Lay the data-model foundation for the Fitbit/Pixel integration over the
Google Health API, mirroring the WHOOP provider. No OAuth, sync, or UI yet.
- Prisma: FitbitConnection (1:1 per user, encrypted access/refresh tokens)
and FitbitOAuthState models, two encrypted BYO-credential columns on User,
and FITBIT appended to the MeasurementSource enum. Migration
0117_v1120_fitbit_integration is additive only — one enum value plus two
tables and two nullable columns, no new MeasurementType (every launch
metric maps onto an existing type).
- Wire the two new encrypted columns and the FitbitConnection token columns
into scripts/rotate-encryption-key.ts.
- FITBIT touchpoints: the read measurementSourceEnum (not the writable set),
the DEFAULT_SOURCE_PRIORITY and workout-source ladders, the sources-section
label map, the IntegrationKey union plus its display name, and a
fitbit()/fitbitStatus() query-key pair. The source-rank SQL and read-tier
collapse pick FITBIT up from the ladders with no code change.
- Add the FITBIT source label to all six locale bundles and regenerate the
OpenAPI contract so the source enum carries FITBIT for the iOS decoder.
- Mood discovery footer no longer pairs a mood-only numerator with the
full behaviour-by-outcome denominator. Reword it to surface the mood
relationships out of all day-to-day pairs the engine tested across
every metric, so the count it reports is honest about its scope.
Update all six locales.
- Rank the better-days board on a comparable scale: standardize the tag
mood-point delta into a Cohen's-d effect (delta over the pooled
with/without mood SD, clamped to [0,1]) so it sits alongside a metric
|r| instead of the old |delta|/2 heuristic. Fall back to |delta|/2 only
when the pooled SD is undefined. The pooled SD is ranking-only and is
stripped before the row enters the snapshot payload; the raw delta and
r the cards show are unchanged.
- Normalize the Withings and Whoop resume buttons to the min-h-11 scale
token, dropping the arbitrary min-h-[44px] spelling.
- Give the API token copy button a 44px tap target on small viewports,
scaling back to the tighter icon size from sm up.
- Drop the dead cached prop from InsightStatusCard and every caller; the
card has carried no cached affordance since the badge was removed.
Apple Health writes an unspecified ASLEEP aggregate alongside the granular
CORE/DEEP/REM partition for the same period. The asleep total already preferred
the granular partition, but the per-night/per-session `stages` map and the
hypnogram `segments` list recorded every stage row unconditionally, so a night
ended up carrying the granular set PLUS its ~equal bare twin — a double-height
green ASLEEP block on the stacked bar and the hypnogram, and a doubled total in
the analytics stage breakdown. Route the `stages` map and the `segments` list
through the same granular-over-bare gate as the asleep total: when a session
carries a granular stage, drop the redundant bare ASLEEP (and stage-less) rows
everywhere. WHOOP never emits a bare ASLEEP, so WHOOP nights are unaffected.
Apply the gate per session, not over the merged night pool, so a granular
overnight no longer strips a same-day bare-only nap's minutes from the night
total. Render the hypnogram per-segment in the stage palette (one ReferenceArea
per segment) instead of a single flat cyan line, and show the breakdown plus
totals footer whenever any per-stage minutes exist — reserving the unavailable
state for a truly empty session — so a legacy stage-summary night still surfaces
its breakdown. Thread the user's source priority into the CSV export's sleep
dedup, and honour offset with a true pre-slice night count on the sleep list.
GET /api/sleep/night?date=YYYY-MM-DD returns one reconstructed night for
the phase-progression view: the canonical source's stage segments (each
with an absolute start/end span), the per-stage breakdown, the asleep/
in-bed/awake totals, the mid-sleep awakenings count, and same-wake-day
naps surfaced separately. Session-clustered, wake-day-keyed, one
canonical source per night so two sources never overlay. date omitted
returns the most recent night. A view over existing per-stage rows — no
schema, no new measurement type. OpenAPI regenerated.
The CSV + JSON export emitted a wall of per-stage sleep minutes with no
night attribution. Collapse SLEEP_DURATION to one record per night by
default — the main session's TIME ASLEEP per the nap convention — with
the stage breakdown, nap count, and awakenings in the notes column so
the single uniform export reads as one night. `granularity=raw` keeps
the per-stage rows for power users. Storage is untouched.
Add reconstructSleepSessions — the session-level sibling to
reconstructSleepNights that keeps each sleep session separate (rather
than merging same-wake-day sessions under one key) and resolves every
stage row to its absolute span (start = measuredAt − duration), picks
one canonical source per session, and counts mid-sleep awakenings.
Add pickMainNightAndNaps to make the nap convention explicit: the main
night is the session with the most asleep minutes; every other session
on the same wake day is a nap, surfaced separately and never folded into
the headline.
These views are pure functions over existing per-stage rows — no schema
change, no new measurement type. They back the per-night list collapse,
the analytics reconciliation, and the hypnogram.
After the graceful collection-403 soft-skip, a genuine grant revoke 403s every
data class: each resource soft-skips (returns 0, records no failure), so the
orchestrator sees no error and stamps recordSyncSuccess — leaving the dead
connection looking healthy until the token-refresh path catches the 401 up to
~1 h later. Track the soft-skip count across the resource loop via an
AsyncLocalStorage tracker scoped to the cycle; when every resource soft-skipped
and nothing imported, leave the status as-is instead of stamping success. A
partial cycle (any import, or any resource that did not soft-skip) stamps as
normal.
A revoked refresh token returns HTTP 400 with an OAuth body of
`error: invalid_grant` from the token endpoint. The status-only classifier
buckets every 400 as persistent, so the connection never prompts a reconnect
even though the grant is gone. Store the upstream OAuth error on WhoopApiError
and lift a 400 invalid_grant to reauth_required in classifyWhoopError; every
other 400 (a malformed request, invalid_client) stays persistent.
Apple Health (and some WHOOP shapes) write BOTH an unspecified ASLEEP
aggregate row AND the granular CORE/DEEP/REM breakdown for the same
sleep period. The per-night reconstruction summed both, so a real ~8 h
night surfaced as ~16 h across every sleep surface — the chart showed
~16-18 h points and the target-range 30-day average read ~16.4 h/day.
Prefer the granular partition: when any CORE/DEEP/REM stage is present
for a night/source, count only those; fall back to the bare ASLEEP (or
a stage-less legacy/manual) row only when no granular stage exists.
IN_BED and AWAKE stay excluded; dual-source nights still collapse to one
canonical source via the user's sleep priority ladder. The same rule
applies inside the per-source asleep tiebreak so a source writing both
shapes is not over-weighted.
Align the surfaces that bypassed the shared helper:
- insights/targets: SLEEP_DURATION was averaged as raw per-stage rows
(minutes) and fed into a target band expressed in hours, producing the
bogus 16.4 h/day. Reconstruct per-night time-asleep, convert to hours,
and classify + roll up consistency on the nightly hour totals.
- coach snapshot: route the sleep block through reconstructSleepNights
so the prompt narrates the same nightly numbers (was summing every
stage row, including IN_BED + AWAKE).
- sleep card chart: read /api/measurements/series?kind=sleep (night
reconstruction, hours) instead of the generic measurement
daily-average path, and label the axis in hours instead of minutes.
Add regression fixtures: a night with both a bare ASLEEP row and the
granular breakdown asserts the granular total (no double-count); a
dual-source night with both shapes per source; and a multi-night 30-day
average that lands at ~8 h, not ~16 h.
Two changes that turn the descriptive relations surfaces into a
statistically-gated discovery view on the mood page.
Mood as an outcome: the FDR-controlled all-pairs correlation discovery
engine treated mood only as a behaviour (a lag source). Promote it to an
outcome channel too, so the scan also surfaces "behaviour today → next-day
mood" relations (e.g. more daylight today → better mood tomorrow) alongside
the existing "mood today → next-day outcome" direction. The self-pair is
skipped and Benjamini-Hochberg already controls the larger pair family this
opens up. The discovery route now feeds the shared mood daily-mean series
into both roles.
Surface the discovery slice: the engine was computed but never rendered.
The mood Insights page now shows the mood-involving pairs that survive the
false-discovery correction, each phrased observationally with its paired-day
count, r, and BH q-value, plus the honest "X of N tested" footer. The
surface self-fetches the correlation route and degrades to nothing when the
operator has disabled it. Observational only, with the standing disclaimer.
Two new observational surfaces on the mood Insights page, both computed in
the single mood-aggregates engine so the charts and the coach snapshot stay
in sync.
Influence on mood (per tag): for each frequent flat and structured tag,
compare the average daily mood on days the tag is present against the
counterfactual baseline of days it is absent, with the delta and a
confidence band. The band comes from a Welch two-sample t-test reusing the
exact Student-t p-value already behind the correlation cards, not a naive
approximation. Both groups must clear a five-day floor; rows rank by
absolute delta and cap per axis. Sparse tags, single-day tags,
all-same-value days, and zero-delta tags are dropped cleanly — never NaN.
Better days board: one ranked list folding the tag deltas and the
mood x health-metric correlations (sleep, steps, pulse, weight, blood
pressure) into a single effect-size-ranked, confidence-gated surface — the
headline answer to "what goes with my good days?". Near-zero and
under-sampled correlations are excluded; the list is capped and observational
only, with the standing "association, not cause" caption.
Both surfaces ride the mood-status snapshot too so the assessment prose
never contradicts the on-screen relations. New copy in all six locales.
The target reference card carried an inline "Adjust target range" link,
which read as a control buried in a read surface. Lift the action to a
gear button in the category-page header, to the left of the Coach launch
icon, so the two sit as a consistent icon cluster beside the heading.
A small page-scoped provider bridges the two: the reference card
registers its editable target (one per page, except blood glucose which
fans out to several contexts and registers each) and the header gear
opens the per-metric editor sheet the provider owns. Single-target pages
open that target directly; glucose opens the primary context, since
target editing lives only on these pages — there is no separate targets
management surface. The card itself is now a pure read surface.
The gear carries the action's existing label as its accessible name,
clears the 44 px touch-target floor on mobile, and exposes a stable
`data-slot` for end-to-end coverage.
Per-stage SLEEP_DURATION rows carry the stage segment's END instant, so a
night where the user falls asleep before local midnight spread its stages
across two calendar days. Keying each stage by its own day split one
overnight sleep into two partial nights, and the headline last-night tile
showed only the post-midnight slice.
Reconstruct nights by clustering stage segments into sessions: a new session
starts only when the next segment's start (end minus its minute duration) is
more than three hours after the running session end. A single long stage
block no longer splits a night, a daytime nap stays separable from the
following overnight block, and the gap is measured on the absolute instant so
the clustering is DST-immune. Each session is keyed by the local wake day of
its latest segment end, matching the convention that a sleep session belongs
to the morning you wake up.
Collapse a multi-source night to one canonical source before summing, using
the user's sleep source-priority ladder (default WHOOP > Apple Health >
Withings), so a dual-source account no longer double-counts the total. Sources
are never blended within a night.
Bound the series sleep read to one year even on the 'Alle' range, matching the
slim slice's sleep window, so a multi-year dual-source account does not walk a
six-figure stage-row set into JS.
Add fixtures straddling local midnight, crossing a DST spring-forward, keeping
a nap separable, and collapsing a dual-source night.
Each Trends-row card carries a one-line caption below its mini-chart.
The first choice is the AI advisor's per-metric sentence, but on a cold
briefing — a web-only account, a pre-briefing cached payload, or the
advisor still generating — that sentence is null and every card fell
back to a static "Awaiting more data" hint, even with a full 30-day
series painted right above it.
Add a deterministic, rule-based descriptor derived from the same series
the mini-chart plots, and apply a three-tier caption precedence:
1. advisor annotation present → the AI sentence (unchanged).
2. else, a computable series → the deterministic descriptor:
direction (rising / falling / stable) plus the first-vs-last
magnitude over the window, e.g. "Rising over 30 days (+8 mmHg)",
"Falling over 30 days (−1.4 kg)", "Stable over the last 30 days".
Mood reads on a categorical 1–5 scale, so it is phrased in plain
terms ("Mood trended slightly higher over 30 days") with no raw
point delta.
3. else (genuinely < 2 points) → the real "Awaiting more data" hint.
The descriptor is observational and neutral by construction: direction
and magnitude only, with no causal, diagnostic, or medical claim and no
value judgement — a rising weight and a rising step count read the same
way here. A per-metric stability floor (the larger of an absolute and a
relative bound) keeps day-to-day noise from reading as a trend.
The descriptor logic lives in a pure, unit-tested helper. The caption
reads the same series the chart shows: mood reuses the existing
mood-analytics cache slot, and numeric metrics take a small dedicated
30-day daily-aggregate read keyed under a new query-key factory entry
that joins the measurement-write invalidation set. New descriptor copy
ships in all six locales. The pending shimmer while the advisor is in
flight is unchanged.
SLEEP_DURATION is stored in minutes, one row per sleep stage per night
(IN_BED / AWAKE / ASLEEP / CORE / DEEP / REM) since the HealthKit
category-sample ingest landed. The value is not cumulative, so reading
the single most-recent row surfaced one stage's minutes — never the
night. The dashboard sleep tile and the sleep series therefore showed a
fragment of the night rather than the night.
Add a shared `summarizeSleepNights` helper that groups the per-stage
rows by the wake-day calendar date in the user's timezone and sums the
asleep stages (CORE + DEEP + REM, plus a legacy bare-duration row),
excluding IN_BED and AWAKE per the AASM convention already used by the
sleep score. Wire it into:
- the dashboard summary tile, which now emits last night's time asleep
in hours with an explicit `unit: "h"` and a per-stage breakdown for a
future detail view;
- the sleep series, which now returns one point per night (time asleep
in hours) plus a top-level `unit` for every kind so clients never
infer the unit;
- the slim summaries slice, so the web dashboard tile and the insights
overview read per-night totals (kept in canonical minutes there; the
web tile converts to hours) — strictly more correct than averaging
across stage rows, and parity with the iOS value.
A measurement logged by hand on the iOS app while offline is mirrored
into HealthKit. On pairing a server, adopt-on-pair uploads it as a
MANUAL row while HealthKit background sync independently re-ingests the
same sample as an APPLE_HEALTH row. The two carry different external ids
and sources, so the composite unique index keeps both — doubling the
whole hand-logged history on first pair.
Collapse the pair at ingest: an incoming MANUAL or APPLE_HEALTH row is
dropped as a duplicate when a same-reading row of the opposite source
already exists (same type and value, measuredAt within two seconds).
The match is scoped to that one cross-source pair, so same-source rows
and rows from server-owned sources keep their existing contracts, and
two genuinely distinct readings at the same minute never collapse.
stats:* cumulative rows stay on their per-day overwrite lane.
The collapsed copy returns status "duplicate" with reason
"cross_source_merge" so the client cursor advances unchanged; no
schema, migration, or OpenAPI change.
WHOOP body fields are self-reported profile data; guard mapBody against
absent, NaN/Infinity, or non-positive values so a garbage reading never
seeds a WEIGHT measurement or a profile height of 0.
Extract handleCollectionFetchError so the five resource syncs share one
soft-skip-vs-rethrow path instead of five copies. Behaviour is unchanged:
a per-resource 403 returns 0 and keeps the connection connected, anything
else records a classified failure and rethrows.
The snapshot builder ran its independent reads sequentially on a cache
miss: feature extraction, the shared measurement query, the mood,
compliance, sleep and workout table reads, and the GLP-1, derived,
trajectory and memory helper blocks each awaited in turn. None consumes
another's result, so they now fire concurrently and resolve in batched
hops — feature extraction and the measurement read first (both gate the
per-metric blocks), then the remaining reads in one Promise.all. The
source guards are unchanged, so a disabled source still issues no query,
and the synchronous block assembly keeps its original order, so the
provenance envelope is byte-for-byte the same. Cuts several round-trips
off the cold build; the warm path is cached and unaffected.
The body-measurement endpoint was fetched by a function nobody called, so WHOOP
weight, max heart rate, and height never reached the database despite being
documented in the mapping table.
Add sync-body.ts and wire it into the per-user sync loop (and, through it, the
backfill):
- weight_kilogram → a WEIGHT Measurement (source = WHOOP) on the stable
externalId whoop:body:weight with overwrite semantics. A body measurement is
a single self-reported profile value, not a time series, so a re-sync updates
the same row in place rather than accumulating one row per poll. measuredAt is
the fetch time; the source-priority picker ranks a real scale above WHOOP.
- max_heart_rate → WhoopConnection.maxHeartRate (a profile constant).
- height_meter → User.heightCm (m→cm), written ONLY when the user has no height
yet — a user-set height is never overwritten and height is never minted as a
Measurement.
Every write is field-by-field and idempotent across reruns. Document the body
mapping and the watch-only blood-pressure stance in mapping.md.
A 403 on a single WHOOP collection endpoint was classified reauth_required,
which parked the whole connection at error_reauth and short-circuited every
future sync. A lower-tier user missing one gated data class would brick the
entire integration.
Treat a per-resource collection 403 as a soft skip of that data class: log,
return 0, and continue so sibling resources still sync. Reserve connection-wide
reauth for a 401 (token rejected) and for a 403 on the token-refresh / profile
path, which run outside the per-resource catch blocks and remain unchanged.
Add isCollectionForbidden as the shared gate and cover both branches in tests.
Extend the SSRF pin inventory to also assert the requirePublicHost pins
already present on the user/operator-supplied moodLog push/sync webhooks
and the ntfy sender, so the full set of overridable-host egress points
is CI-guarded and a future edit that drops one reds here.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Prisma unique-constraint (P2002) narrowing predicate had been
hand-copied into five call sites. Fold them into a single isP2002 in
src/lib/prisma-errors.ts and import it everywhere (the two step-drain
sites keep their local isUniqueConstraintViolation name via an import
alias). Behaviour-identical; one place to chase a future P2002-semantics
change instead of five.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The eight v1.10 additive HealthKit signals (cardio recovery, six-minute
walk, stair ascent/descent speed, breathing disturbances, wrist
temperature, falls, walking steadiness) are writable dashboard widget
ids so the native client's Home-pin PUT validates them, but the web
dashboard renders strip tiles via hardcoded per-id blocks and has no
render path for them. Left as-is, the web Settings -> Dashboard list
would offer eight toggles that do nothing on web.
Name them in a new IOS_PIN_ONLY_WIDGET_IDS set (writable-but-not-
web-rendered) and filter that set out of the rendered settings list, so
the toggles never appear on web while the ids stay accepted by the
widgets PUT enum for the native client. Add tests asserting the eight
remain writable (catalogue membership) yet are absent from the web list.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The synchronous DAY rollup hook carries no pg-boss singleton, so two
measurement writes landing on the same (user, type, day) can interleave
their delete-then-insert and the losing createMany trips the per-source
unique index. The peer rewrote the identical source-collapsed partition
— the aggregate is a pure function of the same underlying rows — so the
rollup is already correct. Catch the P2002 in the single-chunk
transactional path, annotate it, and return rather than failing the
recompute. Any other Prisma error still bubbles.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Drop the unused `now` field from the coach `ExtractOpts` type; fact
extraction never read it (conversation-summary keeps its own `now`).
- Correct the no-ladder tiebreak docstring in measurement-read: the
collapse picks the alphabetically smallest source name, matching the
live-SQL `ORDER BY … source` tiebreak, not max count.
- Replace the hand-inlined canonical-source subquery in the summaries
rollup-path narrows query with `canonicalMeasurementsFrom(rank, "90 days")`
for live/rollup parity with its siblings.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add eight widget ids (cardio recovery, six-minute walk, stair ascent/descent speed, breathing disturbances, wrist temperature, falls, walking steadiness) to the dashboard widget catalogue + default layout (hidden by default) so they can be pinned via the layout PUT.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Load the source-priority blob once and pass it into both
readBestGranularityRollups calls. Completes the B3 fast-path threading.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Load the source-priority blob once and pass it into both BP readers
instead of letting each re-query the user.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Load the source-priority blob once and pass it into the three
readRollupBuckets calls instead of letting each re-query the user.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Source-grep guard asserting the openai/anthropic/geo outbound sites keep
the unconditional DNS-rebinding pin and the local AI client keeps its
conditional LAN escape hatch — so a future edit can't silently re-open
the SSRF surface.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add requirePublicHost to the three outbound safeFetch sites whose host is
user- or operator-controlled and lacked the connect-time pin: the OpenAI
and Anthropic clients (BYO base-URL override) and the IP-geo lookup
(IP_GEO_LOOKUP_URL env). A base URL resolving to a private/metadata
address is now rejected. The LOCAL AI client intentionally stays
conditional (LAN escape hatch). Redirect-follow is already closed by the
safeFetch defaults.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a 'What the Coach remembers' settings section: lists the user's
durable facts grouped by category with a learned-when stamp, forgets a
single fact or clears all (behind a confirm), and notes the rolling
conversation summary. Consumes the v1.11.1 CoachFact routes via the
query-key factory; gated on the Coach feature; localised in all six
locales. Closes the API-only gap from v1.11.1.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Remove the two never-wired *_PROMPT_VERSION exports the dead-code gate
flagged, and consume the summary-length target as a real backstop that
caps the stored rolling summary so a misbehaving provider can't bloat
the encrypted column or the next prompt's history window.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- conversation-summary: fault-isolate the per-message decrypt in the
fold loop so one undecryptable historic turn (e.g. mid key-rotation)
no longer permanently stalls a conversation's summary refresh.
- source collapse: unify the no-ladder tiebreak on alphabetically
smallest source across the JS reader and the all-time rollup SQL, so a
ranked vital whose day carries only off-ladder sources resolves to the
same canonical source on the warm and the live-SQL read paths.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the GDPR/trust surface for the durable Coach facts: GET lists the
caller's active facts decrypted, DELETE /[id] soft-deletes one, and a
bulk DELETE clears all — every verb apiHandler + requireAuth +
requireAssistantSurface("coach") + ownership-scoped, with an unknown or
cross-user id collapsing to an idempotent no-op that never leaks
existence. Registered in the OpenAPI contract.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
buildCoachMemoryBlock gains a third fault-isolated sub-source — the
top-N durable facts — folded under snapshot.memory so they ride the
existing lowest-priority registration and are shed first under the
char budget. Add an EN + DE system-prompt paragraph instructing the
Coach to use facts as the user's own stated, descriptive context —
never as a diagnosis, never invented.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
fetchConversationWithMessages now returns the decrypted rolling summary
(fail-closed → null), and buildHistoryWindow substitutes it for the dead
placeholder so a long conversation keeps memory of its elided turns.
Once a conversation grows past the turn cap the chat turn fires a
fire-and-forget coach-memory-refresh so the next long turn is fresh,
without adding a provider round-trip to the interactive reply.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the coach-memory-refresh pg-boss queue: a fire-and-forget enqueue
contract (coach-memory-shared) and a worker pipeline that runs the
conversation-summary refresh then the fact extraction for one long
conversation, each fault-isolated and budget-gated. Register the queue
in allQueues + a boss.work handler in the worker, with a source-grep
inventory guard so the queue can never ship unregistered.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add the two background generators behind Coach long-term memory, both
reusing the budget-aware provider chain (runStatusCompletion) with a
local fallback and encrypting their output at rest:
- conversation-summary: folds the turns elided past the history window
into a rolling natural-language summary, merging into any prior
summary; short-circuits cheaply until a conversation grows long.
- facts: extracts durable, descriptive personal facts (preferences,
stated conditions, goals, constraints, context), gated by a Zod
schema + de-dup + per-user cap; buildCoachFactsBlock ranks the top
facts for snapshot injection. Never diagnostic, never numeric.
Both are pure/injectable for unit testing and annotate counts only,
never content. Wiring into the queue, route, and snapshot follows.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extract the AES-256-GCM Bytes codec from persistence into a shared
bytes-codec module so the conversation summary and the new personal-fact
store reuse one ArrayBuffer-backed implementation. Add the CoachFact
table (migration 0116) for durable, user-scoped facts the Coach extracts
from conversations — encrypted fact text, plain category/confidence/
source for rank-without-decrypt, soft-delete for a reversible forget.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Complete the source-aware rollup read path: the slim summaries slice
(rollup + live), the comprehensive aggregator (both paths, including the
dailyByType correlation feed), and the /api/measurements route (rollup
read + the live date_trunc fallback) now resolve overlapping sources to
the source-priority ladder's canonical reading per (type, day) before
aggregating. A dual-source vital (e.g. WHOOP + Apple Watch resting heart
rate) surfaces one canonical series instead of a blend or a doubled
count; cumulative metrics keep summing per canonical source.
The collapse runs in SQL where the aggregate must stay a single
round-trip (buildSourceRankCase emits a per-user ladder-rank CASE;
canonicalMeasurementsFrom restricts a live read to canonical-source rows)
and in application code (collapseRollupRowsBySource) for the row readers,
both driven by the same resolved ladders so live and rollup agree.
Also widens the enum-literal SQL whitelist to accept digits so VO2_MAX
no longer throws on the typed write-hook recompute or the rank CASE.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
readRollupBuckets and the WEEK/MONTH/YEAR readers now select the source
column and run collapseRollupRowsBySource before composing, so every
rollup consumer (dashboard sparkline, fast-paths, derived metrics, range
deltas) sees one canonical row per bucket. loadUserSourcePriority loads
the ladder blob once; readBestGranularityRollups threads it through each
granularity probe so the per-floor walk never re-queries the user.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
collapseRollupRowsBySource resolves the per-source rollup rows minted by
the source-aware writer down to one row per bucket, using the user's
source-priority ladder (the same ordering the raw-row picker uses). A
metricKeyForType superset maps every overlapping non-cumulative vital
(RHR, HRV-SDNN, SpO2, body/skin temperature, respiratory rate, weight,
body fat, blood pressure, pulse, VO2max, recovery) plus the cumulative
keys onto a SourcePriorityMetricKey. Unlisted sources fall back to the
highest-count row so a bucket never doubles or goes dark.
Wired into the rollup readers in the following commits.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rollup rows were grouped by (type, day) only, so two sources reporting
the same standard vital for one day (e.g. WHOOP + Apple Watch resting
heart rate) were AVG-blended into a single row instead of resolving to
the source-priority ladder's canonical reading.
Re-grain the rollup table to (type, day, source): the writer now groups
by source and delete-then-inserts each affected partition (so a source
that stops reporting a bucket cannot strand a stale row), and the read
path will collapse overlapping sources to the canonical reading in a
follow-up commit. Cumulative metrics (steps, energy) keep summing per
source. Migration 0115 purges the legacy source-blind aggregates; the
boot backfill + read-time self-heal re-mint per-source rows.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fold the deterministic short-horizon forecasting engine into the Coach
prompt as a compact, lowest-signal snapshot block. For each in-scope
metric the block carries direction + slope + the projected horizon end
with its widening prediction band, the last observed value, and the
fit's R² / server-computed confidence — read straight from the existing
projection engine, never recomputed. A metric below the fit/history/
staleness floor is omitted; the block is dropped entirely when none
qualify, and it sheds first under prompt-budget pressure.
Add a conditional-projection ground rule so the Coach narrates a
trajectory only when the block is present and only for the metric it
covers: direction and range, always conditional, never a certainty, a
risk score, a diagnosis, or a dated event; absent block means no
projection. The rule lands in the safety-contract matrix across all six
locales and inline in the EN + DE system prompts, with an overclaim
probe asserting a projection is never upgraded. Bump PROMPT_VERSION.
Assemble a zero-LLM longitudinal memory block from artefacts already
persisted: the latest period-narrative headline and a per-metric
prior-vs-current band recall. The Coach can now reference what was true
last period instead of re-deriving a cold snapshot every turn.
The block is machine-derived (labels plus numbers only), so it is
unencrypted and adds no model cost. Each sub-source is fault-isolated —
a failing narrative read or context build drops that sub-block and never
sinks the turn. It registers against the lowest-signal cluster so the
budget degrader sheds it first, before any clinical data.
Prove a WHOOP run and the same Apple-Health run inside the 5-minute clustering
window collapse to the APPLE_HEALTH canonical row (richer GPS + HR record),
and a lone WHOOP run survives when no richer source logged the session. The
measurement-side WHOOP/Apple/Withings oracle already lives in
source-priority-whoop-applehealth.test.ts; this closes the workout axis.
