Feat/UI rebuild - #253
Conversation
lib/ui is gone, lib/ui2 replaces it. the old one threw away five of the six fields analytics emits, so every number rendered bare — no confidence, no tier, no reason when it was absent. real bugs this turned up, all live before now: - metric.dart matched 'AUTHORITATIVE' but analytics emits 'AUTH', so that tier always parsed to unknown - every poincare row rendered empty (heart['irregular'] is a plain map read through envValue, and the fixture had the wrong shape so it hid it) - workout avg hr was lost after 3 days, recomputed off the pruned substrate. banked to sessions.avg_hr now - crossDayArtifactUsableToday ignored its own algo_version stamp - a flat series drew pinned to the floor of the card - gen5 stepMotionCounter never reached Substrate circadian was written, tested and called by nothing. wired it on hourly hr means (raw accel is pruned at 3 days) and said so in the envelope note. things the app can't actually produce are deleted rather than explained — blood pressure, whole-day energy, power zones. docs/internal/UI_ROADMAP.md has what each would need. schema 38, algo 66.
MetricRow only drew ConfDots when there was no sparkline, and when it did it drew a hardcoded Conf.none instead of the conf the caller passed. so every health row — resting hr, hrv, the lot — showed three empty dots no matter what ConfX.of(m) actually said. two goldens had baked that in. conf is nullable now. null means confidence doesn't apply to this row, which is the honest answer for anything the user typed in themselves; three green dots on a hand-logged meal is a claim we can't make. also noted in the golden test that they were baked on flutter 3.44.9 — an older sdk anti-aliases hairlines differently and fails a few of them on nothing but sub-pixel blend.
seven agents audited edge, analytics and protocol; this is the edge half. findings are in docs/internal/sweep/. decoded_onehz's sensor columns were NOT NULL, so a missing sample had nowhere to live and got written as a real zero. absent accelerometer became (0,0,0), which enmo, immobility and auto-workout all read as a perfectly still wrist — accelPresentAt existed and had exactly one consumer. schema 39 makes the six sensor columns nullable and the readers absence-aware. the optical and thermal columns were the same shape of bug but less harmful in practice, since every adc reader already gated on v > 0. a single nan was destroying the whole crossday bundle for every user, days 3 to 8. readiness_glassbox emitted percentileOfYou = nan under 7 days of history, jsonEncode threw, and the catch dropped illness, anomaly, ctl/atl/tsb, chronotype, sleep coach, vo2max and every percentile, leaving a debug line. sanitizeForJson drops the offending key now, records it, and the catch is loud. the rollup carried no algo_version and no date and nothing checked either, so weeks-old numbers under an older algorithm looked exactly like today's. it's stamped on write and gated on read, and withheld is said out loud rather than dressed up as a cold start. charts were drawing pictures the data didn't support. nothing carried a timestamp, so a 30-day view over 22 stored days drew 22 evenly spaced points and ran the line straight through the missing week — series are dense now, one slot per calendar day, null for a day that didn't derive. AxisSpec.of looped forever on a non-finite value. a flat week drew pinned to the card floor. fill over a min-anchored scale turned 58 to 60 bpm into a mountain. four painters still take compacted series; that's tracked, not done. dangerousCmds was only enforced in _send, so nine call sites reached _write straight past it. the block lives in _write now with one audited opt-in — setFfValue, which the R22 sequence bypasses deliberately. the ui rebuild had orphaned working features: the alarm had no screen while the band still fired it off its own rtc, the widget lost its foreground refresh, four notification routes landed on home, and swiping back out of a live session threw away every logged set and then blocked all future workouts. the app also explained itself far more than it said anything — 69 statuscards against one trendcard and zero recommendations. cycle was fully built in the data layer with no ui at all. nutrition couldn't delete. that's started, not finished. 22 notification kinds down to the three we actually sanctioned. headless gate has a ceiling. two derivations can no longer race. schema 39, algo 67. needs the new analytics and protocol shas pinned.
six more agents, an adversarial pass over the last round, and the findings in docs/internal/sweep/. the adversarial pass refuted 18 of 44 claims from round one, which is roughly the rate i'd expect and the reason it ran at all. privacy first, because it's the only place the code contradicted a public promise. PRIVACY.md claimed a settings toggle for health-data contribution and setHealthShareConsent had zero callers — its ui died with lib/ui. fresh installs default off and were safe; an upgrade carrying true uploaded the whole gzipped db daily with no way to stop it. and refreshAppStatus GET'd the companion server on every launch and every foreground, gated by nothing, with COMPANION_URL baked into releases. there's a toggle now, an update-checks switch, fetchStatus returns null unless sideload ota is on, and PRIVACY.md says what the code actually does. "reset all data" left about twenty tables standing and cleared two prefs. wipeAll enumerates sqlite_master now, so the next table someone adds can't escape it. export existed with zero callers while the reset dialog said "export first" — it's in settings now, csv and whole-db, with the exclusions listed on screen rather than silently dropped. two ways to brick the install: _init had no try/catch and was called unawaited, so any throw left a spinner forever with no retry; and the migration ladder was one transaction with no guard, where a failure meant reinstall. init failure is a state you can recover from now, and an unopenable db is quarantined under a unique name, rebuilt, and merged back — hand-typed tables first, since those are the ones nothing can recompute. if the salvage also fails the original is renamed back. the app knew what was wrong and threw it away: lastError, spotError and six ble flags rendered by nothing, a bluetooth permission denial reported as "no band in range", a db read failure shown as "nothing derived yet" on nine screens. there's one BandStatus now, built by one pure function, so the phone, pairing and devices rows can't drift into three answers. it also grew forever. the 3-day prune only ran under scope.fullHistory, which is the manual re-analyze button — 12.3 mb/day, about 1.1 gb at 90 days. and sqlite hands freed pages to the freelist, not the filesystem, so there's a vacuum behind a freelist threshold. substrate moved to Float64List: 83 to 48.6 mb held, bit-exact, no derived number moves. health export held 400 decoded bundles at once — measured 298 mb — and now holds one. accessibility: three cards overflowed at 2.0x in the exact places goldens existed, passing only because every fixture was two characters. fixtures are realistic now and there's a non-golden sweep to 3.1x that fails on any overflow. chart palettes went through the contrast solver instead of raw pigment — the hypnogram's light lane measured 1.67:1. expand:false silently dropped the 44pt guarantee on twelve controls including a destructive one. widget extension was pinned to ios 26.5 with 16.0-era code, so widgets and live activities were dark for nearly everyone. four different readiness bandings across phone, widget and watch — 65 was green, orange and yellow depending on where you looked. one banding now, read from the phone. 324 of 348 goldens moved. p.on solves twice now, against card2 and against wash(accent) over it — five of six accents measured 4.30-4.49 on the tinted surface nothing was checking. algo 67 and schema 39 both unchanged; everything here is bit-exact or scheduling. still needs the analytics and protocol shas pinned.
five agents. the widget contract, the live session, the band's own failures, and every word on screen. the widget was shipping grey. a native pass had correctly deleted all four duplicated readiness-banding tables — 65 rendered green on the phone, orange on the widget and yellow on the watch — but the dart side that publishes readiness_tier and readiness_band didn't exist yet, so every surface fell back to unlabelled. it publishes both now, key names checked against all four readers. the widget also couldn't go stale. has_data is a bool frozen at push time, so a phone that hadn't synced in a week showed week-old readiness looking like this morning's, and updated_at was published and read by nobody. freshness is computed at render time now, and the timeline emits a second entry at the deadline so widgetkit flips it with no process wake. 26h for metrics, 24h for battery — a readiness score describes a night that stays true all day, a battery percentage describes now. also fixed a 2038 bug on the way: updated_at read through Int overflows negative, which would have made every widget permanently stale. the live session re-laid-out its whole body, RouteMap included, every second of a workout: 192 rebuilt widgets per tick, now 21. strength opts out entirely, but its one live element asks for the tick itself so heart rate doesn't freeze. the picker built 67 activity rows nobody was looking at. the app knew what was wrong with the band and threw it away. one BandStatus built by one pure function now, so devices, pairing and home can't drift into three answers. pairing checks the blocker BEFORE the string matcher, because adapterOff matches none of its phrases and a switched-off radio was being demoted to a band fault. lastError and spotError deleted — computed, rendered nowhere. copy: always-on prose down 20%, absence copy down 41%, statuscard "why" from 17.6 words to 9.7 across 74 cards. but the real fix wasn't editing text: _Cta drew a tappable arrow whether or not onFix existed, which is why 60% of cards had a dead link and why the prose had stretched to compensate. one bool. settled the vocabulary too — band not strap, readiness not recovery, logged when you type it and recorded when the sensor catches it. three bugs this round were invisible because the fixture was too tidy: SourceRow overflowed for any status longer than "Connected" and every fixture was connected:true. same shape as the two-character cards and the evening-only greeting. fixtures get the longest realistic value and the unhappy branch now. also: go() was a raw PageRouteBuilder, which has no back-gesture machinery, so ios edge-swipe-back was dead on about twenty screens.
you looked at them on the phone and said no. gone: the widget, the Conf enum, the ConfX tier mapping, the conf parameter on SignalCard/TrendCard/MetricRow/ ChartFrame, and all 32 render sites. the tier and confidence stay in the envelope — this was only ever a rendering choice, and lib/data, lib/compute and lib/models are untouched. taking the parameter out (rather than just the render) exposed a chain of dead code behind it: _confOf, confOfEnv, chronotypeConf, cosinorConf, and MetricData.latest — which was the only caller of todayMetric(), which was the only reason metric detail did a getToday() fetch at all. that fetch is gone too. absence is a separate rule and is untouched. still no bare em-dash, still a StatusCard with a reason. one place needed care: in readiness_detail the dots WERE the absence affordance for a missing contribution, so that now renders nothing rather than a dash — the sub-line above it already names the case. the prose caveats matter more now that nothing else carries the distinction, so they stay: skin temp isn't calibrated so there's no °C, deep is the weakest of the three stages, MET is estimated. 194 goldens moved, 4 deleted. 2358 tests.
removed, all your call: - the export-exclusions paragraph on the data screen - "everything stays on this device" on profile - the workouts / records / days / sources tile - activity history (it was a count with no destination anyway) - edit profile from more settings — it lives in quick access now, one door - "stages are a low-confidence estimate" and its "how this is computed" link taking the tile out made four ProfileStats fields dead, and they weren't free: they cost a getRecords() and a whole year-of-workouts query every time the profile opened. both gone with them. the staging caveat still exists where it belongs — metric_detail's method text says a wrist sensor can't see slow-wave activity, and that's the place someone goes when they want to know.
one question — how did my night go — answered in three seconds, then revealed by scrolling. total sleep, onset to wake, in bed and how much of it you were asleep. then the hypnogram, then stages, then how it sat against your own nights, then anything that stood out, then the signals underneath, then one thing to do tonight. no sleep score. there's no composite in the pipeline and inventing one here would mean picking weights in a ui file. what replaces it is the comparison every "86" is a lossy summary of anyway: last night against the middle half of your own recent nights, per measure, with the night count attached. a quartile band needs no thresholds and no population norms — it's your own distribution, so it can't be wrong about somebody it was never fitted to. 7 nights before it compares at all, 14 before it calls anything extreme, and last night is cut out of its own baseline. dropped from the brief: the score and the quality bar, confidence dots (you had those removed a commit ago), sleep replay, spo2 (decoded but never validated, every stored value is null), the data-quality block, and possible contributors — getJournalInsights correlates day-level readiness against journal tags at a 4-day floor, so there's nothing per-night that would survive being asked how it knows. merged the brief's sections 4 and 5: "how did the night go" and "what was different" were the same four comparisons twice. the bar is the delta and the sentence is the verdict. two real bugs on the way. the breathing lane was dotted on every single night — the grid was sized to 1-minute heart rate while breathing and temperature are 5-minute, so real samples read as dropouts. and the hypnogram drew one rect per column with a 0.8pt gap between neighbours, so a solid two-hour stretch of light sleep came out as a picket fence; it draws runs now, with a riser at each transition, which is what a hypnogram actually is. history is three metric_series scalar reads and one window query, not 28 day bundles.
the overnight footnote is gone. the legend already names each lane and its unit, and the lanes are visibly separate — the paragraph was describing the picture instead of letting it do its job. steps: a real 0 already rendered as 0 (isEmpty is value == null, so an unmoved counter is a reading, not an absence). what you were seeing was genuinely nothing recorded — gen4 has no pedometer, so it's phone steps or nothing. it no longer swaps the tile for a paragraph about wrist motion: the tile stays and says None / NOT RECORDED. the answer to "how many steps" is short either way.
you asked why 71.6 never moves when you switch 7/30 days, and whether there's a today. i had an agent trace it with a real harness rather than by reading: 7 days of synthetic 1hz raw, today truncated mid-afternoon like a real mid-day look, through the actual DerivationEngine. today derived clean — partial:0, and the same 24 metric_series keys every other day gets. all four structural suspects ruled out: today isn't filed partial, it's scheduled FIRST when it has raw, metric_series has no version/finalized/skipped column to filter it out, and the local-noon stamp lands it in the right slot. so it was never a data bug. both symptoms were the same presentational one: the headline was UNDATED. it named its day only when it was stale, so a fresh number sitting under 7/30 tabs told you neither that it was today's nor that it wasn't a summary of the range you'd picked. it now always says "Latest reading · Today", or "· 3 days ago" when that's the truth. the range still changes the chart and the lowest/typical/highest block, which is what it was always doing. also fixed the x-axis off-by-one — slot 0 of a 30-slot window is 29 days back, not 30. 71.6 was false precision. one metricValue helper now: bpm, ms and % whole, br/min and degrees to a tenth, minutes as 7h 30m, steps and kcal grouped. metric_detail and health_screen both route through it, so the same reading can't read 71.6 on one screen and 72 on the card that links to it. a delta that rounds to nothing says "in line with your 28-day average" instead of drawing an arrow over "0 bpm above".
you were right and i was wrong twice. first: the headline. i told you it correctly stays fixed because it's the latest reading. that's a bad design, not a defence — a big number under 7/30 tabs is read as a summary of the range you picked, and on steps it was worse than confusing: today's 43 under a "30 days" tab reads as a month's total. the headline is the window's daily average now, and it carries the day count — "1 of 30 days" — which is also the honest explanation for the case that looked broken. with one day of history, seven days and thirty days really do average to the same number. the latest reading is still there, on its own line, dated. second: the flaky test. two agents reported workout_reliability flaking and i told you it was contention from eight of them hammering one laptop. it wasn't. it fails about one run in four on an idle machine, in the full parallel suite, and passes 6/6 alone. the cause: the test slept 150ms and asserted runs == 0. that was flaky AND weaker than it looked — zero is also what you see when nothing was ever queued, so it would have passed before the enqueue landed. it waits for pending_light now, which is the actual parked state, and since _arm() returns early while a workout is live there's no timer at all — so once the job is parked, runs cannot advance however long anything takes. deterministic instead of merely patient. 5/5 green.
"what is it right now" and "what has it been lately" are different questions, and a range list starting at 7 days made the first one unanswerable. Today is its own window. on it the headline is today's reading labelled Today, not a one-day average, and there's no chart — these series carry one value per day, so a one-day window is a single point, and a single point drawn on an axis is a shape pretending to be a trend. "your normal range" reads the whole history there instead of the window, because otherwise its lowest, typical and highest would all be the same number: today's. on why 7 days showed you today's steps and nothing else — that's not the screen picking today, it's that today is the only day with a steps row. gen4 has no pedometer, so steps only exist from the day phone steps started reporting. the headline says "1 of 7 days" now, which is the part that was missing: the number was right, it just never told you how thin it was. nothing recorded today gets its own words too, rather than "no history yet" — one is a gap, the other is a new install.
alignment: the name is the only flexible part of a label→value row; the measurement keeps its natural width and sits flush right. two flex children splitting the row by ratio is why every value block started and ended at its own x — the lab rows already did it the right way, which is why those were the ones that looked correct to you. _vital turned out to be a private re-implementation of MetricRow and is gone. its value/unit stack shrink-wrapped to the QUALIFIER, so "bpm today" and "SD from your own nights" were setting each row's width. a qualifier isn't a unit; it sits under the name now, and 48-142 bpm / 14.2 br/min / +0.31 SD line up. that also caught a shipped bug: respiratory rate on Overview rendered as "1… br/min" — a Flexible gave the measurement a quarter of the row and ellipsised it. splits: the tab is per-archetype now, so flow, match and basic don't get one at all. a route with no gps still gets "no splits for this session" — that's expected-absent, which is different. share had four style names per archetype where ShareCard only branches on one, so three of every four were labels for an identical card; two now, and the second only when there's art to draw. RR, since you asked: it's real, and the two tabs disagreed. getToday gated it on resp['rsa'] being a Map while getDayLungs called the same function with no such gate — so Overview could say "no respiratory rate" on a day whose Vitals tab printed 14.2 br/min from the same bundle. the rsa block only supplies a confidence and _respObj already falls back to 0.5, so the gate was suppressing real readings. removed. it still needs a scored night, and the copy says which of "no night" or "too noisy" applies. also removed the skin-temp caveat card and the sleep architecture deep dive from vitals, and "what interrupts you" is "manage notifications". gallery: seven taps on the version row in settings. 41 components, live, with text scale and light/dark switches — the two dimensions every layout bug we've shipped actually lived in. goldenCases() moved into it, so the goldens now photograph the same list the gallery renders and the two can't drift.
you said it reads "beat timing was too noisy to recover one". that sentence was written by hand in the ui and shown for every kind of failure — it happened to be one of four possible reasons. the estimator already records which gate it failed, in respiration.rsa: too few beats for an RSA spectral estimate (needs 20), artifact fraction over the gate, no stable HF respiratory peak resolved, or a peak that moved between spectral resolutions. that envelope was being read for its confidence and nothing else — when there was no value it returned a bare null and the screen fell back to the hand-written guess. the note comes through now, so the card tells you which gate your night actually failed. that's also the answer to your question: RR needs beat timing clean enough over the sleep window, and the card will now name the specific thing that wasn't.
the workflow died on the usage limit before it could fix anything, so these are by hand, each re-verified against the code first. home was painting a held-over prior night's readiness under today's date and calling it "Today's readiness". getToday holds the last scored night over until today's settles — every morning before the first sync, and the whole of any gap after one — so readiness, sleep, rhr, hrv and skin temp described an older night while steps and energy beside them described today. the deleted ui had a gate for exactly this (settledReadinessScore, still pinned by a regression test) and ui2 never read it. the skeptic found the hunter had UNDERSTATED it too: _latestBundle reads 14 ROWS, not 14 days, so the number has no age cap. withholding it would be wrong — it's real and it's the most recent one there is — so it now names its night: "Readiness · 16 May". the workout tab loaded once and cached for the life of the widget, so a session you'd just finished was missing from history, this week, tracked and the weekly load until you restarted the app. insightsRevision now ticks when a session is durably written, and the screen re-reads when it moves — which covers the manual finish, the gesture path and the live activity alike. pause didn't pause HIIT. the interval timer had no pause check at all: it kept counting down, buzzing and announcing work/rest at a phone in your pocket, and BANKING rounds you never performed. a five-minute pause wrote four rounds of 45s work into the summary while the elapsed clock, which does subtract pauses, said the session was shorter than the rounds it listed. round hr samples are skipped while paused too — standing still would drag the mean toward resting. wellness wrote the whole day's journal from a snapshot taken when the tab last loaded, and putJournalMetrics deletes the day before re-inserting. so: open wellness, write your journal from compose, come back, tick one habit — journal gone. it re-reads the day before merging now. three confirmed S1s left: the live summary's compacted hr curve, ChartFrame's excludeSemantics eating the scrubber, and import reporting rows as days.
the live summary drew a band dropout as a continuous line. perMinuteHr only grew when a minute HAD samples, so a 40-minute session with a ten-minute dropout produced a 30-entry list; the summary maps index to x, so minute 9 was joined straight to minute 21 and every later reading was drawn ten minutes early — while the SAME session reopened from history was dense and showed the gap correctly. the accumulator is dense now, indexed by session minute, and splits into two readers: perMinuteHrDense for the chart, which needs the time axis, and perMinuteHr for strain and the mean, which want the readings. avgHr divides by the minutes that measured something rather than by the slots. import reported a ROW count collapsed to the literal 1, so restoring a year of history said "1 day imported". day_result's primary key is (day_id, algo_version), so rows were never days anyway — a history through two algo bumps has two rows per day. it asks the source file for a distinct day count now, and reports nothing rather than a wrong number if the table can't be read. ChartFrame used excludeSemantics to stop a screen reader announcing the bare axis ticks and the doubled header. it did that, and it also deleted every descendant semantics node — including child, which on the sleep screen is the Scrubber, so the one interactive control inside a chart was invisible to VoiceOver and Switch Control. the decoration is excluded piece by piece now and child is left alone. explicitChildNodes matters too: a container without it merges its descendants, which loses the control the same way by another route — the test caught exactly that on the first attempt. that's all 7 confirmed S1s. 2388 tests.
the first hunt's skeptics for wiring and analytics all died on the usage limit, so those 80 findings were one agent's opinion with no second pass. re-ran it: one skeptic per module told to refute, and anything it couldn't settle went to a second reviewer told to REPRODUCE it instead — opposite priors on the same claim. 67 survived, 5 died there. the worst ones, all S1: after the band drops mid-workout, the 1hz tick kept billing device.liveHr as a live measurement into zone-seconds, calories and strain, and persisted it on stop. nothing nulls that field on an unintentional disconnect — disableLiveStreams only runs on the intentional path — and the freshness stamp that exists, liveHrAt, had zero readers. the stale value also suppressed the honest "band is not connected" card, because that only shows when hr is null. the offloaded second half recomputed headline strain from scalars['rhr'] — the general resting hr that deliberately falls back to DAYTIME hr — and overwrote scalars['strain'] with it, so a day where the pure pipeline abstained still published a strain built on an awake "resting" heart rate. reproduced against a real export: two days with clinical.strain absent, scalars.strain 6.0 and 8.7, rhr ~84 against the same user's real nocturnal 59-65. needsRepairGuide could be set by one failed createBond or two quick disconnects, and the only line that clears it was unreachable unless autoReconnectPaused happened to be true — so a working, syncing band told you permanently to forget and re-pair it. nutrition's 7-day macro means averaged FLOOR totals as if exact: a day counted toward the average on energy completeness alone, and the mean then admitted any non-null protein sum including one that summed past occasions carrying no macro figure. the energy card says "at least"; the macro means had no such channel. medication adherence counted scheduled doses on days BEFORE the medication existed as missed, because nothing bounded schedule resolution by the def's creation date. add a med at 3pm and it immediately read "0 of 7". the widget, watch and siri still published the hardcoded 480-minute sleep need that the phone's own screens refuse to show. also removed, each a validated finding rather than a cleanup i asked for: proximity_policy (262 lines instantiated by nothing but its own test, fed by readRssi which has no caller either) and LiveWorkoutState.firedMilestones (a set documented as suppressing a banner, haptic and confetti that do not exist). insightsRevision was on that same dead-symbol list and is NOT removed — it has a listener now, added two commits ago when the workout tab stopped reloading.
all 83 from docs/internal/sweep/SLOP.md, the 15 hedged ones included. no replacement copy — the instruction was removal, not rewriting. the worst of them: "a signal worth waking you for is worth waking you for", a tautology sitting where a reason should be. "no account to sign out of." "eight hours is a slogan, not your need." a devices card restating the ladder directly above it, and a second card printing the identical string a row ten lines up already printed. "no sensor can see a pool wall", twice in one session summary. twelve rows would have left a card with a heading and nothing under it, so those cards went whole — and their sections with them where the card was the only content. that rule went too far in three places and i put them back, because the prose was slop and the affordance was not: - "allow location" on the gps-denied live card. denied is the ONE branch a user can resolve from that screen, and deleting the card left them with a missing map and no way to fix it. the other three branches kept theirs. - "add weight in profile" on workout. without a weight there is no calorie estimate at all, and that was the only place saying so and offering the door. - "3 of 7 nights so far" on the sleep comparison. that's a count, which the brief protected explicitly — it says the comparison is coming and when. all three are title + action now, no body. StatusCard renders no empty Text when why is blank, which makes "title and a button" a legitimate card rather than a broken one — that IS the minimal-text ideal. what a card must never be is a title alone. dead code out with the strings: kCalorieNeedsWeight, an unused profile import. _routeIssueCard, _versusUsual and _stateDetail are nullable now and Observation.advice is optional, so nothing renders an empty string. 124 goldens moved. one sleep assertion repointed: the section is present with its count, not absent.
the gallery only had about two thirds of the vocabulary in it. added the readiness card, the share card, the live session pieces, the settings / source / activity rows, legend, mono table, unlock contract and the import report. pulled ReadinessHero out of home_screen so it could go in at all — it was inline in a ListView behind the db. putting them in the sweep found four things overflowing at 3.1x that are live right now: home's readiness card, the live hr block, every settings row and the activity list. fixed. first attempt at the row ones used Flexible and quietly shifted every settings value right, so they use the row rule instead and the value drops under the title at big text. added a test so the gallery can't fall behind again — a new component has to be in it or on the not-a-component list.
the muscle map was never a measurement. it was the exercise→group table in catalogue.dart times the volume you typed, painted onto a body — which is the shape a scan produces. gone from the post-activity breakdown and from the share card too, since that one leaves the phone. painter and muscleLoad deleted; the group name stays as the picker label, which is all it ever honestly was. a lift now shows what it knows: volume, sets, top set, and every set as logged. empty session still explains itself. gallery: added the second state of every card — absent, over target, down and bad, a series with a gap in it, title-only status, zero and full consistency, every pill colour — plus a share card per archetype. that is the state a screenshot never catches because a demo phone always has data.
summary.dart already refuses this in words — the live ring is a pacer you breathe along with, so a finished session has no phase to draw and a static ring at .8 is decoration shaped like a measurement. the share card was drawing it anyway, and that is the card that leaves the phone. a flow session shares as the minimal card now, same as a lift.
running and gps sessions get a third share style. 4:5 poster, your photo across the top if you add one (from this phone, nothing uploaded, gradient if you don't), the stats you picked, and the route drawn on actual openstreetmap tiles underneath. the map is fetched and stitched rather than a live map widget — the card is exported with toImage, and an async tile layer exports blank about half the time. every tile is awaited before anything is drawn, so what you see is what gets sent. no tiles = the plain route on the card, never a half-loaded one. osm tile policy: real user-agent, disk cache, hard cap of 24 tiles per card, one zoom, and the attribution is ON the map so a crop can't separate them. tiles are tinted to the card rather than left as a white slab. no cadence, no weather, no training-load band — three things every running card shows and this stack doesn't measure for a gen4 band. carried real lat/lng through to ActivityResult.geo; the painters keep using the normalised box. tests pin the projection against london/sydney/null island, because a wrong constant there still draws a tidy route, just in the wrong country. the overflow test runs without the real font on purpose — it found one, in the footer.
you can't see the map without a run, so the gallery gets a real one: a lap of cubbon park with real coordinates, a load-the-map button and a photo slot. sits at the top of the gallery. not in galleryCases on purpose — every case in that map is shot by the goldens and swept at five text scales, and one that reaches the network would make both a function of the wifi. this is screen-only and fetches nothing until you tap.
i built a stack — photo band on top, stat strip, map at the bottom. the reference is side by side: photo full bleed across the whole card, scrim fading left to right, everything in a column down the left. so: wordmark, activity in accent caps, the big number, a VERTICAL stat list with ringed icons and hairlines, the map, date footer. route runs a pace gradient now, green fast to red slow, with real pins instead of dots. four stat rows, not three. three silently dropped the fourth and the share sheet preselects four — a card that quietly loses a number you ticked is the one thing this card isn't allowed to do. trimmed the map and the padding to make room. PosterStatRow is public so the workout screens can use the same row. gallery preview gets a share button that actually shares — you can't test sharing without a run otherwise. still no cadence, weather or load band. not measured.
…here you were right about the map. in the reference it isn't a boxed widget, it's dissolved into the picture with only the route and the streets around it showing. so it's a full-bleed layer now, on the photo and under the scrim, box and corner radius gone. the dissolve is a mask built from the route itself — a very wide, heavily blurred stroke along the path, applied as dstIn. what survives is the block either side of where you ran. over a photo a second mask fades it upward into the picture; with no photo it keeps the whole card, because then it IS the background. the line stays full strength on top: the tiles are texture, the line is the measurement. had to widen the scrim's hold to the full column — a route runs diagonally across the whole card and was crossing HEART RATE at .62. workout summary and history now use the same ringed rows as the poster. the three-across strip is gone, so the 4th and 5th stats stop being dropped by a layout and pace can print its own unit. three things found on the way, all live before today: - the summary said green = SLOWEST and the poster said green = FASTEST. same run, two maps, opposite meanings. summary now matches the card. - InlineMetrics ellipsized measurements: the journey card printed '+642…' for +642 m, health trends printed '78 / …' for 78 / 100. scaleDown now — a shrunk measurement is still the measurement, a truncated one is a different number. - the summary hero had a Flexible next to a Spacer, so two flex children split the row by ratio and a lift's volume printed '1,…' at 2x. ring scales with text now, clamped at 2x — it's on real screens, not just the fixed-size card.
the share sheet asked three questions and only one was mine to answer. style list gone — a lift never had a texture to pick. stat picker gone — it was asking which of my own measurements to leave off a card with room for all of them. what's left is a photo and a share button, plus post/story so instagram stops cropping it. the basemap was at zoom 2 on every card ever shared. the tile budget loop stepped the zoom down while the count was over its cap, but the frame is a fixed pixel size, so the count doesn't move with zoom — 900x1200 is 25 tiles against a cap of 24 and it ran to the floor every time. the corridor mask hid it, which is why removing the mask is what found it. then it was invisible instead. osm's raster has almost no dark pixels, so mapping 0..1 onto the two ends crushed land, water and roads on top of each other. stretch the window the source actually uses. hero and grid were both printing the distance now that nothing is hand-picked, so the grid drops whichever stat the hero already is. gallery: flows tab walks every activity in the catalogue through the real screens. its light tab was showing dark and its dark tab light, so every review through it was of the palette i hadn't picked. dropped powercurve and movementmap, nothing drew them.
the three that matter most: the share sheet fetched osm tiles. so opening it on a gps session sent that session's location to tile.openstreetmap.org, in an app whose whole claim is that nothing leaves the device — up to forty requests, undisclosed, on a screen you open to look at a picture. it's behind an explicit default-off opt-in now that says plainly what gets asked for, and without it the card still works: your photo, the route line, plain ground. "private session" was a toggle that did nothing. the flag never left the draft, so a session you'd been promised was hidden from summaries and exports was sitting in both. it's a column now. and the pin. kAlgoVersion spans both the pinned analytics and the local one, so a day derived against different analytics at the same version can never be invalidated — which is exactly why nobody noticed that v67 and v68 cite six analytics behaviours the pinned sha doesn't contain, in the paragraph that warns about v43. repinned, and the version test reads pubspec.yaml now, so the next drift is a red test instead of a shipped wrong number. the rest, by shape: fabricated numbers. active_min written twice by two algorithms over two windows under one key. a confident 0 active minutes from a day with no gravity vector, and 0 restless minutes from a night with no accel. wear coverage divided by the span of the records that exist instead of the length of the day. a 100% coverage invented from the existence of one hr sample. the wrong day. notifications stamped with the newest derived day instead of today, so importing old history fired an alert about a night in may. emitOncePerDay keyed its guard on the day the data came from and never checked that day was today. home named the held-over night on readiness only, so the sleep and heart-rate cards beside it still presented that same older night as last night's. forms that ate what you typed. a typo wiped a stored height. a comma turned a macro into a missing one. a lab result vanished with the dialog. every outcome on your data, thrown exceptions included, rendered as a green done. blank and unparseable are different things now, in the one helper they all share. things you couldn't undo, and things you couldn't do. a medication could be added and never removed. a period start was deleted by one unconfirmed tap. pause didn't pause anything but the clock — calories, strain, zone minutes, gps and max hr all kept accruing. and the migration ladder. rung 19 backfilled columns into a table it had just rebuilt without them, so an upgrade from schema 18 or older threw inside the one exclusive transaction and got the database quarantined. the test that should have caught it seeded a row whose hex doesn't decode. schema 40, algo 69, 2447 tests. the 20 summary_* goldens were already red.
schema 41. device_family lands on decoded_onehz, decoded_rr and sessions, stamped at ingest from the generation the link pinned at service discovery — never inferred from the data, never backfilled from the currently paired device, because what's paired today says nothing about which strap wrote a row in march. it carries into Substrate as ONE value for the window: a window that mixed two straps has no answer and resolves to unknown, and unknown refuses rather than quietly reading as gen4. that refusal is the point. every existing install reads NULL until it next syncs, so a metric moved onto the seam goes absent for historical days. that's visible and it's correct — the alternative is one strap's calibration silently answering for another's. it also found a live one on the way in: naming the new column unconditionally in the insert map bricked the v6 upgrade, because _backfillDecodedStore runs mid-ladder long before the rung that adds the column, and a throw inside onUpgrade's single exclusive transaction quarantines the database. the write path only names the column when it has something to say now. sleep_detail's stage catch-all was `_ => SleepStage.light`, so every second the band never recorded was drawn AND tallied as light sleep — a three-hour hole came out as three hours of sleep. unobserved is not a stage; it draws as a gap, and an unrecognised label from an older bundle fails the same way. the fertile window is gone. it was mean_cycle_length minus 14 — a textbook constant printed as her personal date with "not contraception" underneath. the prediction is a median-and-MAD range with n stated now, and it is embarrassingly wide for irregular cyclers, which is the finding. vo2max and fitness age deleted. the water buzzer deleted — complete, armed on every launch, 24 real notification slots, held off only by a pref with no toggle. water_ml stays as a passive logged signal. spo2 is a permanent refusal with the ir-red reason written down, and the "temporarily disabled pending hardware-verified decoding" strings are gone; they were a promise to a future contributor that it comes back. then the things that were already computed and never shown: day strain off the curve that was persisted for nobody, the readiness absence diagnostic that only ever went to a firebase breadcrumb, deceleration capacity, the rhythm screen's history, daytime hrv with the motion gate it never had, sleep onset on forced windows only, wake-up runs, the caffeine timing that was dropped three lines before the analysis. verified against real gen4, gen5 and MG exports — all three are old-schema, all three migrate to 41 with zero rows dropped and integrity ok. 2503 tests. the 20 summary_* goldens were already red and are now stale for a second reason. algo bump and the repin come after the next batch lands.
the biggest one is the smallest diff. decoded_onehz.hr was NOT NULL, and 0 was the off-skin sentinel, which is why every v25 record — 28,395 of them in a two-week gen4 export — went to raw_archive instead of the substrate. the column is nullable now and nothing was turned into a null on the way: hr IS NULL is zero rows on all three real databases after the migration. the audit was the work, not the alter; the risk was a reader treating missing as 0 and scoring those seconds off-wrist. three universal 220-age paths were still handing out a heart-rate ceiling while hr_max refused for an unknown strap, so the same session could band zone_min on one ceiling and anchor its trimp on another. one function now, dispatched by family, and where no ceiling can be established the session says so instead of inventing 190 for anyone who skipped the age field. at 30 the ceiling moves 190 to 187, which moves band membership — 150 bpm was z3 and is z4. that is the bug, not a regression. 'fat burn' is 'easy'. which fuel you oxidise in z2 needs respiratory exchange and no wrist sensor produces it. mid-sleep is persisted, unwrapped, because 23:30 to 01:30 is +2h forwards and -22h numerically and a change-point detector run on the wrapped series finds a modulo artifact. only future nights get the key. hr_valid reaches Substrate. the ambient-light channel finally has a producer — one line, and 0 still means absent, not darkness. pages_behind writes to the table that was created for it instead of a single ledger row that only ever held the last reading. the coach stops being told about odi_per_hour, a column that can never have data again. encrypted backup says both halves in the same breath: unreadable without the passphrase, and gone forever if you forget it, because there is no server holding a key. no hint field, no recovery code — both weaken it to feel safer. GATES.md answers four questions that were marked blocked on hardware nobody had. the headline: gen5 skinTempC is populated on 1,035 of 1,035 real records, 28.8-38.3 degrees, 397 distinct values, 98-100% coverage inside real sleep windows. the coverage gate that was supposed to kill four features didn't. the refusals are in the code now, at the sites someone reaches for them, rather than in a planning doc nobody opens. repinned to analytics 391ede4. 2598 tests, the 20 summary_* goldens still red from before all this.
schema 44. a lot of correct analytics landed last wave and then stopped at the repo boundary — the functions existed, nothing called them, so nothing computed them and the ui agents had nothing to render and rightly refused to fake it. this closes that seam. the observed hr ceiling scores sessions ONE AT A TIME rather than as a concatenated day, because stitching two sessions invents a hold across the gap between them. it needs the 15s hold plus corroborating motion, and the motion floor is per-family: gen4 0.10g, gen5 0.04g, unknown refuses. sri was going to ship an empty list forever. the pairs were serialized, but the analytics floor is half a clock day and this caller's valid mask is only set inside a hypnogram segment — an 8h night marks about 480 minutes, so no pair could ever clear 720. it would have rendered as "no irregular nights" for everyone. the published sri number doesn't move; every accepted epoch still counts. the day breathing curve gets its stillness gate, and the measurement is the point: gen4 attempts 1,573 windows and 201 survive, gen5 617 and 552 survive. so on gen4 the triple lomb-scargle now never runs on 87% of candidates. that also means resp_day changed meaning — it's the day's resting floor now, and it will be empty on most days, which is the honest output. nightly hrv shape runs over the SAME cleaned nn the headline rmssd uses, so the curve and the number can't disagree, and every bin ships lo/hi so it draws as a band. session cost and the overreaching conjunction read outputs that already ran daily. the conjunction is two facts and no verb, and it is structurally kept out of the notification collector rather than just left out of it. and 867 seconds of real MG data came back. 1,035 v18 frames were archived as undecodable before this month's decoder fixes and decode fine now; the re-drive is idempotent, verified by running it twice on the real export, and it never consumes or relabels the archived bytes. the other 168 already had a row and the standing row won. MT-05 stays refused, and now it's refused with a number: the minute-majority split eats 17-21% of a labelled 83-minute walk and 22% of a labelled 50-minute run, measured against real sessions. those are exactly the minutes it would have subtracted from active energy.
home + home_cold: the sparkles button is gone, goldens have no coach config so it reads as unconfigured. these two were already red at 1x before i touched them - something in grammar.dart moved and nobody reshot them - so this picks that up as well. metric_detail: opens on today now. percentile test taps through to 30 days, the rank is a history thing and today has no stored point in that fixture.
workouts on workout > history, height and weight on edit profile. both were three taps deep under more settings next to a database export. second tap says refresh instead of import — the table is keyed on the store's uuid so re-reading the whole window updates rows instead of stacking copies. can't ask ios whether read was granted, so the word comes off our own record of a read that actually returned something. permissions asked on the tap and per set: profile types on the profile button, workout on the workout one. nothing at launch. imported sessions now show up in history in date order with the watch or app that recorded them on the row, and count in "this week". still out of weekly load, ctl/atl, records and every baseline — there's no heart rate trace behind them, so a load number would be made up. said on screen where the two numbers sit next to each other.
home's sparkles is conditional now and the overflow row is gone, so a fresh install had no way to set it up at all. this is that way. reads the config through a try/catch like coachReady does — the golden harness mounts screens without providers and a throw there is a red test about the harness, not the screen.
it printed two words and stopped: "hrv", "rhr". two things were wrong. it read readiness_glassbox.drivers, which is already filtered to the inputs that cleared the swc gate — so an input sitting inside its usual spread was never in the array and the screen couldn't say "and your breathing rate did nothing". reads breakdown now, same array readiness detail renders. and nothing said what the number was. the reading, your own usual, the direction, the size, whether it beat the noise — all of it already in the baselines block on every derive, read by nothing. skin temp shows no numbers, not just no chart. it's a raw adc and "32411, 31 above your usual" is arithmetic nobody can use.
…e from "a baseline, not days" over two paragraphs about a centre and a spread is not a section title. it's "resting heart rate" now, two sentences, and the seeded row reads "usually 54-62 bpm" instead of "spread 4.2". the two facts underneath are unchanged: nothing uses it yet, and it never becomes a reading of its own. also took the import row off profile home — height, weight and workouts already moved to the screens they fill, and what's left isn't quick access. one door, under your data, and more settings names it now. and every nerd stats door dropped the day. you'd step sleep back to tuesday, tap through, and land on last night — same numbers whichever day you came from.
the png had no transparency at all — a flat (249,249,249) plate at alpha 128 over the whole canvas, so dropped on the purple card it was a grey square. flood-filled from the edges instead of keying near-white, because the character is cream and a colour key eats it. opaque pixel count is unchanged, so nothing got cut off it. 279kb -> 42kb across 1x/2x/3x. card is full width, mascot bottom right, copy and start on the left. it's a row, not a stack — stacked they fought for the same pixels. height is a floor now, not fixed, or the button clips off at 2x text. two grounds instead of one purple: a cream mascot disappears on a pale card, so light goes deeper rather than lighter. rings draw solid when measured. calibrating keeps the fade on purpose — if everything's bright you can't tell "this is real" from "this is still filling".
the overflowbox did it. inside a listview the main axis is unbounded, an overflowbox takes whatever it's handed, and it took the whole list down with it. sorry — that shipped without me rendering it in a scroll view. full bleed is the list's job now: it drops its side padding and pads every other child, so the hero is just the one child that doesn't get any. ordinary layout, can't blank anything. card moved to start_card.dart and parameterised so wellness can use the same one with its own mascot and colour.
the method prose was sitting on the screens themselves — how karvonen works, what the alertness model can't see, what prsa averages, why cold shuts the sensor. gone. kept the absence reasons, the one-clause honesty lines and the medical copy, and used plainer words for the rest.
the meditating mascot came in already transparent, so it just needed cropping. 810kb -> 30kb across the three densities. teal because it's C.domMind, the tab's own colour, and the coral cushion and headband sit against green better than they do against purple. count is 3 — kBreathPatterns, what the picker actually offers. the paced breathing ActionCard is gone; this is the same door with the same last-sitting line on it.
Expanded and Flexible both default to flex: 1, so a row with one of each split it 50/50 — that's why "see all" started in the middle of the line and stopped wherever its own text ended. same shape in three shared components, so it was everywhere at once. fix is spaceBetween with both sides shrinkable. making the right side a plain child anchors it too but overflows on a long left side — day strain caught that at 15px. and the mood buttons: a container with a child and no alignment sizes to the child, and the constraints coming down are loose, so they rendered as narrow pills instead of the fifth of the row the Expanded had paid for.
five sub-tabs never fit. on a 390pt phone Cycle shows 3pt and Labs 19pt, on a 360 both show nothing at all, so the row reads as if it ends at the fourth one — which is how the cycle and labs screens went unfound. ScrollHint fades the tail and puts a chevron at the edge, both only while there is genuinely more, both gone at the end of the scroll. wire it into SubTabs separately.
they're sweep reports and measurement runs written against the real exports, so they quote dated readings — a resting heart rate on a named day, an irregular-rhythm flag with its date. this repo is public. they stay on disk. this only takes them off the tip: they're still in 24 commits of branch history, which has to be purged before any push.
it printed the onboarding weight scalar and that was it. the ask was a bar saying whether your weight is normal for your intake and burn, and i can't build that one honestly: - no food log in any real db i have, and a logged day is partial by design anyway (an occasion with no kcal is a valid entry, so the day's intake is a floor). self-report is under by 20-30% on top. - calories_total is an estimate at 0.5 confidence and swings 2454-4545 across a fortnight on my own export. a barely-worn day still prints a confident 1715. - weight here is one profile number, not a series, so there's no usual to compare against. the honest trend already exists in journal (weightTrendEwma, gaps left as gaps). reason is in the code where the section was so nobody rebuilds it.
the fifth tab is off the edge on every phone we ship to — at 360pt it's fully off-screen in both tab sets, and above 1.0x text every set overflows even a 430pt screen. can't be made to fit. hint draws nothing while the row fits and scales with what's left to scroll. icon is the colourful one now, bw set alongside it, and the launch screen stops flashing the old mark before dart starts.
416 pngs, 27mb, none of them on main. they're machine-specific — the two sdks here already disagree on antialiasing — so committing them is a diff nobody can review and a conflict nobody can resolve. test code stays, pictures don't. regenerate with --update-goldens. still in branch history, purged along with docs/internal before any push.
|
Important Review skippedToo many files! This PR contains 233 files, which is 133 over the limit of 100. To get a review, reduce the PR to 100 files or fewer by splitting it into smaller PRs or changing its base branch. Upgrade to a paid plan to raise the limit. This review couldn't start because sufficient usage credits or metered capacity aren't available. Add credits or update usage-based reviews in the billing tab, then retry. ⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Plus Run ID: ⛔ Files ignored due to path filters (264)
📒 Files selected for processing (233)
You can disable this status message by setting the Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
PR Reviewer Guide 🔍(Review updated until commit 85d6490)Here are some key observations to aid the review process:
|
PR Code Suggestions ✨Latest suggestions up to 85d6490 Explore these optional code suggestions:
Previous suggestionsSuggestions up to commit 1ab38b0
Suggestions up to commit 2125dc4
Suggestions up to commit d03e4f9
|
pub get with pubspec_overrides.yaml sitting there rewrote the tracked lock to ../protocol and ../analytics, so ci died at the pin guard before it ran a single test. same thing the guard's own comment says happened twice in july. pins themselves were already right. coderabbit refused the pr outright — 436 files against a 100 limit — so .coderabbit.yaml drops test/, the platform dirs and assets/ and points it at the ~230 files where a bug can actually be. also the pin comments still said UNPUSHED and said the analytics audit wave was uncommitted. both stopped being true last night.
|
Persistent review updated to latest commit d03e4f9 |
it's the one sub-tab that's irrelevant rather than empty for most people, and an empty tab that can never fill reads like a feature you failed to use. profile > preferences > cycle tracking. off hides the tab and skips its query, keeps everything already logged.
ci went red on every golden after the pngs came out of the repo. skip with a reason instead of an --exclude-tags flag, so the run says out loud nobody checked the pixels. the 61 non-golden tests in the activity file keep running, which a file-level tag would have killed.
zz_scratch_render_test.dart says 'deleted before commit' on line one and then got committed. analyze runs --fatal-infos so the two infos were the whole reason the test job was red — it never reached a test.
|
Persistent review updated to latest commit 2125dc4 |
runHeadlessSync self-acquires its BandOwnership lease when the three iOS gate entry points call it with no lease argument, so only that call frame ever holds the token. If the run truly wedges rather than just running long, HeadlessSyncGate's timeout hands the gate back to the next wake but leaves the band lease itself stuck on an orphaned future that never runs its own release. Every later headless wake then silently no-ops, and a foreground connect spins forever in acquireForeground's wait loop with nothing left to complete it. Force-clear the headless owner from the gate's own timeout handler, matching what the boot path already did explicitly for the lease it holds directly.
The pin was still sitting on the PR branch head. PR #46 merged it to main as bfea5e5, two commits ahead of the old pin — both test-only deprecation-ignore annotations, lib/ untouched, so no kAlgoVersion bump needed for this move.
|
Persistent review updated to latest commit 1ab38b0 |
|
Persistent review updated to latest commit 85d6490 |
1 participant
User description
rebuilt entire ui
PR Type
Enhancement, Bug fix, Tests
Description
Complete UI replacement:
lib/uiremoved,lib/ui2introduced with full metric fidelity (confidence, tier, absence handling)Metrictier parsing ('AUTH'not'AUTHORITATIVE'), Poincaré row rendering, flat-series chart floor, and gen5 step counter wiringsessions.avg_hr;crossDayArtifactUsableTodaynow respectsalgo_versionSchema bumped to v38,
kAlgoVersionbumped to 66New
HrSensorLink: GATT Heart Rate Service (0x180D) support for session-scoped external HR sensors, stored in newexternal_hrtable, never merged into baseline analyticsLiveDraftpersists session state (sets, laps, score, pause clock) to Prefs so minimising or killing the app cannot lose typed workout dataDiagram Walkthrough
File Walkthrough
14 files
New live session screens with honest absence and draft persistenceNew GATT Heart Rate Service link for session-scoped external sensorsShared UI grammar primitives for ui2 screensDesign tokens and theme for ui2Rebuilt home screen wired to full metric fieldsHealth screen with confidence-aware metric rowsReadiness detail screen with tier and confidence displayGeneric metric detail screen for trend and confidenceActivity summary screen with retry-save and enriched resultActivity catalogue and archetype definitionsSchema v38 migration: external_hr table and sessions.avg_hr columnNutrition store for food logging persistenceOff-wrist lookup helper for substrate filteringCoach action definitions for AI coach integration4 files
Chart widgets with correct flat-series and absence renderingSleep detail screen with honest stage absence handlingWorkout screen wired to sessions.avg_hr and full resultkAlgoVersion bumped to 66; crossDay artifact algo_version fix2 files
Tests for live activity screens and draft persistenceWiring tests for ui2 metric field propagation101 files