The saved Coach window only seeded the snapshot and each tool's default.
A tool could still read back to all time, the availability probe reported
counts and means for history outside the window, and several blocks read
fixed windows of their own (labs 12 months, correlations 180 days, the
coarse tail, the all-time extremes) whatever the setting said.
coachPrefs.defaultWindow is now enforced as a limit through one module,
history-reach.ts: the turn context clamps the client window, the snapshot
clamps its window and gives every fixed-window block either the shorter
window or, when the block's own window cannot fit, leaves it out. Every
retrieval tool clamps its window, an earlier period or a stored table that
reaches past the limit answers outside_reach, and the probe reads only
inside the limit and names older history as outside_reach without any
figure. Follow-up chips that would read past the limit are not offered.
The default (allTime) is no limit and reads as before. MCP clients pass no
limit and keep their own window arguments.
Adds the twelve-month option, rewrites the picker text to state what is
read with the numbers the readers use, shows the same options in the
managed-profile settings, counts cluster members with a singular, and
drops the opacity on switched-off rows that put their text below WCAG AA.
Under an answer: copy, read aloud, try again and the time as text on the
left; tokens and model on the right, on one line down to 390 px. The
thumbs leave the web (the feedback route stays for the native client),
and so does the "What I'm looking at" disclosure: the method line and
the tables the answer read move into the steps list, and an older
message without steps lists the areas it drew on, never a raw key.
Remember becomes a bookmark icon in the question's row and confirms in
place with a toast. Follow-up chips render inside the latest answer's
column at its text size. The answer column takes the full width below
`sm` and a fixed 80 % above, so nothing moves sideways as a reply
streams. A round button fades in when the thread is scrolled away from
the end and brings it back.
The hero led with the first sentence of the cached briefing, which was
often a greeting ("Good morning.") while the sentence with content was
never shown. Without an AI provider it fell back to "Your health score
today is N.", which the hero then removed because the ring already
shows the number, leaving a ring and "Nothing needs your attention".
The daily digest now publishes the lead resolved on the server and a
short list of facts about the day:
- lead: the reaction line or the first briefing sentence that is not a
greeting and does not repeat the score while AI text is available,
otherwise a deterministic sentence about the strongest signal (a vital
outside its personal range, a night an hour off the usual, vitals in
range, last night). A fact the lead was built from leaves the list.
- today: up to five statements in priority order (Rest Mode, doses taken
today, an appointment today or tomorrow that is not already on the
rail, last night's sleep against the usual, vitals against their
personal range, cycle day). Module-gated on the server, never padded,
never asking for anything. A phone shows the first four.
- restMode: passed through while an illness episode is active.
- score.steadyWeeks: how long the score has held, read off the stored
daily scores, shown under the ring in place of a delta.
briefingLead skips greetings and score sentences the same way, so the
morning push improves with it. Visit times are rendered in the profile
timezone and hour-cycle preference; the cycle day comes from the same
verdict the cycle page uses. The ring, the rail and the charts are
unchanged. OpenAPI regenerated; additive fields only.
Measures that the thread and composer keep the same width with the panel
open and closed, that the choice survives a reload, Escape and focus
return for the panel, the sheet and the settings overlay, rename and
delete with undo from the row menu, the settings deep link, and axe on the
panel and the overlay in both themes. Screenshots are attached to the
report.
The Coach page gets a conversations panel on the right of the thread. From
1280 px it docks beside the thread and slides open and shut; the reading
column keeps its width either way and only its centring moves. It is open
on a first visit and remembers the last choice per device. Below 1280 px it
is a sheet from the right. The toggle sits at the trailing edge of the top
bar, through a small page-owned slot the top bar now renders.
The panel lists every conversation grouped by recency, with a row menu to
rename in place (Enter saves, Escape cancels without closing the panel) and
delete with the existing undo toast. The undo window now survives the sheet
closing. New chat is the round button at the panel's bottom right, and
picking or starting a conversation keeps ?c= in the URL.
The gear in the panel header opens quick settings: a popover from md up, a
bottom sheet on a phone. It switches the answering provider (chain reordered
through the existing route), the model and reasoning level where the person
owns them, and shows what the Coach can see. Keys, the Codex sign-in and the
fallback chain stay in Settings > AI. The drawer's gear opens the same
overlay. The composer's + menu is gone.
Settings > Coach links to /coach?settings=data, which opens the overlay on
what the Coach can see instead of a rail that no longer exists.
The first case spent 14 s per full-snapshot build in the all-time pulse
aggregate, planned against a table with no statistics, until autovacuum
analyzed it; on a loaded runner that took it past its 60 s budget. Analyze
measurements and rollups after seeding, as a production table would be.
The cases and the comparison they make are unchanged.
The parity test built a fresh Intl.DateTimeFormat for each of ~36k
comparisons and crossed the 5 s timeout on CI runners. One formatter per
zone halves its time; it also gets an explicit 20 s budget, since it is
meant to be exhaustive.
backupPayloadSchema is built when validations/backup.ts is evaluated and
holds about 12 MB of heap. The restore job, the stored-backup readers and
the export writer import that module, and they sit on the boot path, so
every server process built the schema at start whether or not it ever
touched a backup.
The schema version, the compatibility check and the per-section summary
move to validations/backup-summary.ts, which never builds the schema. The
call sites that validate a file (upload, restore, the summary and download
routes, and the streamed reader's element check) load backup.ts with a
dynamic import at that moment; everything else takes only types. The
OpenAPI registry keeps its static import, since only the spec generator
loads it.
After warm-up on a production build under --max-old-space-size=524, the
used heap drops by about 29 MB, and the schema no longer appears in a heap
snapshot. A structural test refuses a new static value import of the
module outside the registry.
get_metric_series (and get_metrics, the metric resources and the review
prompts, which fan out over it) answered with one section of a Coach
snapshot, but built the whole snapshot to get it: the all-type feature
extraction, the derived wellness layer, trajectory, correlations, memory,
GLP-1, cycle, illness, labs and visits. On a seeded account with about
215k measurements a resting heart rate read ran 208 SQL statements and
blocked the event loop for 175 ms to return under 2 KB.
The MCP surface now asks the executor for a source snapshot. It runs the
same prelude (prefs, module gates, admitted sources, cutoffs, now shared in
snapshot-prelude.ts) and the same block builders over the requested
source's reads only. The core clinical aggregates come from extractFeatures
scoped to the one block, whose bulk read keeps its all-type cap so the rows
a block sees do not change. A section large enough for the full builder's
prompt-budget pass to trim, and the environment cluster, go to the full
builder, so their results are the old ones by construction.
The integration test compares the tool's serialised result with the full
builder's section and grounding for nine sources over four windows, the
get_metrics batch and the metric resource, on a seeded account dense enough
to put the feature row cap into play.
Node sizes its heap from the container's memory cap. The 1g default gave
524 MB, which a long-used account fills in normal operation: on a live
instance 490 to 507 MB stayed live after a full GC, garbage collection
stalled requests for up to 8.5 s, and the process ran out of heap 26 times
in a week, taking the weekly backup with it. 1536m gives 792 MB. The
example env and the backup runbook say how to lower it on a small host.
Both grids carry an aria-label while their data loads, on a plain div,
which axe reports as aria-prohibited-attr: the label has no role to
name. A scan that lands while the scores strip is still loading fails on
it. The loading grids now take role=status, the role a loading region
has and which already implies the polite live region; the loaded vitals
grid carries neither.
Entering the Insights overview fired every section's reads at once, about
two dozen requests on a seeded account, and the server worked through
them in a queue: the hero and the first sections waited behind trends,
the period review, cycle, signals, rhythm events, health status,
breathing, labs and ECG, all far below the fold.
The first three sections in the account's own order still mount with
the page. Every later one renders a 1 px sentinel until it comes within
400 px of the visible area of the shell's scroll container, then mounts
and fetches; once mounted it stays. Measured on a seeded account under a
4x CPU slowdown: 24 -> 13 requests on entry, the entry network span
1.3-1.8 s -> 0.5-0.9 s, LCP about 1.35 s -> 0.9 s, blocking time about
550 ms -> 200 ms. Scrolling to the end still loads all 25 and renders
all ten sections.
Specs that assert on, measure or scan the lower sections scroll them in
first through a shared helper, so their coverage is unchanged.
On a short window (around 785 px and below) the nav list is taller than
the room between the logo band and the footer group, so it scrolls. With
nothing marking the edge, the last visible entry was cut off right at
the Collapse row and read as the footer covering it; macOS hides the
scrollbar that would have said otherwise.
While the list overflows, the footer group draws the same hairline the
user section does. A ResizeObserver on the nav and its list keeps the
state current as the window or the entries change; with room to spare
the rail is unchanged. The e2e spec checks the edge, that the last entry
scrolls fully clear of the footer, and that a tall window draws no edge.
The in-memory day-aggregate fake re-filtered every row for every day,
converting each timestamp to the profile timezone again, so a 1000-day
series took about five seconds and the graded pulse-status test hit its
timeout. It now groups by day and segment in one pass (4.8 s to 85 ms).
The medications list and the compliance summary refetch on every mount,
so a take or skip logged on another device shows on return (#316). The
nav link's hover / touch / focus intent and the route-commit preload
prefetch both reads too, so every visit through a link sent each of
them twice: the prefetch, then the identical request from the mount a
few hundred milliseconds after the first answer landed.
The mount now skips its refetch only when the cached data answers a
prefetch that went to the network for this visit: within the last 10 s,
and since the route last left /medications. A direct load, a return
with an older entry, or a quick return whose prefetch found the entry
still fresh keep the refetch on mount.
A Coach snapshot, a derived-metric batch or the Insights page fans out
into a dozen readers that each probe the same account's rollup coverage
at once. In production one MCP get_metrics call held 59 of these probes
open on the database together, each taking about 320 ms under that
contention, and the same burst repeated every 15 minutes.
Concurrent probes for one account now join the query already running.
Nothing is kept after it settles, each caller gets its own map, and a
measurement write drops the in-flight entry so a probe that starts after
a write always reads afresh. The registry lives on globalThis so every
server bundle shares it.
A document whose preview render fails for good (a PDF whose first page
renders blank) was queued by the boot-time backfill, refused, and queued
again at the next boot, without end.
The thumbnail job now keeps the failure with the document
(document_thumbnail_failures, migration 0374): the reason, the attempts
since the last success and when the newest failed. The backfill, both the
boot discovery and the per-account pass, leaves a document out after one
deterministic refusal (raster-failed, empty-render, pixel-cap) or three of
any other kind, until the document changes. A direct enqueue still renders
it, and a preview that is made clears the record.
A weekly or nightly backup job whose process died (a restart, a
container out of memory) stayed active until its two- or four-hour
expiry, and its retry walked the accounts in the same order: the account
that killed it has the oldest copy, so it led again, and every account
after it went without a copy for as long as that record stayed too
large.
Both passes now send with a five-minute pg-boss heartbeat, so a dead run
fails within minutes and its retry picks it up. Each pass records when it
started and finished every account (backup_pass_attempts, migration
0373); an account whose last attempt never finished goes behind all the
others. The backups page shows when the run in progress started and names
an account a run died under, for the weekly and the off-host card, and a
heartbeat timeout reads as what it means.
Since v1.39.6 the preview scanner parsed the JSON on its way into the
pieces, kept every section but the three bulk tables, and validated the
whole document at the end: a second copy of the record beside the
writer's own, inside the weekly job. On an account with a document vault
that is most of the record.
The writer now shows each section and each bulk row to an observer just
before writing it. The preview collects key uses from that and takes its
counts from the writer's own FullBackupCounts; nothing is parsed or kept.
An upload hands over the summary and key uses it has already validated.
The round trip pins the writer's counts and key uses to a full read of
the copy on an account with a row in every model, and a memory test holds
what the pass keeps once the producer is done.
Measured on 70,000 measurements, 3,000 mood entries and 40 one-megabyte
documents under --max-old-space-size=524: held after the producer
56 MB before, 3 MB after; peak heapUsed 398 MB before, 334 MB after.
The drill read only the alphabetically last object of the newest date, so
one account stood in for all of them. It now checks every object of that
date one at a time, records a failure against the account it belongs to,
keeps going, and pages once per failing object with its key and how many
accounts were checked.
A text column written before key ids existed holds the bare base64 of
nonce, tag and ciphertext. The key check read such a value as an encoded
Bytes column, decoded it once too often, and accused the v1 key of not
opening a copy it opens. The instance's stored AI key has exactly that
shape and is the shortest v1 value in every disaster-recovery copy, so it
was picked as the probe and refused every restore of an account with v1
values, even with the instance settings left out of the verdict.
A legacy value now reads as a string unless its decoded head is base64
text. Samples are kept per section and only the sections the verdict
counts are probed; a key is unreadable when none of its probes opens.
Binary values open with the label their column is sealed under, taken
from the encrypted-column registry, and are not probed when no label is
known. A preview stored by an earlier release is still read, and its
sample's layout is derived again from the value.
Day keys and wall-clock fields are computed per row in the analytics,
digest and personal-record loops, and each call ran
Intl.DateTimeFormat#formatToParts plus a resolvedOptions() allocation
to re-validate the zone. In a CPU profile of an Insights page load the
zone helpers were 17-19% of the busy time on the request thread.
Probe the offset once per UTC hour at both ends of the hour and derive
the wall clock arithmetically; an hour that contains a transition, an
instant outside 1971-2199 or an invalid date keeps the formatter path.
Zone validity is memoised per name. userDayKey drops from 9.3 to 1.5 us
per call, wallClockInTz from 6.4 to 0.5 us; a parity test checks the
result against Intl around every 2026 transition in ten zones.
The health score card passed the episode onset, an ISO instant, straight
into its sentence, so it read "Rest Mode active since
2025-08-04T09:55:00.000Z". It now shows the day that instant falls on in
the profile timezone, in the user's date format. The test that pinned the
raw value now pins the formatted day.
probeRollupCoverage runs ahead of the dashboard snapshot, the analytics
slices and the comprehensive insights. It read every live row of the
account for SELECT DISTINCT type and counted every DAY bucket per type,
which the planner served with a sequential scan of measurement_rollups.
Walk the distinct types with a recursive CTE along the live index and
test coverage with EXISTS. On a copy of a long-lived account the probe
drops from 24-50 ms to under 2 ms and returns the same map; an
integration test keeps the old query as the oracle.
Marc-André Bombeck