matters because each channel runs best-effort and the user
sees notifications fan out in that sequence; sorting in `dispatchNotification`
keeps the delivery sequence stable across dispatches even when Postgres
returns the channel rows in physical order. Unknown channel types sort
last so a stale enum value can't preempt a real channel.
The `CHANNEL_TYPES` tuple, `CHANNEL_TYPE_LABELS` map, and
`ChannelConfig` union pick up `APNS` / `Apple Push (APNs)` /
`ApnsChannelConfig` (an empty record by design — the per-device APNs
token + environment live on the `Device` row, not on the channel
config). On `shouldDisable: true` from the sender, the dispatcher's
existing channel-state machine flips `disabledReason` and bumps
`consecutiveFailures` through `recordChannelHardReject()` — no new
state-machine code needed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds `src/lib/notifications/senders/apns.ts` so the dispatcher can fan
notifications out to a user's iPhone via Apple's push gateway. The
sender exposes two surfaces:
* `sendApnsPush({ deviceToken, environment, payload })` for one-shot
delivery — the helper the iOS-targeted reminder pipeline calls
directly.
* `sendViaApns(userId, payload)` mirrors `sendViaWebPush(userId,
payload)` so the dispatcher can route APNS through the same
`SendOutcome` contract: at-least-one-success → ok, every device
permanently rejected → hard channel reject + auto-disable, anything
else → soft failure that feeds the existing exponential-backoff
schedule.
Provider lifecycle is one `apn.Provider` per gateway (`sandbox` vs
`production`), lazy-initialised so the JWT-signing-key parse doesn't
happen at boot. `@parse/node-apn` regenerates the bearer JWT every
~50 minutes inside the Provider, so the singleton is safe — the
explicit `resetApnsForTesting()` helper exists so unit tests can
re-read env vars between cases.
Permanent failure detection covers `Unregistered`, `BadDeviceToken`,
and `DeviceTokenNotForTopic`; the dispatcher branch deletes the dead
Device row before returning, mirroring the web-push 410 cleanup. The
env-var loader is all-or-none with a one-shot warning when the
operator started configuring and stopped — half-finished Coolify env
blocks surface in the logs the moment they matter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New `src/lib/measurements/apple-health-mapping.ts` translates inbound
HealthKit samples into the row shape the `Measurement` model expects.
Symmetrical to the Withings `MEASURE_TYPE_MAP` but keyed by HealthKit
string identifier so the upcoming batch endpoint can do a single
table lookup per entry.
Each mapping carries its canonical DB unit, a unit-conversion helper
(identity for SI-aligned types, ×100 for HealthKit's 0..1 fraction
quantities), an aggregation hint for downstream summarisation, and
flags the explicit-consent metrics (HRV, VO2 Max, sleep) so the audit
trail can pick them out.
Sleep is the one category type — the helper returns one row per stage
with the new `sleepStage` column populated; per-stage minutes go
straight in as `value`. Unknown identifiers return `null` so the
batch endpoint can mark the entry as skipped instead of failing the
whole batch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extend MeasurementType with the seven HealthKit-shaped values the iOS
batch ingest needs (HRV, resting HR, active energy, flights climbed,
distance, VO2 max, body temperature) and add APPLE_HEALTH to the
MeasurementSource enum. Sleep duration unit shifts from hours to
minutes so per-stage HKCategorySample rows persist without precision
loss; the new SleepStage enum mirrors the existing GlucoseContext
column shape and a CHECK constraint scopes it to SLEEP_DURATION rows.
The composite unique index `(user_id, type, source, external_id)`
gives the new batch endpoint an idempotent dedup key while the legacy
`(user_id, type, measured_at, source)` index continues to back the
manual UI's no-duplicate-at-the-same-wall-clock guard. NULL externalId
rows stay distinct under Postgres semantics so manual entries never
collide.
Migration is strictly additive: new enum values append, new columns
are nullable, no rows mutate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Six MEDs grouped per the W5 reconcile brief:
- Sec-MED-1: `hl_onboarding` cookie now `SameSite=Strict` (no
cross-site flow depends on it; sibling `healthlog_session` stays
`Lax` for the Withings OAuth callback).
- Sec-MED-2: `/onboarding` PUBLIC_PATHS match tightened from loose
`startsWith("/onboarding")` to exact path + subroute guard
(`pathname === "/onboarding" || pathname.startsWith("/onboarding/")`).
A future `/onboarding-export` route can no longer inherit the
silent-pass-through. Same guard mirrored on the redirect short-
circuit. Two new proxy tests pin the contract.
- Code-MED-3: lift `berlinDayKey()` from
`src/app/api/analytics/route.ts` to a shared
`src/lib/analytics/berlin-day.ts` and route the targets
sparkline-bucketing through it. The targets route used to bucket
by UTC `YYYY-MM-DD` so a 23:30-Berlin reading on Tuesday landed
in Wednesday's bucket. Four new unit tests pin DST + UTC-midnight
edge cases.
- Code-MED (streaming race): on slow connections SSE `done` fires
before the invalidate-refetch resolves, producing a 200-500ms
duplicate-bubble render. Add a 150ms grace window on the
message-thread that suppresses the persisted twin while the
streaming bubble is still in flight.
- Code-MED-2: drop dead i18n keys `insights.coach.settings` and
`insights.coach.settingsTooltip` (W3 removed the cog; the keys
had zero call sites and would carry confusing legacy if v1.4.23's
prompt-tuning panel picks a different key shape).
- Code-LOW-5: centralise `["user", "dashboardWidgets"]` query key
in `queryKeys.dashboardWidgets()` and migrate the three call
sites (dashboard, insights, settings/dashboard-layout-section).
Same query-key-collision discipline as `queryKeys.analytics()`.
Closes Sec-MED-1, Sec-MED-2, Code-MED-2, Code-MED-3, Code-MED-3
(streaming race), Code-LOW-5.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Coach chat route had three identical `new ReadableStream({
start(controller) { try { ... } finally { controller.close(); } } })`
blocks (refusal / provider-error / happy path), differing only in
which frames they enqueued. Extract the boilerplate into
`src/lib/sse/create-stream.ts` so the v1.5 iOS streaming endpoints
inherit the same try/finally close semantics without copy-paste.
The helper accepts sync or async emit callbacks; the controller is
always closed in `finally` so a thrown emit never hangs the response.
Three unit tests pin the sync, async, and throw paths.
Closes Sr-H2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The `setOnboardingPendingCookie` helper was called from six different
auth routes (login, passkey-verify, register, password, /me,
/onboarding/complete) by convention. A future auth surface (SSO,
magic link, social login) added without remembering the helper would
silently re-introduce the dashboard flash that v1.4.22 C4 was
designed to fix.
Make `onboardingPending` a required parameter on `createSession`
(not optional) so issuing a session without anchoring the cookie is
type-impossible. The five primary auth surfaces now thread the
user's `onboardingCompletedAt == null` value through; the
`/api/auth/me` mirror remains as a fall-back for legacy sessions
that predate the cookie. The `/api/onboarding/complete` clear stays
where it is — it is a value flip after a DB write, not a session
issuance.
Three new integration tests pin: createSession sets the cookie when
pending, clears it when completed, and destroySession clears it
alongside the session cookie.
Closes Sr-H1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The tile computed `bp30 - bpAll` regardless of which baseline the
user picked under Settings → Dashboard, while rendering a caption
keyed on `comparisonBaseline` ("vs. last month" / "vs. last year").
The math therefore lied to the user. Every other trend tile uses
`tileCompareDelta()` against period-aligned `avg30LastMonth` /
`avg30LastYear` baselines.
Add `priorMonth` (now-60d…now-30d) and `priorYear` (now-395d…
now-365d) windows to `computeBpInTargetWindows()`. Ship the pcts on
the analytics envelope as `bpInTargetPctPriorMonth` /
`bpInTargetPctPriorYear`. Route the dashboard's `bpCompareDelta` to
subtract the prior window matching the user's baseline; fall back
to null for `none`.
Three new bp-in-target unit tests (priorMonth, priorYear, null
windows) and two insights-polish guards (envelope contains the
fields, page no longer subtracts `bpAll`).
Closes Code-H2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Maintainer-voice sweep across `src/`. ~190 references to the
maintainer by first name in code comments and JSDoc neutralised to
"the maintainer", "the live tenant", "users", or "the user" depending
on context. Three sed passes covering ~50 distinct patterns
(`Marc reported`, `Marc wants`, `Marc kept`, `Marc rolled back`, etc.).
Test fixtures (`displayName: "Marc B."`, `username: "marc"`,
`<HeroStrip userName="Marc">`) kept as opaque test data per the brief.
CHANGELOG normalisation. The v1.4.14 + v1.4.15 entries paired a
German bullet with its English translation by historical design — the
in-app copy is English-default and the user docs are English, so the
mixed-language style was a holdover from the early-marathon notes.
A one-off node script extracted the `_..._` English paragraph from
each bilingual bullet and rewrote the entry as that text. ~44 bullets
reduced; one Docs bullet that lost a leading word in the regex
fixed by hand.
CLAUDE.md → CONTRIBUTING-AI.md rename so the filename is no longer
AI-vendor-specific. `CONTRIBUTING.md` reference updated. `AGENTS.md`
stays in place for multi-agent compatibility (the conventions doc is
shared with Codex / Cursor / etc.).
Wave-4 state wrap. STATE.md now records C1–C5 + D + D2 as complete;
v1421-backlog.md grew a "Closed in v1.4.22" header listing the 20
items the marathon landed; `.planning/phase-W4-report.md` documents
the per-commit decisions, test deltas, and the 23 items deferred to
v1.4.23 with reasoning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The post-hydration redirect for users with `onboardingCompletedAt ===
null` caused a brief dashboard flash before the client-side effect
could fire. The redirect now lives in `src/proxy.ts` so it lands on the
first server response.
The proxy runs in the Edge runtime and can't reach Prisma, so the auth
routes (login / passkey-verify / register / password / /api/auth/me)
mirror the DB `onboardingCompletedAt` column into a non-httpOnly
`hl_onboarding` cookie. The cookie is a UX hint, not a security
signal — a user editing it locally only skips the dashboard flash;
they still can't bypass any server check. `/api/onboarding/complete`
clears the cookie so the next navigation drops the redirect
immediately, and `destroySession` + expired-session purge wipe it
alongside the session cookie.
The e2e onboarding-flicker spec can return to testing the actual
incomplete-onboarding case: the proxy reads the cookie (absent for
the fully-onboarded storage-state user), not the mocked `/api/auth/me`
response, so the test now sets `onboardingCompletedAt: null` honestly
instead of pinning it to a non-null value as a workaround.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a text-marker sentinel parser (`---KEYVALUES---` … `---END---`)
that strips the Coach's evidence block out of the prose before it
streams to the client, validates each line against a Zod schema, and
attaches the parsed entries to `provenance.keyValues`. The persisted
envelope round-trips so the disclosure renders on conversation
reload. Hard-capped at 8 lines and 1 KB to guard against
prompt-injection padding; malformed blocks log a wide-event and pass
the prose through unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the v1.4.20 "conversational summariser" persona with a partner
who sits alongside the user — warm, curious, conservative, with seven
numbered ground rules that prefer invitation over refusal when data is
missing and reserve specific numbers for a structured evidence block
at the end of each reply. The day-level timeline section from v1.4.21
stays verbatim so weekday questions still resolve cleanly. Bumps the
shared PROMPT_VERSION from 4.20.2 to 4.22.0 and broadens existing
schema tests off the 4.20.x pin onto the 4.x train.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two related leak fixes traced to the same root: every other AI-prose
surface ran user-visible text through `stripChartTokens()`, but the
recommendation card's always-visible text was rendered directly. The
model embeds `metric:WEIGHT` and friends in the rec text expecting an
inline mini-chart, but the chart slot lives in the EXPANDED rationale
block — so the token had nowhere to be consumed and leaked verbatim
to the DOM.
- `<RecommendationCard>` now wraps `norm.text` in
`stripChartTokens()`, matching the contract every other
rec / finding / summary surface already uses.
- DE Health-Score component labels surfaced English nouns ("Mood",
"Medis", "BD") instead of full German. Rename to "Blutdruck",
"Stimmung", "Einnahmetreue", "Gewicht" so the four sub-bars read
natively German. The "Mut" the maintainer reported on prod was
voice-to-text rendering "Mood".
- i18n-integrity test pins the four component labels so a copy-paste
regression can't reintroduce the leak.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New `.planning/phase-v1421-hotfix-report.md` summarises the six
fixes (Daily Briefing regenerate · Coach day-level snapshot · Coach
prompt update · streaming-bubble dedup · drawer header · sources rail
scope picker), the test-count delta (2026 → 2036 unit), the
token-budget impact of the day-level snapshot (~190 → ~3000 tokens
per full-scope turn), and a recommendation to ship as a v1.4.20.1
hotfix.
STATE.md grows a Phase F section + status block tracking the six
commits on develop. Includes prettier polish across the touched
coach surfaces (sources-rail, drawer, message-thread, snapshot,
chat route).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously the Sources rail listed BP / Weight / Pulse / Mood /
Compliance as a static legend. Marc asked for a real "what I can
see" surface — per-source checkboxes (include/exclude) plus a
window selector (last 7d / 30d / 90d / all-time) that flow through
the Coach request body.
Wire the controls all the way down:
- <SourcesRail> grows scope + onScopeChange props. Each row
becomes a 36px tall flex row with a real checkbox; the row
fades to 60% opacity when its source is excluded.
- A Select trigger above the list drives the window. Both control
surfaces share the existing insights.coach.window.* i18n keys;
add a new windowLabel key so the dropdown has an EN+DE label.
- <CoachDrawer> owns the scope state (resets on each mount per
the v1.4.20.1 plan — no conversation-level persistence yet) and
forwards the picked scope through useSendCoachMessage to the
chat request body. The payload only ships when the user has
narrowed away from the all-source last30days default so the
server can tell "no opinion" from "intentionally narrow".
- buildCoachSnapshot already accepts the optional CoachScope from
the previous commit; it now flows end-to-end.
No new dependencies — checkboxes are plain inputs styled with
Tailwind + Dracula tokens.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous Coach prompt only said "ground every number in the
SNAPSHOT" without telling the model that the snapshot now carries a
timeline.recent block of day-level rows with weekday labels.
Without that pointer the model kept defaulting to aggregate-only
phrasing on day-specific questions.
Add a DAY-LEVEL READINGS section to both EN and DE prompts:
- explain the timeline.recent / timeline.weekly shape
- tell the model to answer day-specific or weekday-specific
questions out of timeline.recent and to cite the actual reading
with date + weekday
- when no row exists for the day named, say so plainly and offer
the closest available day rather than the aggregate
- older weekday questions fall back to timeline.weekly
- the scope.sources field is the conversation's source-allowlist;
a question about a non-listed metric should be acknowledged as
out-of-scope rather than answered from the aggregate
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Coach snapshot used to ship only aggregated statistics (mean,
slope, SD, range, count) per metric. A turn that asked
"why was BP higher on Monday?" had no day-level rows to ground in
and the model fell back to apologetic aggregate-only phrasing
("from the snapshot I cannot isolate the specific Monday").
Extend buildCoachSnapshot to fold in:
- timeline.recent: one row per UTC day for the last 14 days, each
tagged with its weekday so the model can answer day or weekday
questions directly. BP rows pair systolic + diastolic; days with
only one side measured are dropped to avoid fabricating a
complement.
- timeline.weekly: ISO-week buckets covering days 15..N where N is
the configured window. Keeps the prompt budget tight on 30/90-day
windows while still letting the Coach cite older weeks.
- per-day medication adherence rows from MedicationIntakeEvent so
the Coach can answer "did I miss my dose Tuesday?" without
inventing the schedule.
Also add the optional `scope` field to coachChatRequestSchema (sources
+ window) so the upcoming sources-rail picker can narrow the snapshot
without breaking older clients — missing fields fall through to
all-source last30days.
A typical 30-day, 5-source snapshot lands around 2.4-2.8 KB JSON
(≈700-900 prompt tokens) — comfortably inside the 25k/day Coach token
budget.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- weekly-report eyebrow chip drops uppercase + tracking-wide for the
sentence-case rule the rest of the app follows; print stylesheet
now neutralises bg + colour so monochrome output reads clean
(Design-M4, Design-L3).
- correlation-card "Try a 7-day experiment" CTA flips from default to
outline while disabled. Disabled-primary read as "the system is
broken right now"; outline reads as "this is on the roadmap"
(Design-M7).
- coach-drawer settings cog aria-label switches to the coming-soon
tooltip copy so SR users hear the same "arrives in v1.4.21" message
sighted users see (Design-M9).
- sources-rail static "fresh" dot becomes aria-hidden — v1.4.20 has
no real freshness state behind it; SR users no longer hear "Fresh"
five times in a row (Design-L4).
- summariseTitle slices on Unicode code points rather than UTF-16
code units so a user message ending mid-emoji ("🩺") no longer
produces a half-character on disk (Code-LOW-04).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three live byte-identical formatRelativeTime copies (hero-strip,
daily-briefing, coach history-rail) lifted to src/lib/i18n/relative-
time.ts. Both correlation-card and trend-annotation duplicated the
exact same low/moderate/high Dracula classnames; extracted to
src/components/insights/confidence-badge.ts (per-component i18n
labels stay local — the copy differs).
Drop the dead "null" placeholder branch in useCoachConversation's
queryKey: the ternary picked a different key when id was null, but
enabled: id !== null kept the query disabled so the placeholder never
fired. Key directly on id.
Hoist STORYBOARD_COLOR_BY_CATEGORY in /insights/page.tsx to module
scope so the map isn't recreated on every render.
Fixes Simplify-S-02, Simplify-S-03, Simplify-S-04, Simplify-S-05,
Code-LOW-02.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Every POST /api/insights/chat called extractFeatures(userId, false),
which ran prisma.measurement.findMany({ where: { userId } }) with no
date filter — fetching every measurement the user had ever recorded.
For power users with multi-year Withings imports that's tens of
thousands of rows on every Coach turn.
Add an optional sinceDays bound to extractFeatures (default unchanged
= full history; analytics dashboard + insights/generate keep their
existing behaviour) and have the Coach snapshot pass sinceDays: 90.
The Coach already only narrates the last7 / last30 / last90 windows,
so the 90-day floor preserves every metric the snapshot surfaces.
Fixes Senior-HIGH-3.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Composite 0-100 score with three bands (green >= 75, yellow 50..74,
red < 50). Server-deterministic so the same inputs always produce the
same output.
Sub-components:
- bpInTargetRate (30 %) — reuses the existing all-time figure
- weightTrendAlignment (20 %) — slope vs target band over 30 days
- moodStability (20 %) — `100 - CV*100`, requires >= 5 entries
- complianceRate (30 %) — mean of medication compliance30
Null components scale the remaining weights proportionally so the
score never lies about what was actually measured. Pure helpers
(linearRegressionSlope, coefficientOfVariation, weightTrendAlignment,
moodStability, complianceRate) live alongside the composite.
Tests cover strong/mixed/poor scenarios, every null-redistribution
shape (bp null, bp+mood null, all-but-compliance null, all null) and
determinism (same input -> same output).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a newsletter-style report route at /insights/report/[week] that
parses the YYYY-Www slug, reads the cached weeklyReport block off the
AI advisor payload, and renders Summary / Going-well / Worth-watching /
Tips / Data-quality sections in a print-friendly layout.
Print export uses window.print() with Tailwind print: variants on the
toolbar (hidden) and article wrapper (full-width margins). The route
honours ?print=1 to deep-link the export PDF flow from the upcoming
hero banner card.
Includes:
- src/lib/insights/week-iso.ts — parseWeekISO / toWeekISO / weekISOToRange
- WeeklyReportView wrapper + WeeklyReportPresentation pure layer
- Empty-state with Generate CTA when the cached report's weekISO
doesn't match the route param
- EN + DE i18n keys under insights.report.* and insights.heroBanner.*
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extend aiInsightResponseSchema with weeklyReport (ISO week + summary +
goingWell/worthWatching/tips bullets + optional dataQualityNotes) and
storyboardAnnotations (date + label + category + detail; max 20).
Both blocks are nullable + optional so legacy 4.20.1 caches round-trip.
Bump PROMPT_VERSION 4.20.1 -> 4.20.2 and add GROUND RULE 10 (weekly
report sections match layout, conservative phrasing, no causal claims)
+ GROUND RULE 11 (storyboard annotations are factual user-logged events
with neutral prose) to both EN and DE prompts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extend `aiInsightResponseSchema` with an optional, nullable
`trendAnnotations: { bp?, weight?, mood? }` block — one short
sentence per metric, capped at 200 characters. Each field is
independently optional so the model emits only the metrics it can
ground in the snapshot.
Insight-generator prompt picks up GROUND RULE 9 (EN + DE):
single-sentence, observational, ≤ 200 chars, banned causal
phrases enumerated. PROMPT_VERSION bumps 4.20.0 → 4.20.1 to
signal the format change so cached payloads from B1 still parse
through the `.nullable().optional()` field.
13 tests cover the boundary (200 chars), the empty-string
rejection, the legacy-payload round-trip, and the EN/DE prompt
rule presence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add `src/lib/insights/correlations.ts` with three pre-defined
hypothesis runners (BP × medication compliance, mood × resting
pulse, weight × weekday) plus pure Pearson + one-way weekday-ANOVA
helpers. Each runner gates the surface on n >= 14 paired rows AND
p < 0.05; below the bar a typed `insufficient` result lets the
card render an EmptyState rather than a misleading pattern.
Interpretations stay conservative ("a pattern worth watching" —
never "X causes Y"). 95 % confidence intervals come from a Fisher
z-transform on r and a discrete band-from-eta-squared mapping for
the weekday case. No new dependencies — the t-distribution +
F-statistic p-value approximations are inlined (Wilson-Hilferty +
Abramowitz-Stegun).
16 unit tests cover positive / negative / no-correlation / under-n /
non-significant / no-variance branches and the Monday-outlier
weekday flag.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The route gates on requireAuth (cookie or bearer), enforces the daily
token budget, runs the prompt-injection / off-topic refusal scanner,
and only then walks the user's provider chain.
The conversation is created via withIdempotency() when no
conversationId is supplied, so a retry under the same Idempotency-Key
replays the original assistant reply rather than spinning up a second
thread. Existing-thread turns skip the wrap.
Replies stream as SSE: one "token" frame per word-sized chunk, then a
"provenance" frame describing the windows + metrics the assistant
could see, then a "done" frame carrying the persisted message id.
Provider failure cascades to a single "error" frame and the assistant
message is never persisted. Conversation history is capped at 20 turns
in the prompt window via a synthetic summary placeholder so cost stays
bounded.
`buildCoachSnapshot()` reuses the analytics features pipeline so the
Coach narrates the same numbers the dashboard tiles render.
`getCoachSystemPrompt()` extends the strict-insight ground rules with
conversational framing while keeping the zero-hallucination rule.
5 integration tests pin the round-trip, refusal short-circuit, budget
429, foreign-conversation 404, and validation 422 paths against the
postgres testcontainer.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wire the foundation layer for the AI Coach:
- `types.ts` defines the SSE event shapes (token / provenance / done /
error), the chat-request zod schema, and the conversation/message
DTOs the route + UI share.
- `persistence.ts` encrypts every message body with the active
ENCRYPTION_KEYS entry before insert and decrypts on read.
`summariseTitle()` keeps the rail snippet ≤80 chars with a smart
word-boundary cut.
- `budget.ts` introduces a per-(user, UTC-day) ledger with a 25 000-
token daily cap. `enforceBudget()` throws HttpError(429,
"coach.budget.exceeded") so the route can short-circuit before
spending. UTC alignment matches the LLM-provider billing cycle.
- `refusal.ts` runs prompt-injection and off-topic detection on inbound
messages. Pattern bank covers EN+DE injection wording, jailbreak
variants, and obvious off-topic asks (weather, code-help, recipes).
Health-allow tokens override the off-topic deny list so mixed
questions land on-topic.
- i18n keys `insights.coach.refusal*` + `insights.coach.error*`
added to both locale bundles.
- 28 unit tests cover the helpers with mocked Prisma.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends `aiInsightResponseSchema` with an optional `dailyBriefing`
block (paragraph + 0-5 keyFindings, each with tone / headline / detail
/ delta / sourceWindow / sourceMetric). Nullable + optional so v1.4.19
cached payloads still parse cleanly.
Bumps PROMPT_VERSION 4.19.0 → 4.20.0 and adds GROUND RULE 8 in both
EN and DE system prompts, instructing the model to emit a briefing
block grounded in the user's own snapshot numbers (no extrapolation,
no medical advice claims, hard-cap at 5 findings).
TDD: new daily-briefing-schema.test.ts covers legacy back-compat,
well-formed parsing, > 5 findings rejection, empty-paragraph
rejection, finding-field validation, and the PROMPT_VERSION 4.20.x
match. Existing medical-reference-prompt.test.ts assertion stays
broad enough to ride the bump.
+31 tests (1672 → 1703); typecheck clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Run prettier on the test, e2e, and component files this v1.4.19
A8 wave touched. The .planning/* and docs/audit/* prettier
warnings are pre-existing baseline noise — left alone to keep
this commit narrow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The over-intake and skipped-intake achievements shipped with the
literal titles "Idiot" / "Lazy Boy" in both EN and DE bundles —
the German file never translated them. For a personal health app
the user gets called an idiot any time they double-dose because
they forgot they already took it. Replaced with neutral,
descriptive titles ("Double take" / "Stepped back" — "Doppelt
genommen" / "Pause eingelegt"). Locked in by a regression suite
that scans every badge title against a banned-word list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`formatTimeWindowRange` used to hard-code the German "bis ... Uhr"
suffix, so an English-locale user saw "Today, 19:00 bis 23:00 Uhr"
on the medication card and the medication form. Pass the active
locale through and emit "19:00 – 23:00" for English while keeping
the existing German output for DE users (and unchanged for legacy
call sites that still don't pass a locale).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Marc reported the Zielwerte page (`/targets`) still surfaces English
status pills even when the rest of the UI runs in German: "Low /
On Target / Stable / Moderate / Normal" stayed in source language
because the server emits stable English category strings and the
page rendered them verbatim. Card titles ("Resting pulse", "Mood
stability", "Medication compliance" …) had the same issue.
Add two i18n key blocks under `targets.*`:
- `targets.label.<TYPE>` — one entry per measurement type, used by
`<TargetCard>` to localise the card title. The server keeps
emitting English labels (so logs / external consumers don't
churn); the page swaps them in via `t("targets.label." + target.type)`.
- `targets.status.<key>` — every classification category from
`lib/analytics/classifications.ts`, `pulse-targets.ts`, and the
inline maps in `targets/route.ts`. A small `STATUS_CATEGORY_KEY`
map on the page normalises the human-readable English category
to the i18n key.
Glucose targets keep their existing key path (the server already
emits `targets.glucoseFasting` etc. as the label, resolved at the
top of `<TargetsPage>` before cards mount).
Fall-through is graceful: if a future server adds a category we
have not translated yet, the verbatim English string still
renders — the page never blows up.
Also adds `targets.bmi`, `targets.label.BMI`, `targets.status.normal`,
`targets.status.optimal`, `targets.status.fitness` to the i18n
integrity test's PLACEHOLDER_ALLOWLIST: those terms are identical
in EN and DE (BMI is a global acronym; "Normal" / "Optimal" /
"Fitness" are technical terms German clinics use verbatim).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Marc reported a raw "metric: blood_pressure_sweet" string at the
bottom of /insights — looked like a debug code or unrendered
template. Root cause: the v1.4.16 strip regex
`/metric:[A-Z_]+/g` only matches uppercase letters and underscores,
but the model occasionally emits lowercase / snake_case forms like
`metric:blood_pressure` or `metric:weight_trend`. Those slipped
through `stripChartTokens()` unchanged and the literal token text
landed in the DOM.
Split the regex in two:
- `STRIP_TOKEN_REGEX` (`[A-Za-z0-9_]+`) — permissive; cleaves any
case from prose so unrenderable / hallucinated tokens never
surface as text.
- `PARSE_TOKEN_REGEX` (`[A-Z_]+`) — strict; the render path still
only mounts charts for the uppercase allowlisted vocabulary.
Allowlist semantics are unchanged — `parseChartTokens()` still
returns only entries from `ALLOWED_CHART_TOKENS`. The change is
purely about defending the visible-prose surface from the wider
class of token shapes the LLM occasionally emits.
XSS-style trailing junk (`metric:WEIGHT' onclick='alert(1)'`)
still cleaves cleanly: the apostrophe terminates either class.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pixel-5 audit against PROD (393 px) showed two issues with the
chart cards on /dashboard and /insights:
1. The bucket-aggregation chip ("Wochendurchschnitt" / "Weekly avg"),
the optional comparison-period chip, the four range tabs and the
cog dropdown all sat on a single justify-between row in the chart
header. On viewports below the sm breakpoint the chips ate enough
horizontal space that the tabs wrapped to a second row inside the
same flex container, doubling the header height. On Galaxy Fold
compact (280 px) the tabs split across three or four rows and the
medication card overflowed horizontally.
2. The medication compliance chart drew one tick per data point on
the x-axis. A 30-day Pixel-5 window painted thirty overlapping
labels into a 393 px gutter while the weight / BMI charts looked
sparser by accident, so the chart row read inconsistently.
The header is now stacked into two rows below the sm breakpoint:
title + chips on row 1, range tabs + cog right-aligned on row 2 with
flex-nowrap so the tabs always fit. The decorative bucket-aggregation
chip and comparison captions hide on mobile (the range tabs already
communicate the visible window). Above sm the original side-by-side
layout is preserved.
A new helper at src/lib/charts/x-axis-density.ts caps visible ticks
at 4 (≤360 px Fold), 6 (≤480 px Pixel-5 / iPhone-12), 8 (small tablet)
and 10 (desktop) and is wired into HealthChart, MoodChart,
MedicationComplianceChart and ComplianceLineChart so every chart row
reads the same density. A new useViewportWidth hook keeps it reactive
without an SSR penalty (desktop default during server render, real
measurement after mount).
The scatter-correlation chart was deliberately left alone — its
x-axis is numeric (BP / weight values, not a time series) and already
supplies an explicit `ticks` array.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Marc reported the BD-Zielbereich tile pinned at exactly 50/50/50 for
7T, 30T, and the headline ("total") on his real production data — the
fourth attempt at this surface (v1.4.15 → v1.4.16 → v1.4.18 → v1.4.19).
Root cause: the analytics route routed `bpInTargetPct` (the headline)
through `windows.last30Days?.pct` — a literal copy of the `30T`
sub-value. Since `computeBpInTargetWindows` only returned 7d + 30d
windows, the headline could never legitimately differ from `30T`. For
Marc's real data (572 paired readings since 2022, recent 30d = 50 %,
all-time ≈ 11 %) the tile read 50 % / 50 % / 50 % and looked
algorithmically pinned — exactly the symptom he reported.
The previous three passes treated this as either a predicate bug
(v1.4.16 ceiling-vs-narrow-band, real and merged) or a missing
`avg7`/`avg30` wiring (v1.4.18, also real and merged). Both fixes
landed correctly, but neither addressed the headline duplication.
Fix:
- `computeBpInTargetWindows` now returns a third `allTime` window
that aggregates EVERY paired reading in the input series.
- The analytics route fetches all paired BP rows (no 30-day cutoff)
and routes the headline through `windows.allTime?.pct`. The route
only re-pairs once internally so the cost is paid once per request.
- The 7T / 30T sub-values are unchanged (still windowed slices of
the same input).
Failing test reproduces the bug with Marc's data shape (`allTime` MUST
not equal `last30Days` once older history is present); fix turns it
green. Integration test seeds 40 paired readings against real Postgres
and asserts all three windows compute independently.
Hand-counted against Marc's prod data:
- 7T: 50 % (1/2) — matches what he sees, real coincidence
- 30T: 50 % (5/10) — matches what he sees, real coincidence
- Headline: ~11 % (62/572) — was 50 %, now correct
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Every insight summary opened with the same "Datengrundlage ist sehr
stark" / "Your data foundation is strong" filler — the user does
not see what fields the snapshot carries, so the opener reads as
awkward stock praise.
Add ground-rule 7 to both locales of the system prompt: don't open
with a compliment about the data quantity or quality, mention data
quality only when it materially limits the analysis (n<7 in the
analyzed window, recencyDays>14 since the last entry, or a coverage
gap). When data is fine, dive straight into the analysis.
Banned-opener phrases listed verbatim in both EN and DE so the
model does not paraphrase the same sentiment.
PROMPT_VERSION bumped 4.16.1 → 4.19.0 so feedback aggregation can
distinguish responses generated under the new rule.
New test file pins the rule, the trigger thresholds, the banned
phrases, and the dive-straight-in instruction in both locales.
Existing PROMPT_VERSION pins relaxed to allow any 4.x.x ratchet so
future minor bumps don't sweep up these assertions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-existing format drift accumulated across v1.4.x; running
prettier brings the whole tree back to baseline so future diffs
read clean. No behaviour change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
