What’s new
in RxDaily.

Version 0.83.4

Version 0.83.3

Version 0.83.2

Version 0.83.1

Version 0.83.0

Version 0.82.0

Version 0.81.0

Version 0.80.2

Version 0.80.1

Version 0.80.0

Version 0.79.0

Setup is a step shorter, a medicine's timing is chosen with one tap, and Care shows your clinician and clinic as one.

• Setup no longer stops to choose a scene.
• Choose how often you take a medicine from chips on the form.

Version 0.78.0

The clinic's plan reaches Today, a tour points at the real screens, and the content studio is tidy, validated and runs on Payload 4 beta.

• What the clinic sets up, readings to take and plan items, appears on Today.
• A first-run tour lights up the real screen, one step at a time.
• The content studio uses plain names, and quizzes and scenes now live there.

Version 0.77.0

allie opens with your day, the clinic sees every high reading, and administrators read a clinic without typing a reason.

• allie opens with one short line about your day and brings back your last conversation.
• Needs a look checks the readings of everyone in active care, typed or from a cuff.
• Pick any period on a patient's progress, from 30 days to a custom range.

Version 0.76.0

The admin and clinic consoles are redesigned on one system: RxDaily color and glass, borderless cards, figures as cards, and a native bottom bar on phones.

• Both consoles stand on one design system: a colored ground in RxDaily's palette with glass over it, white or night cards with no borders, and one icon family.
• On a phone the consoles have a bottom bar with three destinations and a More sheet with every destination as an app tile.
• Each page leads with its key figures as cards, and says far less.

Version 0.75.5

allie answers every emergency with the emergency response, and the monthly billing build runs again.

• A message about self-harm or an emergency always gets the emergency response and its call card, also when it states a high blood pressure.

Version 0.75.4

allie's message field works on iPhone.

• On iPhone you can type to allie and send a question.

Version 0.75.3

A natural voice on the phone, and never the robotic one.

• Lessons, sessions and allie are read aloud on the iPhone in the natural voice.

Version 0.75.2

A picture of its own for every Explore read, session, program and recording.

Version 0.75.1

The whole app distilled to the essentials, a health book you can explore, and clinical data that only the right people read.

Version 0.75.0

A redesigned app, medicines that count at once, safer clinic invitations, and a clinic that sees only what it should.

Version 0.74.18

Privacy and safety: attachments deleted on schedule, a covered voice only, and quieter logs.

• Photos and documents sent to allie are deleted on their schedule every night.

Version 0.74.17

The RxDaily brand book at /brand, and the server side of error reports from the iPhone app.

• The brand book: the wordmark, color, type, voice and motion, drawn from the product itself.

Version 0.74.16

Readings from Apple Health, an image-led first-run guide, and a moving ocean at sign-in.

• Blood pressure, weight and glucose come in from Apple Health, labeled as such.

Version 0.74.15

The news desk writes a long read every day, and the iPhone build goes to TestFlight with Apple Health.

• News stories are long reads that go to the paper behind the news and say how sure we can be.

Version 0.74.14

Practice sound in one tap with Lauren's breath cues, and one natural carousel everywhere.

• Lauren says breathe in, hold and breathe out.

Version 0.74.13

RxDaily's own icons and pictures are back, and Explore opens on Walk more and the first sessions.

• The living tab icons and allie's orb are back.

Version 0.74.12

allie's Save reading works in the browser again.

• Save reading on allie's cards works again.

Version 0.74.11

Walk more asks its talk test aloud, and content pictures arrive smaller and faster.

• Content pictures arrive at the size they are drawn.

Version 0.74.10

The admin Enrollment desk opens about three times faster.

• A faster Enrollment desk.

Version 0.74.9

Walk more speaks its offer for next time, first sessions open on their own picture, and the health book's chapter lines read as sentences.

• Walk more says its offer for next time aloud.

Version 0.74.8

Lauren's first sessions are in the app, opened from a row in Explore.

• 22 first sessions in Lauren's voice, one tap from Explore.

Version 0.74.7

Walk more for everyone in Lauren's voice, a cleaner setup, Today, Explore, Health and measure screens, and an allie that answers from your health book and plan.

• Walk more, in Lauren's voice, for everyone.
• Setup puts your medicines first, and the tour is three stops.
• allie answers from your health book, your medicine times and your plan.

Version 0.74.6

Food quizzes, an allie that saves the readings you tell it and draws its cards from your data, When to get help in Health, and Lauren's voice across the guided sessions.

• Five food quizzes with a warm word for a right answer and no score.
• Tell allie a reading and confirm it to save.
• When to get help in Health, for everyone.

Version 0.74.5

A Getting started card on Today shows a new person only what applies to them and ticks itself as they go.

• Getting started on Today: medicines, a first reading of your own measure, your meal plan and one thing for your day, in any order.

Version 0.74.4

Explore opens with fewer requests, a due medicine leads Today, allie speaks in US units, and the clinic sees its people and its own time zone first.

• Explore makes 6 requests on open instead of 9, and 2 when you come back within a minute.
• A due medicine is the first thing on Today.
• Clinic times in the clinic's own time zone.

Version 0.74.3

Setup is clearer to get through, the app opens with a short guide, and RxDaily's emails look like the app.

• A first-run guide that points at each part of the app once.
• A consent step with a short summary and one I agree and continue button.
• Date of birth fields that cannot be missed, and every error beside the field it is about.
• Emails with the real wordmark and the app's photographs.

Version 0.74.2

Joining from an invitation takes five screens instead of nine, and the clinic patient page opens on what the person just did, with times.

• Accepting an invitation is five screens: consent, about you with your password, confirm, a short welcome, and your first item.
• The clinic patient page starts with Recent activity, so a dose, a reading or a practice shows with its time as soon as the person logs it.
• The CMS shows what an article still needs before it can be published, instead of refusing one requirement at a time.

Version 0.74.1

Breathing practices from the library run for their full length again.

• A breathing practice opened from Explore now breathes with you for every planned breath, instead of finishing the moment you press Start.

Version 0.74.0

allie's answers about your readings now speak in words and show your week, programs open with a cover and a Join you can always reach, and Health shows your own usual range. For clinics, the panel opens on the whole clinic for owners and admins, the patient page shows logged doses and blood pressure with its range, and the consoles read more clearly.

• allie answers "Blood pressure 126/81, this morning at 8:11, you entered it" instead of codes and timestamps, and shows the week when you ask about the week.
• Programs open with a cover picture, a preview of day 1 and a Join that stays at the bottom of the screen.
• Health draws each measure across your own readings with the middle half of your values as a band, so you see your usual range, and daily medicines say "Logged on 24 of the last 28 days".
• Clinic owners and admins open the panel on the whole clinic, and the patient page shows logged doses per day and blood pressure as two lines with the person's own range.

Version 0.73.0

One player now runs every guided session. Meditations and other recordings play over a moving scene with Lauren's words filling in as she speaks, breathing and exercise share one footer with Stop in the same place on every step, and every session you finish or stop is saved to your record. Clinics can create patient invitations again.

• Meditations, body scans and program day recordings play full screen over a moving scene, with the words filling in turquoise as Lauren speaks and staying on screen through her pauses. One tap hides them for eyes-closed listening.
• Breathing and exercise share one player: your action on the left, Pause in the middle and Stop on the right of every step, rests you can stretch by 30 seconds, a short safety list that waits for Ready, and a check-in you answer by choosing a row.
• Every session you finish or stop is saved to your record with the minutes you actually practiced, and a bedtime session that plays to its end is counted without a prompt or a sound.
• Clinics can create patient invitations again.

Version 0.72.3

Phone cards stack cleanly and Read aloud shows preparation and buffering. Background book updates now exclude old non-patient accounts without a health record.

• Activity suggestions, setup starters, scene choices and Today cards stack on phones. Browse the library keeps its compact grid.
• Read aloud responds immediately, keeps loading visible until audio is ready, and lets you cancel or retry.

Version 0.72.1

Today, Explore and allie's program suggestions now use the same service. Each suggestion explains why it was chosen, and your recent choices help avoid repeats.

• Suggestions follow your recorded choices and recent activity, with a reason for each offer.

Version 0.72.0

allie now follows the conversation and sees your recent readings and labs, the health book carries its personal letter and sources with titles on its pictures, and every suggestion has a picture of its own. News stories are longer and told as stories, the conditions list reads again when an onset is a year, and the app can now carry guided program days and meditations.

• allie follows the conversation instead of treating every message on its own, sees your recent readings and labs, and explains a reading against your own history.
• The health book puts each title on its picture, says what every chapter is for, and carries your personal letter, section headings and sources.
• Every suggestion has a picture of its own subject and a line written for it, each medicine has its own picture on Today, and Explore's library shows each icon beside its name.
• News stories are longer and told as stories, built from the original research where it can be read, with at most one link to another site in three.

Version 0.71.0

RxDaily is organized into Today, Health, Explore and allie, with three navigation styles and icons that respond to the day. Medicines are grouped by time of day, visit preparation brings questions and recent readings together, and chat, news, reading and exercise players have been redesigned. For clinic and admin teams, the update includes ACCESS enrollment and monthly billing implementation previews, an adapter-based billing architecture, and verified eligibility, enrollment and baseline acceptance in the CMS implementation environment.

• Four sections: Today, Health, Explore and allie, with Glass lens, Dock and Island navigation styles available in Settings.
• Medicines grouped by time of day, a dedicated page to prepare for a visit, and a redesigned allie chat with more room for the conversation.
• Redesigned Explore carousels, news cards and reading, with new exercise players and Read aloud controls that follow the spoken text.
• ACCESS enrollment and monthly billing implementation previews, shared billing checks with input and delivery adapters, and an end-to-end CMS implementation-environment test.

Your numbers. Your days.

RxDaily has a public front door, and a person's own numbers became something to read. rx-daily.com opens on the app's own scenery with a live Today card, and a clinics page explains how a clinic works with RxDaily. In the app, each measurement opens a year of the person's own readings, Your days shows every day as one square, and every insight is held to one rule: it may say what a pattern could mean in general, never what it means for this person. In the clinic, the care panel is ordered by rules a clinician can check. On the 25th the app took a new shape around it: onboarding stays in a person's own scene, Today opens on the week with the next meal's own photograph, recipes and reading got proper layouts, and the old web patient app left the codebase.

• Your numbers: each measurement opens a week, month, six months or a year of the person's own readings, with the average and its reading count, their usual range as a band, and where every reading came from.
• Your days: one square for each day, shaded by how much was recorded compared with the person's own busier days, filtered by medicines, habits and readings. A month is a calendar whose days open with a tap, and Today's week card opens it.
• Insights that know their limits: insight text is held to decision 0051 in one check. A pattern may be described, never a verdict; a Health Book chapter a model wrote opens with "Written by allie, an AI assistant"; an analyzer's signal is never an alarm to the person.
• The care panel, ordered by the clinic's rules: each person once, with the reason on their row: changes waiting, higher than their own usual this week, a week that will not count for ACCESS, nothing recorded in 5 days, steady.
• rx-daily.com is the RxDaily landing, with a clinics page. A visitor without a session meets the app's own ocean scenes and a Today card that ticks its items off; a signed-in person opens straight into the app.
• Onboarding in a person's scene: sign in, starting on your own and every step of a clinic invitation happen over the person's own scene, on one reading sheet. The invitation is a card that fills in with the name and date as they go, and the welcome offers the invitation first.
• Today opens on the week: the week and its recorded days sit at the top, today's mark fills as the day's items are answered, Up next shows the meal's own dish photograph, and a finished part of the day says what was recorded and what comes next in one line.
• Recipes and reading: a recipe opens on its photograph with nutrition per serving, its ingredients and steps as a checklist, and articles and news share one reading layout with a line on how confident the source is.

A safer allie, measured.

Allie's safety net was measured, found wanting and fixed the same day. The second crisis layer had returned nothing since the move to Vertex; it now runs on a model that answers inside its budget, and one widened vocabulary catches what the rules missed. An evaluation model now checks the places where the platform makes a judgment. Progress became something both sides can read: Your health in the app, and in the clinic a roster by condition, a progress panel on every record and a progress page with every reading against the guideline.

• Crisis detection, measured. On 29 synthetic crisis messages the pipeline escalated all 29, 70 of 70 messages were handled correctly with no false escalation, and the second layer answered in 470 ms at the 95th percentile against a 1.5 second budget.
• One crisis vocabulary for the distress gate and the query router. Escalation of the same crises rose from 13 of 29 to 28 of 29 before the second layer, and idioms no longer raise false alarms.
• Your health, first in Me: a person's conditions, readings with their charts, medicines as the days a dose was logged, and how they have been feeling, in words.
• The clinic roster by condition: one filter per condition group with its count, a conditions column and a search, each person's last fourteen days of entries drawn on their row, and the clinic's ACCESS standing above the list.
• Progress in the clinic: each record opens on a panel of the last 90 days, with every measure's latest reading and source, 28 days of doses per medicine and questionnaire scores. Open progress shows three or twelve months per measure against the guideline, every reading with its source and whether ACCESS counts it, and the target the clinician sets, which the person's app then shows.

The roster, filtered by condition.Progress on the record.Against the guideline.Fasting and after a meal, apart.

One console, rebuilt.

The console, one tool since v0.64, moved onto one component kit: every clinic and operator page in light and dark, and three audits of the result fixed what they found. A clinician can set a patient up before the person arrives, a person can bring their own record into setup, and Explore and Today became one loop.

• One console: eight admin destinations and four clinic destinations behind one navigation, with tabs above the content, data tables, fields and charts that behave the same way, in light and dark.
• Set a patient up before they arrive. Whoever invites a patient can enter their medicines, conditions, allergies, preferences and plan; it is held with the invitation and written through the clinic the moment the person accepts, so their first day is already made.
• Explore and Today are one loop: a first run ends with a real item on the day, and a practice or recipe chosen in Explore lands on Today.

Bringing daily care together.

The first release after the move to Expo, with 34 changes across the app, allie, the health record, clinic, admin and content tools. Explore brought the published library into the app, allie and Today started reading the same daily plan, readings became easier to add, and content gained review before publication.

• Explore brings published articles, recipes, activities, programs and audio into the app, with search, saved items and recently opened content.
• Allie and Today draw from the same daily plan, and a suggested habit can be reviewed and edited before it is added.
• Blood pressure, weight, pulse, blood glucose, blood oxygen and temperature can be entered with when they were measured, and a failed save says so.

Version 0.66.0

RxPulse became RxDaily. The patient app was rebuilt in Expo and React Native for iOS, Android and the browser, with a simpler structure shaped by testing with older adults: platform tabs and sheets, gesture-driven task actions, your own to-dos and habits, moving scenes, and an allie closer to the day you are having.

• A native app in Expo, with platform tabs and sheets, system pickers, swipe actions and haptics, and a browser export alongside it.
• Your day can include your own life: personal to-dos, repeating habits, editable suggestions from allie, and activities that keep their link to the program they came from.
• A new name and identity, with a live brand book built from the same tokens and components the app uses.

Allie, on every screen.

Allie became something you can see and talk to from anywhere in the app, reacting to the conversation as it happens. Around that, the app stopped making claims it could not support: home dropped scores and streaks for one next action it can name a source for, profile stopped asserting readings it did not have, and the copy across the app describes what is on the screen rather than how it feels. Separately, for the accounts that hold more than one workspace, there is now an explicit choice instead of a guess.

• Allie is available throughout the patient shell and uses the same living Cloud renderer wherever she is intentionally shown. She contracts while you speak and swells while she answers; closing her returns you to where you started with the thread intact.
• The app stopped saying things it could not back up. Scores, streaks and invented profile insights are gone, numbers carry their source and date, and an unknown is no longer drawn as a zero.
• For an account that holds more than one workspace, a patient record, a clinic, or platform access, the app now asks which one you are in rather than inferring it. Most people have one workspace and will not see this. It matters for staff accounts and for the consent-first journey RxDiet can use to invite real patients, which ships switched off.

Three tools became one console.

Admin, clinic and practice were three separately invented tools with no shared sense of where you were or what you were allowed to do. They are one organization-aware console now. Operators get the ACCESS, consent, submission, deadline and pipeline views in a single frame instead of hunting across separate tools, and a provider working across two clinics can move between them without patient identifiers following along.

• Admin, clinic and practice share one frame that always knows which organization you are in and what you may do there.
• The ACCESS, consent, submission, deadline and pipeline views operators needed are in one place. The named PHI-bearing paths use strict audit writes; universal read coverage remains an explicit hardening item.
• A provider working across two clinics can switch between them, and nothing identifying a patient crosses from one to the other.

Who touched this record, and when.

The operator side of the platform got its surfaces. The access log can answer "who opened this patient's record", and export a bounded, complete result when someone formally asks for it. Consent, HIE fetch, ACCESS submission, PCP notification, webhook and nudge paths that previously had no operator view have pages now. The console remains off by default, and vendor/CMS integrations keep their own go-live gates.

• The access log is readable, and a request for an accounting of disclosures can be answered with an export rather than a database query.
• Consent, HIE fetch, ACCESS submissions, PCP notifications, webhook delivery and nudge dispatch each have an operator surface, including honest empty, dry-run and unmet states.
• Every console page announces what it is and which organization it belongs to, and a slow one shows a skeleton instead of a blank screen.

Version 0.62.0

Two days, 27 commits, 235 files: the interface caught up with the machinery underneath it. Generic icon cards are gone from every patient surface, replaced by one image-led editorial system with real photography, porcelain surfaces and readable glass panels. Today is rebuilt as white period cards on a warm clay canvas, one explicit action per task, no hidden gestures. Explore is one calmer surface with Learning and Community each in their place, and enrolled programs moved into Plan where the day actually happens. And Allie got her voice back, literally: sessions start only once the live connection is confirmed, the conversation runs on one continuous canvas, and she can guide you to every patient-facing part of the app.

• Every patient surface moved from generic icon cards to one image-led editorial system. Home, health, nutrition, medications, goals, community, learning and care all draw from the same visual language now, with real photography, consistent Geist typography and accessible image descriptions.
• Today is rebuilt. Each period task is its own white card on the warm clay canvas, with the task image faded into the background and one explicit action at the edge. Hidden swipe-to-complete gestures are gone; everything you can do is visible.
• Allie answers reliably again and her voice runs on one continuous conversation canvas, with her orb stepping back into the header when plans or health information need the space. She can now guide you to every patient-facing part of RxPulse and show cards for progress, breathing, movement and home recommendations.

Version 0.61.0

The last few releases built machinery: outcome tracking, proactive reminders, caregiver and clinician plumbing, photo capture. v0.61 is the first full pass on the front of the app that machinery serves. Circle folds into two panels behind one toggle, and reading is the default. A caregiver can hold a revocable, audited view of the day. A photograph of a cuff or a medication label becomes a confirmed reading. And the small things that decide whether a phone app feels usable, tap targets, dead ends, mislabeled cards, the word "quiet", were fixed one by one.

• Circle is now two panels behind one toggle. Catch up is what the circle has been up to; Share is where you add something of your own. Reading is the default, so you can catch up and leave without writing a word.

One generation up.

Every layer Pulse stands on moved up a generation at once: Next.js 16.3, React 19.2, TypeScript 7 with its new Go-based compiler, and version 7 of the toolkit behind Allie. On their own those are version numbers. Together they are the groundwork for the things we want to build next, from instant page transitions to real spending limits on Allie rather than estimates. The landing page now scales with the window instead of snapping between two fixed layouts, the breathing exercise lost a distraction and gained a sound that behaves like breath, and a clinical data adapter that would have failed in front of a patient now fails safely at startup instead.

Pulled out by the root.

Green finally stopped coming back. Mint had been chased out of one surface after another for two releases, and it kept returning, because it was only ever patched per component and never removed at the source. This release fixes it at the token root: the accent family rotates to the warm clay the app already anchors on, the two color tokens that were referenced everywhere but never actually defined become real, and the news categories stop shipping two greens that could not tell health apart from environment. Alongside it, the housekeeping that had been deferred: Next.js patched for four high advisories, production audit highs back to zero, patient report pages taken off static prerendering, and a build pipeline that had been red for two weeks made green again.

• Green at the root, not the leaves. The Quartz theme had been overriding the accent to a deep mint across 162 consumers, the chip-and-well ramp was mint (which is why recipe sections, news cards, ingredient lists and loading skeletons all kept a tint), and the focus ring, selected state and accent glow were mint as well. All of it rotates to warm clay, holding lightness per step so every contrast pairing still passes. Allie's orb keeps its own identity tokens, deliberately untouched.
• Two tokens that never existed. A dozen call sites referenced `--pulse-accent` and `--pulse-surface-2`, neither of which was ever defined, so the hardcoded mint fallback won every time. Both are now real tokens on the warm ramp, which is the other half of why green kept reappearing after each sweep.
• ACCESS outcomes open with an answer. The card now leads with a ring showing how many tracked targets are on track, reusing the same gauge instrument as the health score higher up the page. Counting on-track targets is deliberate: ACCESS metrics run in mixed directions, so a per-metric ring would fill backwards on the greater-than target and could not express a submit at all.
• Patient reports off static prerendering. The /me/reports pages now opt out via `connection()`, matching every other pulse page, so report data cannot land in static HTML.
• Security and the build. Next.js moves to 16.2.12, clearing four high advisories in middleware and proxy handling, transitive floors bring production audit highs back to zero, and the GitLab pipeline is green after two weeks of red.

Recipe chips on porcelain, no leftover mint.ACCESS outcomes open with how many targets are on track.Four categories, four genuinely distinct hues.Warm clay everywhere the accent used to be mint.The sparkline stays in its lane.

Something worth opening.

The Health Book stops being a document that was written once. It refreshes on demand, marks the chapters that changed since you last read them, remembers where you stopped, and prints a one-page summary you can hand to a clinician. Around it: the book is redesigned with ink chapter bands and white cards, recipe pages stop showing three different greens, and Allie's chat moves from green to ink and coral. Breathing, audio sessions and movement now take the whole screen with no nav on top. A set of layout bugs is fixed, from the desktop nav spanning twice the app column to a Start button sitting on its own label.

• The Health Book is useful now. A printable doctor summary collects conditions, medications, latest labs with dates, care items, allergies, and emergency contacts on one clean page: tap Print for your doctor and hand it over. The book is a living document now: a Refresh button recomposes it on demand (authenticated, audited, rate-limited), chapters that changed since your last visit carry an Updated chip, and a Continue reading card takes you back to where you stopped.
• The book is redesigned. Each category (Personal, Health, Plan, Wellbeing) opens with a near-black ink band and a numbered eyebrow, chapters sit on white cards like the Me tab, and empty states follow the house convention. The sweep also killed every decorative mint in the book while preserving the greens that mean something clinically in the illustrations.
• Allie's chat drops the green. The orb keeps its identity, but the chrome around it is now ink bubbles, neutral surfaces and coral accents, with a follow-up rail and recommendation cards. Two color tokens that always resolved to mint are gone.
• Practices take the whole screen. Breathing, audio sessions and active movement sessions hide the bottom nav, the Allie orb and the spacer, so timelines and pose readouts are never covered. The movement pages also lose a layout bug that crammed content into half the width on desktop.
• The Circle composer expands inline in the page where its trigger sits, with no overlay and no scrolling inside itself. Vitals flatten to white rows you can scan. Recipe pages settle on one coherent palette with dedicated macro colors.

The redesigned Health Book index.Vitals as scannable rows.The Me tab.

Ready for ACCESS, shaped for the stores.

The platform release: work that was already true becomes visible. ACCESS is submission-ready on our side, and members now see their outcome track, with trends, right in the app. The iOS and Android shell passes its store-readiness audit (and the audit caught a real bug: voice sessions had no microphone permission on either platform). And programs, quietly shipped weeks ago, finally get their announcement.

• ACCESS, visible and ready. Members see their outcome track (blood pressure, HbA1c, LDL, eGFR) with trends and on-track status right on the health page. Underneath, every CMS submission lands in a persisted ledger for receipts and resubmission, the draft-IG assumptions live in one place for the day CMS finalizes the spec, and WHODAS is mapped for the behavioral-health track. The first live transmission waits only on credentials.
• The mobile shell is store-shaped. A full store-readiness audit of the iOS and Android apps verified icons, splash screens, native background audio, deep-link config, privacy manifests, and in-app account deletion, and caught a real bug on the way: Allie voice sessions had no microphone permission declared on either platform. Fixed, with regression tests. The remaining gaps are paperwork and accounts, not code.
• Programs, announced at last. Multi-week tracks (the 6-week calm anxiety program, the 14-day reset, and more) slot their daily lessons and practices straight into the Today plan, with guided audio, recipes, and movement along the way.

The ACCESS track, tracked with you.Programs join your day.

Color bugs fixed, rough edges filed.

A founder walkthrough turned into a deep cleanup pass. Two token aliases were quietly rendering mint green wherever coral or purple was asked for; both are fixed, and with them a long list of rough edges: low-contrast charts, a half-finished breathing screen, a dead Circle button, first-person poll answers that read in third person, a misfiring crisis disclaimer, and a profile preview that could show a broken image. The colors that remain (porcelain, ink, coral, copper) are the ones that were actually designed.

• The end of accidental mint. Buttons are ink, accents are coral, charts are copper on porcelain. Beyond taste, this was a bug hunt: two token aliases (purple and coral) silently resolved to the mint accent, so green kept appearing where nobody designed it. Both now point at their real colors, and the health score, vitals, and trend charts moved onto proper chart primitives along the way.
• Fit and finish that was overdue. Breathing is a full-bleed immersive stage with settings behind one quiet button. Circle's first-post button works, the composer grows with your writing, and your own poll answer reads in first person. The crisis-numbers disclaimer shows once, where intended. The profile preview can no longer render a broken image. And Allie stops jumping around the home hero: she is one persistent orb, bottom right, everywhere.

Copper gauge, vitals with trends.Breathing, actually immersive.Ink, coral, porcelain: home.

Press play, put the phone down.

A same-day polish and fix wave shaped by founder and designer walkthroughs, on top of a quarter of shipped-but-unannounced platform work. Guided audio comes to the front: narrated meditations and wind-downs play on, screen locked, inside the new mobile shell. The app warms up visually (a porcelain ground replaces the green wash), Today's Plan grows purposeful icons, and staying signed in finally respects that you were, in fact, using the app.

• Guided audio, front and center. Narrated meditation and wind-down sessions ship with real human-quality narration, and the native mobile shell now hands playback to the platform audio engine when the app backgrounds: lock-screen controls, uninterrupted audio, on iOS and Android. The breathing exercise joins the same immersive visual language as the session player.
• You stay signed in while you are actually here. The HIPAA 30-minute idle logoff stays, but genuine activity now keeps the session alive, and if a session does expire, sign-in says so plainly and returns you to the page you were on.
• The app got warmer. A porcelain ground replaces the green-tinted wash across the patient app, Today's Plan sits on a proper card with icons that match each task, Reflect opens a thought box the moment you pick a mood, and a set of mobile layout faults (overlapping vitals tiles, clipped Health Book sentences, unreachable buttons behind the bottom nav, scroll-locked movement pages) is fixed.

A warmer place to land.Breathing, in the session player language.Narrated sessions, lock-screen ready.Vitals that fit a small screen.Mood to thought in one tap.A natural voice on the front door.

Ready to leave the house.

v0.55 lays the mobile foundation. The secure iOS and Android shell that will carry the app into the stores is built: biometric sign-in enrollment, strict origin handling, universal links, offline recovery, and an update screen that stops an outdated shell before it meets web features it cannot handle. Account deletion, the store-compliance work that started last release, is hardened end to end with immediate lockout and atomic cancellation. Privacy moves with it: the language models behind Allie now work inside a boundary that never carries a patient's legal name, request-isolated opaque tokens stand in for identity, and database row-level security now covers every remaining patient-linked record. And behind flags, all off by default, the clinician side begins to take shape: a provenance-preserving clinical signal contract, fail-closed organization scoping for the clinician cockpit, confidence-aware trajectory evaluation for kidney, blood-pressure, and A1c outcomes, and the first renal nutrition intelligence.

• The native shell is built. A secure iOS and Android foundation now wraps the web app: biometric credential enrollment, strict origin handling so the shell only ever loads what it should, universal links that open the right screen, offline recovery, and mobile observability. An installed app checks a cached compatibility setting and shows an accessible App Store or Google Play update screen before an outdated shell can meet incompatible web features. Nothing is in the stores yet; this is the foundation the store submissions will stand on.
• The AI never sees your name. Allie's language models now receive context through a privacy-shaped boundary that excludes source legal names, with request-isolated opaque tokens standing in for patient identity. Alongside it, database row-level security coverage is completed for the remaining patient-linked records: memory, Health Book, feedback, reporting, HIE, synchronization, and authorization.
• Account deletion grew teeth. The lifecycle that shipped in v0.54 is now hardened: a deletion request locks the account out immediately, cancellation and anonymization are atomic, audit history is retained in a privacy-preserving form, and the whole journey is verified end to end against a disposable database on every change.
• The clinician side begins, behind flags. A provenance-preserving clinical signal contract for care gaps and outcomes, fail-closed organization scoping and explainable panel triage for the clinician cockpit, bounded confidence-aware trajectory evaluation for eGFR, UACR, blood pressure, and A1c, minimum-sample cohort evidence with benchmark comparisons, and the first renal nutrition intelligence. Every one of these is off by default and stays off until it passes review.

The pieces come together.

v0.54 is the consolidation release. The engagement systems that grew up scattered across the app, XP, streaks, achievements, celebrations, now run as one server-side loop, and Allie joins it: logging a vital or a meal in conversation extends the same streak the home screen shows. Breathing and movement move onto the same clinician-gated content backbone as recipes and audio sessions, and two new foundations land behind flags: a measurement layer where momentum can only come from real outcomes, and an evidence pipeline that syncs published research through a fail-closed review queue. On the partner side, the CMS ACCESS pipeline now has everything it needs for go-live except the registration credentials: a persisted submission ledger, a scheduled dry-run rehearsal in production, and conformance hardening. The mobile wave starts with in-app account deletion and a native bridge layer, and the whole app sits on a refreshed light-mint visual foundation, deliberately name-agnostic ahead of the rebrand.

• One engagement loop. XP, achievements, and streaks are now computed in a single server-side seam instead of separate client-side approximations, so the streak on the home screen is the same streak everywhere. Mechanics got more forgiving: a short gap earns automatic rest credit, a rest day can be declared with one tap, and the streak can be hidden entirely. Level-ups and achievements that used to be computed and discarded now actually celebrate.
• Allie joins that loop. Allie sees the patient's engagement state (server-computed, never trusted from the client), answers "how am I doing" from real data, and logging a vital or a meal in conversation extends the same streak the home screen shows. A milestone is acknowledged exactly once, and Allie never resists a goodbye: no guilt, no re-engagement hooks.
• ACCESS, ready for credentials. Every submission attempt now writes a receipt into a persisted ledger, the submission cron rehearses in production as a no-network dry run, and conformance hardening landed. The one remaining go-live step is the CMS participant registration; once credentials arrive it is configuration, not engineering.
• The mobile wave starts. In-app and web account deletion with a 14-day grace window clears the app-store compliance blocker, and a shell-agnostic native bridge means the app is ready for its native shell before that shell exists.

One uninterrupted canvas.Pick the pattern that fits.Deletion without the app installed.The consolidation release, live.

The app that reaches back.

v0.53 gives the platform its first way to reach back. If an evening is passing and a streak is about to go quiet, the app can send one short, gain-framed email; a weekly digest sums the week up in plain counts. The notification settings page now tells the truth: every channel shown has a real switch, and each renders a live preview of the exact email it controls, with your name and your current streak, so what you see is what we send. The reminder channels ship dark behind a flag until the forgiving streak work lands, so the first nudge a patient ever receives will point at a streak that forgives a missed day. Underneath, Allie carries far less instruction text on every turn, the synthetic demo corpus can no longer reach a patient's browser, and clinic access grants are now organization-scoped at the database layer.

• The app can reach back. A streak-at-risk email in your own evening window and a weekly digest in plain counts, both off by default, quiet hours on by default, at most one email per day, one-click unsubscribe. Message content is limited to a first name, a count, and a link, enforced by a CI check.
• Honest notification settings with a live preview. Every switch is real and wired to the backend, and each channel renders the actual email template it would send, with your name and current streak. A channel you turn off keeps its preview, dimmed, telling you it will not send.
• A leaner Allie on every turn. The static instruction block dropped from about 20,000 characters to under 8,000, and simple turns register 20 tools instead of 35, cutting per-turn cost without changing behavior.
• Two permanent guardrails. The synthetic demo corpus is now server-only and provably absent from every client chunk, and a clinic coverage grant can no longer reach a patient outside the granting organization.

Every switch is real.The preview is the real email.

What it costs, finally in view.

v0.52 turns attention from what the platform does to what it costs to run. Every AI turn now records its token counts against a per-model price table, and a cost report turns that into per-day, per-tier, per-provider dollar figures, so spend is something we can see rather than guess. On top of the measurement, a set of zero-risk savings lands: per-tier caps on the expensive output side of each turn, read-aloud audio cached by content hash so an identical line is never synthesized twice, and the background memory extractor held back on turns that cannot teach it anything. Around the AI work, the Circle gets a consume-first pass and a couple of quiet reading surfaces, Allie's orb now leads the home header and hands off cleanly as you scroll, and Gemini voice connects through a steadier path.

• AI cost is now measured, not guessed. A per-model price table plus a cost report (a script and an admin-only endpoint) turn each turn's recorded tokens into per-day, per-tier, per-provider dollar estimates, and the cached-token count is recorded on every turn so cache savings show up.
• Zero-risk savings on top of the measurement. Per-tier output caps bound the costly side of each turn, and any truncation is recorded so it never silently degrades an answer; read-aloud audio is cached by content hash; and the background memory extractor no longer runs on a factual lookup or when a clinical write already saved the fact.
• Allie's orb leads the home. Tapping the top-left orb opens her surface with an iOS-style reveal, and on scroll it hands off to the bottom-right companion orb so only one orb ever shows.
• The Circle keeps getting easier to read. A consume-first, card-based layout, one consistent page gutter, and a "Worth a read" rail so even a quiet circle has something to drop into.

Version 0.51.0

v0.51 stands up two backbones. The first is Studio, our own editorial workspace, and with it the app starts moving off the sample content it shipped with onto real content we manage. The articles, topics, news, and daily inspiration a patient sees are now written and published from one place, and every surface reads them live, so an edit lands in someone's hands right away rather than waiting on a release. Alongside it, a preview-as-patient tool shows exactly what any given person would see, by condition, journey stage, and role, so we always know who sees what. The other backbone is reporting: the path that rolls each attributed patient's monthly outcomes into the quality reports the CMS ACCESS model expects is now built end to end and proven against the published specification, ready to submit the day participant credentials arrive. Around the two, the Circle reads like a place you visit rather than a feed, sign-in hands you to Allie, and another pass of security hardening lands.

• Studio, our editorial workspace, and the move off sample content. The articles, topics, news, and daily inspiration the app shows are becoming real content, written and published from one place, and every surface reads them live, so an edit shows up right away instead of waiting on a release.
• AI drafts, people decide. New content can be drafted automatically from vetted health sources, but nothing reaches a patient until a person reviews it: every draft is cross-checked by a safety panel on a separate model, then held in a review queue to approve, edit, or reject. Nothing auto-publishes.
• Preview as a patient. A governance view lets us enter a synthetic profile (condition, journey stage, role, time of day) and see exactly what every surface would show that person, with the reason each item ranked where it did, so who-sees-what is no longer a guess.
• CMS ACCESS quality reporting, end to end. Each attributed patient's monthly outcomes are rolled into the per-measure FHIR reports the model expects and validated against the published specification, ready to send the day credentials are issued.
• The Circle reads like a place you visit. A consume-first, card-based redesign of the community, and a sign-in that morphs Allie's orb straight into your home header.

Version 0.50.0

Our production-readiness release, and the foundation for going live. Underneath everything is safety: every table that holds health data now enforces row-level isolation in the database itself, so a query can only ever reach the patient it is authorized for, not by convention in app code but by Postgres. On top of that, Allie learns to see: attach a meal or lab photo in chat and she reads it, matches a meal to today's plan, or transcribes a lab for your confirmation. Allie also becomes yours to shape, and a long run of correctness, resilience, and safety hardening lands across the platform.

• Your health data is isolated at the database. Row-level security is now forced on every table that holds it, so a read or write can only ever touch the patient you are authorized for, enforced by Postgres, not just by application code.
• Allie learns to see. Attach a meal or lab photo in chat: a meal is matched to today's plan and checked off with its real nutrition (or logged with an estimate), and a lab is transcribed and shown for your confirmation before anything is saved. Rolling out behind a flag as the vision integration is finalized.
• Allie becomes yours. Give her a name and a warm tone and it changes how she actually talks with you, in text and in voice, not just the label. Rolling out behind a flag.
• Steadier all over: reading a lab photo fails over between AI providers, clinical pages load in one database round-trip instead of four, and a deploy can no longer run with authentication off.

Version 0.49.1

A hardening release: your data stays put and Allie answers reliably. Profile details, reflections and progress now stick to your own account across devices instead of drifting, and the health overview fills in with your real vitals, labs and conditions. Allie replies to clinical questions with clean, captioned cards of your actual numbers and trends, and the chat itself is steadier. Underneath: graceful production limits, repaired migrations, and a June 2026 model refresh.

• Your details stay put. Height, weight, age, reflections and progress now persist to your own account and rehydrate across devices, instead of vanishing or showing up under the demo patient.
• Allie answers in cards, reliably. Ask about your vitals, labs or medications and Allie replies with a clean, captioned card of your real numbers and trends, and the chat no longer drops turns.
• Your health overview fills in. The health page and profile now show your real vitals, labs and conditions (and the care track they imply), matching how medications already appeared.
• The home header runs edge-to-edge, with steadier plumbing underneath: graceful rate-limit fallback, repaired migrations, and a refreshed model catalog.

Version 0.49.0

The headline is The Circle: one closed community of neighbors on the same road. You join with a single warm opt-in, choosing your name and exactly what others see, read a weekly note and question from Allie, keep one small private promise drawn from your care plan, and watch a shared bar fill as everyone does their real daily tasks. Allie is always badged as a helper there, never a member, and a published charter says exactly what she does and never does. Around it, Allie gains a third voice, the health overview becomes a set of tappable widgets, the bottom navigation puts you at its center, and the redesigned pages now read correctly in Swedish and Bosnian.

• Your Circle is here: one closed community of neighbors on the same road, with a weekly note from Allie, a small private promise you keep together, and a shared bar that fills as everyone does their real daily tasks.
• Allie is a badged helper in the circle, never a member, and a published charter spells out exactly what she does and never does.
• Allie gains a third voice through Gemini live, with real-time captions and natural interruption.
• The health overview becomes tappable vitals, labs, and medication widgets, and the bottom navigation puts you at its center with Allie always a thumb-reach away.

Allie's safety moved from guidance to real enforcement, a quietly broken photo meal log now actually saves, and a daily Reflect landed on the home screen.

A safety-and-solidity follow-up to v0.48. Allie's safety rules moved from guidance in her prompt to real enforcement: a reply that drifts toward telling you to stop a medication or implying a diagnosis is now stopped before you see it, a separate reviewer logs anything that strays for the care team, and a real emergency can no longer be hidden by an unrelated number. Photo meal logging, a headline of v0.48 that was quietly failing to save, now genuinely works, and a new daily Reflect moment sits on the home screen: one tap on a mood face, a warm word back from Allie, and a forgiving streak that never shames a hard day.

A warmer look, and an app that grows with you. Allie's new teal orb is her whole presence, every screen shares one warm palette, and the home wraps your avatar in a ring that fills as you finish the day.

Our biggest release so far, and the change that matters most is one you grow into rather than spot at once: RxPulse now reads where you are on your path, from a brand-new diagnosis to steady momentum, and quietly reshapes the home, the daily plan, what you read, and how Allie speaks to fit. Two foundations moved with it. The numbers stopped pretending, so your health score, your streak, your medication adherence, and the read on any vital or lab you log are worked out from what you actually did, and a safety engine that was built but never switched on finally reaches your home screen. And logging got far easier: point your camera at a meal or a paper lab report and the app reads it back for you to confirm before anything is saved. Allie moved into a full-screen chat that handles the keyboard properly and talks more like a person, with a code-enforced safety layer underneath, and the writing grew up with honest authorship and a public security page. All of it lives in a warmer, lower-contrast look: a soft off-white background behind crisp white cards with one deep teal-green accent, and Allie's new glowing teal orb as her whole presence.

• The app grows with you. RxPulse now reads where you are on your path, from a brand-new diagnosis through to steady momentum, and quietly reshapes itself to fit. A fresh account opens to a "you are in the right place" welcome and a shorter day (one small step is plenty at the start). The home puts the right things first, the articles and topics you see lead with what helps at your stage, and Allie leans into reassurance early and names your progress later. When you reach a new stage, the home marks it once with a warm "look how far you have come" moment. Nothing is asked of you; the app simply uses the signals it already has.
• Your numbers stopped pretending. A lot of what the app used to show was placeholder text. Not anymore. Your health-score breakdown now lists the real reasons behind the number and your single best next step. A vital or lab you log gets a plain, condition-aware read instead of just being repeated back. Your streak and medication adherence are counted from what you actually do, across days and devices. And a safety engine that was fully built but never switched on now powers the home's "needs attention" card, so a rule like fluid retention for heart failure finally reaches you.
• Snap a photo to log a meal or a lab report. Logging a meal used to mean hunting through a food list. Now you can take a photo of your plate and the app reads back the foods, the calories, and the full nutrition picture: protein, carbs and fat, plus fiber, sugar, sodium, and the minerals that matter for heart and kidney health. A new "today's food" card on the home keeps a running tally. You can also photograph a paper bloodwork report and have the values read into a table. Both ask you to check and confirm first: a camera can misread, so nothing touches your record until you have looked it over.
• Allie has a new face and a clearer voice. Her chat opened up into a full-screen conversation that handles the phone keyboard properly, and her presence is now a single glowing teal planet, rebuilt to feel alive while barely touching your battery. Her answers read more like a person and less like a wall of text, her cards match the look of the rest of the app, and if you greet her twice she warms up ("I hear you", then "I am listening") instead of repeating herself. Underneath, a code-enforced safety layer escalates a real emergency or a risky medication change even if the model under-reads it, and the voice screen raises a one-tap hotline on any crisis language.
• A daily check-in that actually listens. The check-in now asks the questions that matter for your conditions (breathing and swelling for heart failure, blood sugar for diabetes, and so on), and a worrying answer no longer disappears: it brings up guidance at the right level, the line to call, and a path to Allie. Finishing your whole day is celebrated, and your effort builds a real streak.
• A warmer, lower-contrast look, top to bottom. The whole app was retuned to one warm palette: a soft off-white background behind crisp white cards, barely-there hairline borders, and a single deep teal-green accent in place of the old patchwork of purples, corals, and blues. Green, amber, and red are kept only for what they mean (good, watch, alert), and a proper dark mode arrives alongside. Every screen now feels like part of the same quiet, considered place.
• Honest writing, and a page you can check. Every education article now has real, warm, plain-language content (16 were still placeholders), invented "Dr." bylines are gone in favor of the RxPulse Care Team, and the home news feed lost its boardroom tone. A new public security page lays out exactly which protections are live today and which are still launch commitments, with nothing overstated.

A home that opens with your day, in the warm new palette: the living greeting, the daily-completion ring, and a welcome for a brand-new account.Allie's full-screen conversation, with the glowing teal planet as her whole presence and contextual follow-ups under each reply.Her in-chat cards now carry the same shapes as the rest of the app.On your profile, the health-score arc now hugs your photo as a tappable ring.Reading that meets you where you are: stage-ordered topics, a warmer news feed, and honest authorship.

Today, in your own numbers. The home opens with a daily insight, nutrition reads as three rings, trends speak in plain language, and your progress finally shows up as points and levels.

Our biggest refresh of the patient app yet. The home screen now opens with a short daily insight written for you, instead of the same fixed banner every day. Nutrition shows your carbs, protein, and fat as three simple rings. Lab and weight trends read in plain language, like 'your best week was six days ago'. You can finally see your progress as points and levels, a new Insights page gathers everything worth your attention in one place, and we laid the foundation for a real, moderated peer community, with Allie ready to introduce you to others walking a similar path. Across the app, things got simpler and easier to read.

• A daily insight that speaks to your day. The home screen now opens with something true about today, like your current streak or how your week is going, instead of the same encouraging banner that ran for weeks.
• Your nutrition at a glance. Carbs, protein, and fat now read as three nested rings, so you can see the shape of your day in a single look.
• Trends you can actually read. Your weight and lab charts are smoother and come with a plain-language takeaway underneath, like "your best week was six days ago".
• See how far you have come. The points and levels you earn for showing up every day are now visible on your profile and achievements, where before they were tracked quietly in the background and never shown.
• One place for what matters. A new Insights page gathers the patterns worth knowing, like a reminder to weigh in or a nudge about your blood pressure, so nothing important gets buried.
• The beginning of community, done carefully. We built the foundation for a safe, moderated place to meet others managing the same conditions. Allie is the bridge in, so no one ever lands in a cold, empty feed; she can introduce you to a peer on a similar path when you feel ready. It is being readied behind the scenes and will open up soon.
• A lower-contrast look, end to end. Shouty all-caps headings, hard borders, and loud chips gave way to plainer, more readable type across settings, nutrition, movement, community, and the home screen. Earth, the ocean-and-sage companion, is now the default Allie.

She answers, and listens back. Every reply can be copied, regenerated, or rated, follow-up prompts appear under the latest turn, and a new ocean-and-sage orb named Earth becomes Allie's face.

A round of polish on Allie's chat surface, paired with a fifth orb skin and a live waveform on the voice surface. Every assistant message now carries copy / regenerate / thumbs feedback, suggested follow-ups appear under the latest turn, and the loading states feel composed rather than mechanical. Underneath: a tool-argument self-heal hook, a reusable cross-fade primitive, and a new Earth skin for the orb.

• ActionBar on every assistant message. Copy the response, regenerate it, or send a thumbs up / thumbs down. Feedback writes a SHA-256 hash of the message body (never the prose) to a new chat_feedback table so the clinical director can review where Allie missed the mark each week.
• Follow-up chips under the latest assistant turn. Up to three short prompts patterned on what was just said: side effects of a mentioned medication, a 'when was my last X' lookup for a named lab, a 'share this with my doctor' cue for symptoms, a 'one small thing today' nudge for habits. Rule-based, conservative, returns nothing when nothing matches.
• Push-to-talk dictation in the chat composer (feature-flagged). Tap the dictation mic, speak, stay silent for a beat, and the transcript routes through the same path a typed message would. Built on ElevenLabs Scribe v2 realtime. Off by default until the partner sees it; flips on per-env.
• Streamdown swaps in for react-markdown on the chat surface. Incremental parsing so deltas stop re-parsing the full tree, and the rehype-harden + rehype-sanitize HIPAA-safe defaults come bundled. Ten vitest fixtures pin the sanitizer in CI so a future upgrade can not quietly weaken the surface.
• Earth orb is the new default Allie. The fifth skin (ocean-blue to warm sage with atmospheric edge falloff, evoking the planet seen from orbit) replaces Aurora as the visible brand default for new users. Existing users on the auto-backfilled Aurora value are promoted via a one-time store migration; anyone who had explicitly picked Eleven / Vortex / Creative stays on their pick. Aurora is still the reduced-motion-safe fallback inside useAllieSkin (any patient with prefers-reduced-motion: reduce sees Aurora regardless of stored choice) and remains a free pick in /me/settings -> Appearance.
• Live waveform strip below the orb on the voice surface. Twenty-eight frequency-bin bars driven by the existing AnalyserNode buffer, painted only while Allie is actively listening, thinking, or speaking. Hides when muted, respects prefers-reduced-motion.

Held by shadow, not lines. The decorative lines come down across the app, cards rest on shadow alone, and the little section dots become themed orbs.

A second pass at the less-rounded, less-em-dash, more-custom direction the home tab started in v0.44. Three threads land together. The remaining inline hairlines across nine marketing routes are gone, replaced by a single shared Eyebrow component that uses the squircle title mark already running in the patient app. The canonical .card surface drops its inset 0 0 0 1px hairline so cards float on shadow alone, Apple Notes / Stocks style; about thirty-six decorative borders swept off marketing card surfaces where shadow was already doing the work. And the pulse-title-mark itself got upgraded from a flat dot to a themed radial-gradient orb that echoes Allie at micro scale, tinted per-surface so marketing keeps coral orbs, health-book lists keep sage orbs, and patient-app section anchors get purple ones.

• Nineteen inline coral hairlines across marketing routes (contact, accessibility, insights hub, for-providers, for-aco-enablers, blog index, per-article pages, blog-interactive cards) are gone. One shared Eyebrow component at components/marketing/shared/eyebrow.tsx is now the source of truth so the pattern cannot drift back into "Eyebrow defined inline on every page" again. Three approved group-hover bars stay (the contact form, the accessibility feature cards, the v2-startpage card footer) because they encode interactive feedback.
• Cards across the patient app and marketing dropped the inset 0 0 0 1px hairline layer on the canonical .card surface and about thirty-six decorative `border border-zinc-N` lines on marketing-card variants. Cards now read like Apple Notes or Stocks tiles, defined by shadow alone on light surfaces. Sub-cards on dark zinc panels keep their white/10 hairlines because contrast is load-bearing there; form inputs, buttons, and pill toggles are untouched.
• The pulse-title-mark dot became a themed orb. A two-layer radial gradient: top-left highlight halo over a body that picks up the surface-local --pulse-title-mark-color and color-mixes it with a darker accent for depth. Same shape DNA as the cards, same orb DNA as Allie, no separate asset per call site. The squircle @supports rule also broadened to cover .vital-card-md, .glass-surface, .glass-pill, .btn-glass, and .btn-square; pill buttons stay pure circular intentionally.

Listening, not keeping. A voice-first onboarding in test, and a written-down posture that the llm input layer never sees identifying patient data.

Two architectural moves land alongside the polish that v0.43 started. A voice-first onboarding in test where allie introduces herself by voice and walks the patient through eight conversational steps, live at /onboarding/voice-preview. A written-down phi-free posture that the llm input layer never sees identifying patient data, with three implementation phases scoped. The health book engine moves from fall-through to compose so the hand-authored narrative keeps its tone and the dynamic data (live medications, condition-pack red flags, trial-citation chips) stays current. The home tab also finishes its scandinavian polish pass: 14px squircle radius across 18 surfaces, and a daily-plan editor that finally keeps completed tasks in place with strikethrough instead of having them vanish on tap.

• Squircle radius system: 18 home and marketing surfaces now share the same 14px shape, and browsers that ship corner-shape: squircle render true superellipses for the apple-2026 look. Section titles lose their flanking em-dash hairlines in favor of a single tiny squircle dot, the same shape language as the cards.
• Daily plan editor stopped feeling like things vanished when tapped. The period chip is a picker now, the done toggle keeps completed rows in place with strikethrough, and the duplicate inner today label is gone. Completion clears on natural day rollover.
• Voice-first onboarding sandbox at /onboarding/voice-preview. Allie introduces herself by voice, asks what to call you with explicit nickname framing (so the name you give is non-phi by your own choice), and walks through focus, conditions, and food preferences conversationally. The algorithmic food preference engine scales from the 44 current recipes to thousands as the catalog grows. Stage 4 (promotion to /onboarding behind a ?voice=1 flag) is the follow-on.
• Robert mitchell's health book is now a real, deeply personalized book grounded in his actual lab trajectory: 867 lines of hand-authored content across 8 body sections, 12 medications, 10 contraindications, 80+ wellbeing blocks. Plus a reusable patientEducation field on condition packs (HF, CKD, T2DM, obesity) so any patient with the same condition gets a canonical explainer with trial citations.

The ceremony opens. Audio and mic consents run pre-flight before the eight conversational steps begin.Allie's opening line lands the same way for every patient, with the nickname framing set up explicitly.Five focus options. Each card carries a short descriptor for the kind of help allie can lean into.The taste vector card stack picks the next card to learn the most from the fewest yes-or-no votes.Allie reads back what she heard before the journey closes.Done toggle keeps the completed row in place with strikethrough, instead of vanishing on tap.The engine layers live medications and trial-citation chips on top of the hand-authored narrative.

v0.43.7

First chunk of the CMS ACCESS API client scaffolding, the critical path to the July 5 Model launch via Testing Cycle 2 (May 26-29) or Cycle 3 (June 9-17). Two new modules: an OAuth 2.0 client-credentials TokenManager with proactive refresh + single-flight cache, and a submit-and-poll AccessApiClient that handles the async Content-Location pattern documented in the v0.9.8 Operations Manual. Token expiration buffer is 2 minutes; polling backs off exponentially from 5 to 30 seconds with a 5-minute timeout; HTTP 401 triggers one token-refresh retry; OperationOutcome bodies are parsed into typed AccessApiError. 17 mock-based tests, all green.

• TokenManager class with proactive refresh (2-minute buffer before expiry), single-flight cache (concurrent callers share one fetch), in-memory only (no disk persistence; tokens never logged), and an invalidate() method for forced refresh on 401.
• AccessApiClient with the submit-and-poll workflow per the Operations Manual: submit() reads Content-Location from the 202 response and returns a typed AccessSubmissionHandle; pollSubmission() walks the exponential backoff and returns the result Parameters on HTTP 200 or throws AccessApiTimeoutError after 5 minutes.
• Types module captures the result-code value sets (EligibilityResult, AlignmentResult, UnalignmentResult, ReportDataResult) as union literals with string fallback, so new codes encountered in testing cycles do not crash the client.

v0.43, home v2. The day held in one quiet container, the way the home tab now holds today.

Home v2. The patient home tab is built around four timescale containers now: today, this week so far, your care team, for you. Each section reads as its own moment in one anchored scroll; the bottom nav steps out of the way when you scroll and reappears when you stop. Inside the today section, the daily plan card sits where the day actually lives; swiping a row right to mark it done takes a deliberate gesture now instead of a hair-trigger. WeeklyProgress shows the streak number as the headline of the tile so the rhythm of care is visible at a glance, and the care-team section lifts to a lighter editorial weight. First full accessibility audit on this surface returned 95, and two contrast fixes landed in the same release.

• The home reads as one quiet scroll. Four timescale sections (today, this week so far, your care team, for you) each get their own moment; the bottom nav slides out of the way when you scroll down so the immersion holds, then reappears when you stop.
• Daily plan card lives inside the today section now. Marking a task done by swiping right takes a deliberate gesture, around half the row width, instead of a hair-trigger swipe that disappeared the task on the slightest touch.
• WeeklyProgress shows the streak number as the headline of the tile. A one-day streak reads "Tomorrow keeps it going."; a fourteen-day streak reads "That is the rhythm of real care." The care-team section lifts to an italic serif headline so it reads as considered, not labeled.
• First full accessibility audit on the patient home returned 95. Two contrast fixes landed in the same release (the allie microlabel was hard to read on white; the section eyebrow accents needed a little more depth). One build fix that had been blocking deploys for a few releases also landed.

Home, the immersive opening: editorial hero, then today inside one quiet card.Today, one anchored scroll. Daily plan plus Today's Nutrition all live inside the TODAY section.Daily plan, morning afternoon evening all day. Apple Mail style tab row with a quiet dot and a fused underline.

v0.43.1

Hotfix for an Allie routing bug: asking 'show me my meal plan' (or 'what am I eating this week', or 'what's for dinner') was getting classified as a nutrition-focus haiku-tier question instead of routing directly to the show_meal_plan tool. The tool itself was fully wired (registry + builder + meta + system prompt all correct); the gap was in the tier-0 query classifier's TOOL_PATTERNS regex, which enumerated medications / labs / vitals / care plan but skipped meal plan, weekly menu, today's plan, and exercise plan. Patient got a haiku reasoning loop instead of the rendered 7-day meal grid. Fix extends both regex alternations in `packages/ai-context/src/query-classifier.ts` to cover the missing show_* tool surfaces.

• Allie now routes 'show me my meal plan' / 'what am I eating this week' / 'what's for dinner' / 'monday breakfast' directly to the show_meal_plan tool instead of running through the nutrition haiku reasoning. Same fix covers show_workouts ('show me my workouts', 'what is my exercise plan'), show_today ('what's my day plan', 'show me my today's plan'), and the day-of-week meal lookups ('what's for lunch wednesday').
• No behavioral change for the canonical lookups (medications, labs, vitals, care plan) that were already correctly routed. The regex stays anchored on the same `show / list / give / pull up me my X` and `what is my X` phrasings; just the X surface expanded.
• 14 new test cases in `packages/ai-context/src/__tests__/query-classifier.test.ts` cover the new phrases. Total query-classifier suite is 177 tests, all green.

v0.43.0

Variant polish pass for the orb. Each of the three non-Aurora skins (Eleven, Vortex, Creative) was harbouring a Gemini-era one-off that the v0.42 hooks foundation finally lets us retire: hardcoded hex colors, an unmounted spectrum DataTexture leaking GPU memory, a fixed-speed CSS morph decoupled from amplitude. Aurora picks up the same status debounce so all four variants now cross-fade through state changes uniformly, and a new amplitude-tier attribute lets pulse.css pause the sweep rotation when the orb is fully idle, saving GPU on every default-skin patient.

• Vortex no longer leaks GPU memory. The spectrum DataTexture is disposed on unmount via a paired useEffect (was created in `useMemo([], [])` and never cleaned up). Sphere geometry resolution now picked by `detectGpuTier` from v0.42: 32x32 segments on low-end devices (~75% vertex throughput saved), 48 on mid, 64 on high. `uOpacity` initial value bumped from 0 to 0.25 so the heavily-displaced mesh isn't fully invisible during the fade-in (killed the "sudden mesh pop" the audit caught).
• Creative finally reacts to voice amplitude beyond a scale transform. The morph keyframe speed is now driven by a `--orb-morph-period` CSS custom property the variant sets from amplitude: zero amp keeps the calm 6s period, full amp tightens to 3s. The iridescence layer's `inset: -50%` (which forced a 4x oversized conic-gradient paint every frame) drops to `-20%` with a slightly larger blur for similar visual reach at ~1/3 the paint cost. Hardcoded `oklch(0.8 0.12 X)` stops route through `var(--allie-*)` tokens so the iridescence tracks any future palette shift.
• All four variants now cross-fade through status changes via the v0.42 `useStatusTransition` hook. Aurora gains a `data-orb-amplitude` attribute (zero / low / mid / high) and `pulse.css` pauses the conic sweep-rotate when amplitude is zero, saving the ~1% GPU paint cost on every idle patient.

v0.42, polish and groundwork. A simpler home tab and some shared plumbing under the orb.

A cohesion pass on the home tab and a foundation pass on the orb. The home now reads as one quiet surface: the edit pen moves off the cramped Today plan tabs row up to the section eyebrow ("Edit" sits inline-right of TODAY), three tiles drop redundant headers that were repeating their section eyebrow ("Your day", "This week"), the weekly progress tile no longer says "Today is a fresh start" while sitting inside the This week so far section, and a sweep of arbitrary inline values (text-[10px], rounded-[22px], fontSize: 9.5px, hardcoded #5B51B8 gradients) gets routed through the design system. Under the orb, three new shared hooks (useAlliePalette, useStatusTransition, detectGpuTier) replace inline duplicates that the v0.41.2 hotfix had to scaffold; Eleven wires up immediately, Vortex / Creative / Aurora pick them up in v0.43 alongside the Phase B arrival choreography.

• The "Edit" affordance moves from a 44px circle Pencil button on the right of the Today plan tabs row to a small inline link next to the "Today" section eyebrow. Tab row gets its spacing back; the eyebrow carries the link where it reads as a section-scoped action.
• Three tiles drop redundant headers that were duplicating their section eyebrows: TodayBanner loses its "Your day" header (the section eyebrow already says TODAY); NutritionThisWeek loses its "This week" header (the section eyebrow already says THIS WEEK SO FAR); WeeklyProgress's zero-state copy "Today is a fresh start" becomes "Your week starts whenever you do" so the tile stops naming "today" while sitting inside the weekly section.
• Three new shared hooks land at lib/allie/ as the foundation for the orb quality pass: useAlliePalette reads the brand tokens off document.documentElement so all four variants can stop hardcoding hex approximations; useStatusTransition is the canonical 300ms debounce for voice-state changes (Eleven was doing it inline); detectGpuTier classifies the device into low/mid/high based on WebGL maxTextureSize so Vortex can scale its sphere geometry in v0.43.

The TODAY section eyebrow carries the new "Edit" link; the tabs below sit looser inside their pill.The 44px circle Pencil that used to crowd the tabs row is gone; Edit lives next to the eyebrow now.The zero-state copy stops naming "today" inside a weekly tile.The four anchored sections that v0.41 introduced now read evenly without internal headers fighting their eyebrows.For you reads as its own quiet section, not a tail on the previous one.Eleven is the first variant on the new shared hooks; the inline 300ms timer and the hardcoded color array are gone.Aurora is the default; the same shared hooks land underneath it in v0.43 alongside Vortex and Creative.Tap "Edit" at the TODAY eyebrow and you land here, where the order, the period chips, and the patient-added items live.

v0.41.2

Hotfix for the Eleven orb skin, which rendered as a fully white circle for some users instead of the shader-driven aurora-style sweep. Root cause was a one-line shader bug in `components/ui/orb.tsx`: the fragment alpha was set to `uOpacity` directly, so background pixels (outside the painted ovals) rendered at full opacity. In dark mode, the luminance-inversion mapped that empty background to the brightest color ramp stop, painting the entire orb white. Fix multiplies alpha by the pixel's drawn alpha so background stays transparent. Also added a 300ms cross-fade on the Eleven variant's status transitions so listening / thinking / talking blend instead of snap-cutting.

• Eleven skin renders correctly in dark mode again. The shader bug at `components/ui/orb.tsx` fragment shader was modulating final alpha by uOpacity only, so background pixels stayed fully opaque and were painted by the dark-mode luminance inversion into a flat white. `gl_FragColor.a` now multiplies by `color.a` so untouched background pixels stay transparent.
• Initial uOpacity bumped from 0 to 0.5 so the orb fades in from half-visible instead of fully transparent, removing the 250ms blank-orb period on mount.
• Status transitions on Eleven now cross-fade. The underlying ElevenLabs Orb component changes agentState instantly, which made listening to thinking to talking transitions read as snap-cuts. A 300ms debounce in the Eleven variant smooths the shader response.

v0.41.1

A quiet follow-on to v0.41 that re-shapes the home tab into a single anchored scroll. The Day/Week segmented toggle is gone; below the hero, modules now group into Today, This week so far, Your care team, and For you. Patient-added daily-plan items get a soft "you added" chip so the home reads as joint ownership of the day, not just the clinic's orders. The weekly tile is rebuilt as movement framing instead of a three-number scoreboard. Motivated by multi-lens analysis (UX, psychology, business, clinical) for the 65+ chronic-care cohort landing on ACCESS in July.

• Single scrollable home with four anchored sections (Today, This week so far, Your care team, For you), driven by a new `homeSection` field on the module registry. Each section emits a small eyebrow + divider so the scroll narrative is legible without a toggle.
• Source badge on patient-added items in the daily plan. The /me/daily-plan editor (v0.40) lets you add your own goals; those rows now render with a small amber "you added" chip alongside the clinician-pushed ones, so the home honors joint ownership of the day.
• WeeklySummary numeric scoreboard replaced with a movement-framed WeeklyProgress tile. One warm sentence ("3 days in a row. Keep showing up.") plus small amber/sage pills that only render when there is real signal. Avoids the fixed-achievement frame that backfires for chronic-illness patients whose effort doesn't always show in numbers.

v0.41.0

A refreshed visual layer for Allie and a quiet platform pass underneath. Allie now wears one of four skins, the home tab gets an immersive hero, and the codebase is brought up to current versions with turbopack as the production bundler. Aurora is the default; Eleven, Vortex and Creative load on demand so the main bundle stays light. The change is foundation for the broader visual refresh; phase B (choreographed arrival) and phase C (audio UI on non-Allie surfaces) follow in the next two releases.

• Allie can wear one of four skins picked from /me/settings → Appearance. Aurora is the refined CSS-native default that runs everywhere and respects reduced motion. Eleven uses the ElevenLabs UI orb. Vortex is a custom R3F shader with vertex displacement bound to your voice amplitude. Creative is an iridescent morphing blob. Non-default skins load on demand, so the main bundle stays at the size of Aurora alone.
• The home tab opens on a new immersive hero with day/week segmented toggle, a weather pulse, a community live signal, and a weekly summary tile right under the fold. The compat shim under the orb means every surface that already mounted Allie picks up the new visual without any code change on the caller side.
• Platform pass: next.js 16 turbopack is now the default for both dev and production builds; the dead webpack alias was removed; ~40 packages were bumped within semver and three risky majors were taken individually (@assistant-ui/react, @google/genai, dropped the deprecated @types/mapbox-gl).

v0.40.0, the table is set before the first cohort arrives.

Allie starts using the tools, and we begin experimenting with UX alternatives. v0.38 quietly moved your plans and your check-ins into your account; v0.39 was allie picking up the tools (read your record, write to it with a Yes / No gate); v0.40 closes some loose ends the first cohort would otherwise trip on and starts trying out a swipeable tab layout for the today plan with a real editor underneath. Every cohort attribution has a 10-day care-initiation clock running on it, identity-proofing is wired as an adapter (waiting on a vendor pick), and observability is in place with PII redaction first. Allie can also start drafting summaries for your doctor. The today-plan UX is being experimented with; the original accordion stays available in settings while we learn which variant fits best.

• Allie can show you your care plan, vitals, or labs when you ask. Less hunting through tabs; the same chips and rhythm as the home cards.
• A first version of the doctor's summary view, with a send button that hands a clean letter to the right clinic inbox. We are honest that the downstream Direct delivery is still a stub waiting on the partner contract.
• A new swipeable tab layout for the today plan is being experimented with: Morning / Afternoon / Evening as scroll-snap panels with progress dots, drag-left-to-done and drag-right-to-snooze per task. The accordion view from v0.39 stays available as a toggle in /me/settings so you pick the variant that fits.
• Attribution status has a quiet badge on /me/settings/access for the CMS 10-day care-initiation clock, framed as "your care plan can start any time before X" rather than a deadline countdown. The clinic admin queue lists overdue rows as "needs follow-up", not punitively.
• Identity-proofing is wired as an adapter: persona, stripe identity, id.me scaffolded behind one interface with an HMAC-verified webhook handler. Vendor pick is a notion decision doc; activation is one env-var flip in v0.41 once the LOI is signed.
• Observability has a PII redactor before it has a dashboard: SSN, card, email, patient-name, CMS MBI shapes are replaced before any vendor sees a payload. The care-initiation overdue worker emits a captureMessage already, so the alert pipeline is wired the moment a vendor is picked.

Ask Allie "show my care plan" and you get the same surface the home tab shows, no detour to a tab."Show my vitals" lands the same tile group the home page uses, inside the chat."Show my labs" reveals the recent lab list with the normal-range chips the standalone labs page uses.

v0.39.0, allie picks up the tools she has been watching us use.

The release the last one was groundwork for. Allie has read tools to show you today's plan, your meal week, and your workouts in chat; write tools to log a vital, add a medication, or update a care-plan step on your behalf; and a plain Yes / No confirmation gate that puts a deliberate beat between her proposal and your record. Two new home tiles (your day + this week) glance you at how the day and the week are going, a once-per-locale modal confirms the right emergency numbers for where you live, and the production scaffolding for partner clinic onboarding is in place ahead of the Jul 1 cohort.

• Allie can do things now. Add a medication, log a vital, mark a care-plan step done, all on your say-so, with a Yes / No card that lets you confirm before the write lands and a five-minute Undo if you change your mind.
• Allie remembers across conversations. She can carry today's plan into chat, surface your meal week or your workouts without re-asking, and write a small note to her own memory after each turn so the next conversation picks up where the last one stopped.
• Your home tab fills out. A single "your day" tile rolls up today's meds, meals, workouts, and care-plan ticks; a seven-day "this week" strip shows how you tracked through the week.
• A once-per-locale acknowledgment of the emergency numbers where you live. If you ever change your language, the same modal surfaces in settings so the number you would actually dial is always the one you saw last.
• Partner clinic onboarding is ready to receive its first cohort. Self-enrollment with last-4 MBI verifies against the CMS ACCESS roster, a clinical verifier picks up anything that needs review, and your /me/access page shows attribution status and a one-tap revocation.

v0.38.0, persistence groundwork lands first; the Allie tools that ride on it are next.

A quiet release that does loud things later. Your workout schedule and daily check-ins now live with your account instead of just in your browser, the foundation that lets allie remember them across conversations starting next release. Same release also lays the groundwork for integration with partner meal-plan sources, including a potential rxdiet plug-in down the road. We also pulled the Listen-to-article button off the article reader for now, it was misbehaving on iPhones and on some article routes (the personal-letter Listen pill in your health book is unchanged).

• Your workouts and daily check-ins are saved to your account now, not just to your browser. Tomorrow you can clear your cookies, sign in from another device, and find them right where you left them.
• Preparation for integration with partner meal-plan sources. The persistence layer is in place so a future plug-in (potentially with rxdiet) can drop in cleanly when the integration is ready.
• Listen-to-article is paused while we sort out the iPhone glitch. The personal letter in your health book still reads aloud cleanly.
• Coming next, on the way to July 1: allie remembers prior conversations, learns to log your vitals and medications for you, and the partner clinic onboarding flow goes live.

v0.37.9, the English disclaimer stops telling everyone to call 112.

Safety hotfix. The English medical disclaimer used to hard-code "call 112 immediately" regardless of patient locale. A US-based patient in cardiac distress on the English surface was being told to dial a non-routing number. The emergency number now resolves per-locale from the CRISIS_RESOURCES map; US patients on both English and Spanish surfaces correctly get 911.

• Per-locale emergency number injected into the medical disclaimer via DISCLAIMER_TEMPLATES + CRISIS_RESOURCES lookup at runtime.
• US-English patients now correctly see 911 in the disclaimer instead of the previous hardcoded 112. US-Spanish patients also get 911.
• Verified end-to-end by switching patient locale, triggering an emergency refusal, and confirming the response references the correct number.
• Audit follow-on (v0.38.5 PR2) takes the same fix into customInstructions sanitization and prompt-sandboxing for the broader BAA-readiness wave.

v0.37, smaller things across the chapters that add up to a cleaner surface.

Smaller things across the health book that add up to a cleaner surface. Your Chapter 1 personal letter can now be read aloud. Every chapter shows when allie last refreshed it. Empty states across conditions, vitals, wellbeing, and the diary got a plainer editorial redesign. And a tiny but real fix: when you visit /es/ your health book stays in Spanish even if your account language is English.

• Tap-to-listen pill on the Chapter 1 personal letter. The voice strips the formatting before reading so asterisks and other markdown characters are not pronounced out loud.
• Every chapter shows a "last refreshed" stamp in plain language (e.g. allie refreshed this 11 minutes ago). Tap it for the chapter-change summary.
• Empty states across Chapters 4 (conditions), 5 (vitals), 9 (wellbeing), and 10 (diary) got a plainer editorial redesign: hairline border, serif title, warm body, optional secondary action. No centered icon, no oversized button.
• When you visit /es/ the health book stays in Spanish across every chapter, even if your account language is set to English. Small fix; closes a real mixing bug.

v0.36, chapter prose moved from raw HTML to Markdown so AI can compose it.

Chapter prose migration from raw HTML to Markdown across the health book. Same content for the patient, but the new shape unblocks AI authoring of the personal letter (Ch01) and the wellbeing chapter (Ch09) behind feature flags. Four PRs landed in this wave: the deterministic composer started emitting Markdown (PR1), the LLM authoring path for Ch01 went live behind a flag (PR2), the LLM Ch09 mental-health + Ch07 schedule-notes paths went live behind a second flag (PR3), and the legacy ContentBlock html variant was deprecated alongside every dangerouslySetInnerHTML callsite in the chapter renderers (PR4). The content-security policy got tightened in the same wave.

• Deterministic composer emits Markdown for body prose (PR1). ProseMarkdown component renders it via react-markdown@10. Patient-visible output stays identical to the previous HTML rendering.
• HEALTH_BOOK_LLM_LETTER_ENABLED gates LLM authoring of the Ch01 personal letter (PR2). Default off in prod until a human-review workflow exists.
• HEALTH_BOOK_LLM_MENTAL_ENABLED gates LLM authoring of the Ch09 mental-health chapter and the Ch07 schedule-notes block (PR3). Default off in prod; the mental-health flag stays gated indefinitely until a human-in-the-loop review workflow lands.
• ContentBlock html variant deprecated (PR4). All paragraph + infoBox content shapes now carry markdown: string instead of html: string. The legacy hand-authored persona letters migrated from HTML to Markdown in the same PR.
• Every dangerouslySetInnerHTML callsite removed from chapter renderers. The content-security policy was tightened to forbid unsafe-inline scripts entirely.

v0.35 to v0.37, the health book speaks Spanish end-to-end.

Spanish in-app end-to-end. /es/* routes now resolve the deterministic health-book composer with locale=es; chapter copy, intro paragraphs, the index header, and the chrome (back button, top-nav, footer language switcher) all render Spanish without falling back to English. The locale also threads through CRISIS_RESOURCES so emergency-resource numbers route correctly for US-Spanish patients. This wave landed the foundation that v0.37.x polished and that v0.37.9 hot-fixed.

• Deterministic composer threads locale through every chapter assembly path. Spanish entries added to Ch02 body-explanations, Ch03 medications, Ch07 daily-schedule, Ch08 when-to-call, Ch09 wellbeing, Ch10 diary, and the diary configuration generator.
• Index header copy (eyebrow, title prefix, intro paragraph) renders in Spanish with a polished "Libro de salud / Para {firstName}" header. Top-nav title and back-button aria-label localized as part of the same pass.
• CRISIS_RESOURCES gained a Spanish entry that routes Spanish-locale crisis flows to 911 + 988 for US patients. Foundation for the v0.37.9 911 disclaimer hotfix.
• next-intl es message catalog expanded to cover the patient surface, marketing pages, and the health-book chrome. Deep-merge English fallback still applies for any key the Spanish catalog has not yet landed.
• Read-path flip from shadow → persisted (bundled with v0.34.1 PR2). The health-book version table that v0.34.0 created in shadow mode is now the canonical source; the deterministic composer falls back only when no persisted row exists.

v0.34.2, the marketing site quietly grows up.

Marketing site V2 polish pass against the partner punch list. The homepage hero widget cycle gets a humane rework (longer fades, overlapping cards, ease-in-and-out) instead of abrupt pop-on/pop-off, and the trust chips below the CTAs lose the button shape they were borrowing. Coral hairlines that flagged "included" list items become coral check icons across Three movements, Outcomes, and the whats-included program cards, so a feature being included finally reads as a check, not a separator. New scroll-driven motion arrives across the marketing surface: SVG line glyphs draw in on the capabilities tiles, partner-proof badges pop in with an achievement-style bounce, glass cards floating over hero imagery drift on scroll, and the engagement-section feature dividers now vary in width so the row reads as four beats rather than four stamped copies. The journal hides categories with no posts. The insights Let's talk CTA splits onto two rows with equal-width pills. The auth submit button stops looking like an external link.

• Hero widget animation reworked. 36s cycle, ease-in-and-out, 3.5s offset per card with roughly 2s overlap so adjacent widgets fade through each other instead of popping. Single-card visible window is now around 5.8s, up from an abrupt 3.2s. Reduced-motion still shows only the first card.
• Coral hairlines used as list bullets become coral check icons across Three movements, Outcomes the whole system can feel, and the whats-included program cards. A checkmark reads as "feature included"; a horizontal line was ambiguous next to the CTA buttons above it.
• New scroll-driven motion primitives: SVG glyph draw-in on the capabilities monitoring + insights tiles, achievement-style pop-in on the partner-proof metric badges, parallax drift on the four glass cards over hero imagery, and four different divider widths inside the engagement-section feature cards.
• The journal topic strip hides empty categories. Environment and Family disappear today and will reappear automatically the moment a post is tagged with them, no future maintenance required.
• Auth submit button (sign in, sign up, forgot password, reset password, pending) renders as the same solid dark pill the marketing CTAs use instead of an underlined arrow-up-right link.

v0.34.1, two deterministic guardrails on every health-book compose.

Clinical safety polish for the health-book compose engine. Two deterministic guardrail layers now run on every composed chapter set before persistence: a medication whitelist + dose check (Layer A) catches hallucinated drug names or mismatched doses against the patient's active medications; a contraindication fixture (Layer B) flags the top five interaction pairs (anticoagulant + NSAID, GLP-1 + insulin, ACE-I + ARB, SSRI + MAOI, beta-blocker + non-DHP CCB) that show up in chronic-care med stacks when the composed content does not acknowledge them. Rejected composes still persist (audit + replay), the read path falls back to the previous accepted version. Compose status is now persisted | rejected; v0.34.0's shadow-mode is over.

• Layer A guardrail catches hallucinated medication names and dose mismatches. If the composer emits a med not in the patient's active list, or a dose that differs from the snapshot after normalization, the chapter set is rejected before persistence.
• Layer B guardrail covers the top five clinical interaction pairs from the fixture (anticoagulant + NSAID, GLP-1 + insulin, ACE-I + ARB, SSRI + MAOI, beta-blocker + non-DHP calcium-channel blocker). If both legs of a pair are in the patient's active meds and the composed content does not acknowledge the interaction in dailyScheduleNotes or contraindicated[], the chapter set is rejected.
• Compose-job status now persisted | rejected instead of shadow-everything. The audit row carries the layer-by-layer verdict and finding codes; the row stays around either way for forensic replay.
• Wave C v0.34.1 PR1 ships the two deterministic layers; Layer C (LLM-backed scope check) + voice read-aloud + the persisted-canonical read-path flip land in v0.34.1 PR2 once the BAA environment flag is in place.

Version 0.34.0

Engine plumbing for the persistent health book, shadow mode, plus an editorial design refresh of every health-book surface. Every onboarding completion, chat turn, and voice session now enqueues a compose job that writes the patient's composed health-book to a new health_book_versions table. The patient-visible read path stays live for v0.34.0; v0.34.1 will flip canonical to the persisted versions once a week of shadow corpus diff confirms the persisted rows agree with live compose. Alongside the engine, the health book index + every chapter page get a more refined editorial layout aligned with the rxpulse vibe.

• health_book_versions table lands as the audit-trail + future canonical read source. Every compose run writes a row with content hash + snapshot fingerprint + chapter-change summary, tagged status="shadow" in v0.34.0.
• Compose job runs after Next response (after() helper) so user latency stays at zero. Migration to Trigger.dev with durable retry + observability is queued for the next push.
• Debounce + dedup keep cost bounded: per-(member, triggerReason) lock with 30min floor for memory drift, 6h for clinical change, bypass for onboarding. Fingerprint dedup skips compose when inputs unchanged; content_hash dedup skips insert when output unchanged.
• Editorial design refresh of the health book: index gets a typography-led cover, numbered section openers (01 / 02 / 03 / 04), refined chapter cards with tinted icon-wells, and a serif signature sign-off. Every chapter page gains a small-caps section eyebrow (PERSONAL, HEALTH, PLAN, WELLBEING) above its serif title.
• Health-book/__tests__/compose-job.test.ts adds 14 pure-function tests covering fingerprint determinism, content hashing, and chapter-change diff. Vitest goes from 1831 to 1845 passing.

Version 0.33.0

Allie remembers across devices and channels. Three foundations land together: voice now teaches Allie about the patient the same way chat does, mood writes persist to a durable cross-device row instead of localStorage, and a Mastra-compatible working-memory blob seeds at onboarding completion and follows the account. A new /me/about page lets patients see what shapes how Allie speaks with them.

• Voice teaches Allie. Every post-session voice transcript pair feeds the same fact-extraction pipeline chat already uses, so the patient does not have to repeat themselves when they switch between voice and text.
• Mood follows the account. localStorage stays as a transient cache for the active tab; the durable mood row lives in a new patient_mood_events table.
• Working memory seeds at onboarding. A Mastra-compatible markdown blob synthesizes from pronouns + focus + style + conditions + locale, ready for the v0.36+ Mastra cutover.
• New /me/about page shows what Allie remembers: identity, what brought you here, how Allie speaks with you, conditions, and the running list of memory facts. Read-only first; inline-edit chips ship in v0.33.1.

Version 0.32.1

Cleanup wave on top of v0.32.0. Demo patient content (Robert, Jagoda, Sven, the template, the synthetic roster) consolidates under data/members/_demo/ so the real patient surfaces and the curated walkthrough live in clearly separate places. Chapter 8 (when to call) drops the last hardcoded contact name from its Bosnian copy in favor of a generic contact-person placeholder. The on-site package version finally catches up with the shipping artifacts.

• Demo data moves under data/members/_demo/. Real patients see only their own data; the partner walkthrough at /demo keeps working unchanged.
• Chapter 8 contact line no longer carries an individual's name in any locale. Generic 'contact person' phrasing everywhere.
• package.json bumps 0.30.3 to 0.32.1 to reconcile with the v0.31.5 + v0.31.6 + v0.32.0 + v0.32.1 visible artifacts.

Version 0.32.0

Scheduled for May 14, 2026 (not yet announced; code deployed, formal release waits on a final polish pass). Health book stops being a static document. Every signed-in patient gets a personalized chapter set on every render, built from their onboarding answers + clinical snapshot + memory facts Allie has been accumulating across chat and voice. The hand-written prose in data/members/{id}/health-book.ts becomes reference material; the live view always composes from current data.

• Personal letter (Ch01) greets by name and tone, acknowledges what brought the patient here, lists their conditions, and includes a running list of what Allie has learned in recent conversations.
• Four dynamic chapters now source from real patient data: Ch03 medications schedule, Ch07 daily flow (with GLP-1 injection-day callouts when the patient is on semaglutide, anticoagulant bleed-watch when on Eliquis/Warfarin/Xarelto), Ch08 condition-scoped red flags, Ch10 diary tracking columns.
• Cold-start empty states across the clinical surfaces. /health/medications and /health/labs now show a warm "no data yet" card with a one-tap path into Allie chat to add the first medication or share a lab printout.
• /demo route as the explicit partner-walkthrough toggle. Sets the demo cookies and redirects to the curated Robert Mitchell home; the only sanctioned path for a non-admin patient to view demo content.

Health book index now composes the header from the authed user, no more "Patient" placeholder.Ch01 personal letter, composed from current onboarding answers + conditions + memory facts.Post-onboarding deep-link into Allie chat with ?from=onboarding.Cold-start empty state on /health/medications, one-tap path into Allie chat.Cold-start empty state on /health/labs.

Version 0.31.5

Real patients are first-class. The home, Allie, and health surfaces were defaulting to Robert Mitchell's data for any signed-in user whose pulse-mode cookie wasn't set, which turned out to be every authed session. Fixed by inferring me-mode from the auth state itself; the cookie remains as an override knob for admin and family preview. Same release rebuilds onboarding into a lean five-step flow (welcome with name + pronouns, focus picker, conditions, communication style, recap with Open-Allie deep-link) and adds three new user-record fields (pronouns, focus, tone) that follow the account across devices.

• Demo leak fix. getMode() now returns "me" for any authed session, was defaulting to "demo" whenever the cookie was absent. getPatientId() no longer falls back to DEMO_MEMBER_ID for authed users either.
• Lean five-step onboarding. Welcome with preferred name + pronouns chips, focus picker (managing / recovering / staying healthy / supporting someone / curious), conditions (skippable), communication style (warm / clinical / coach), recap that deep-links into Allie chat with ?from=onboarding. 30 to 45 seconds end to end.
• Server-truth reconciliation on the onboarding page. Admin resets, cross-device sign-ins, and stale localStorage now work cleanly. Patient is never stuck in a "local says done, server says not" loop.

Step 1 with preferred name and pronoun chip selected.Step 2: focus picker (managing / recovering / staying healthy / caregiver / curious).Step 3: condition chips, skippable. Blood sugar + weight selected.Step 4: tone preference. Coach selected here.Step 5: recap of all answers + Open Allie CTA that deep-links into chat.Home as the real patient. No more Robert Mitchell content for signed-in users.

Version 0.31.0

Voice + Allie modernization wave. ElevenLabs is now the primary voice provider (WebRTC, handles iOS Safari gesture rules cleanly), Hume drops to secondary. Voice on iPhone works reliably; the AudioContext is pre-unlocked on the mic button click before the SDK's AudioWorklet loads. Typed repair-phrase registry (mic-denied, mic-muted, mic-init-failed, network-drop, audio-blocked, config-missing) with en/sv/bs catalogs so a patient sees a clear next step instead of "something didn't connect." Voice transcripts now persist to the same conversation_messages table as chat, and the voice context route loads recent chat turns so Allie picks up where the patient left off in chat without an explicit recall step.

• ElevenLabs is now the primary voice provider; iOS Safari voice works reliably; LiveMood pill returns under the orb showing prosody-derived mood; pinned transcript replaces the chat-bubble layout for stable scrolling.
• Voice transcripts persist to the same conversation_messages table as chat (metadata.channel="voice"), and both surfaces now load recent cross-channel turns so Allie picks up where the patient left off regardless of which surface the conversation started on.
• AllieSurface visually polished: opaque dark teal surface in both light and dark themes (was transparent in light, content from underneath bled through), z-index raised so it paints over the chat-sheet header and composer, drawer scale effect dropped to fix the iOS "zoomed-in white screen" bug.

Version 0.30.1

Post-v0.30.0 audit fixes. Administrative surfaces inherit the design system instead of fighting it, two new test suites cover RLS and authenticated accessibility, and the GitLab CI pipeline matches GitHub Actions parity.

Version 0.30.0

Partner API surface. Clinic EHRs integrate with rxpulse over HTTP rather than the v0.29 portal: bearer-key auth, six endpoints covering attribution roster + Co-Management visits + appeals + outcome snapshots + dry-run submission export + marketplace lead intake, plus a signed webhook outbox going the other way. /admin/partners gives super-admin a per-org integration dashboard with one-click key issuance. Pack A foundation work on PROM persistence + performance-year ramp + Co-Management billable math + the audit-trigger extension lands underneath the partner surface.

• POST /api/partner/{slug}/attribution attests a beneficiary into an ACCESS track. POST /co-management/visit logs a billable visit (4-month cap + annual cap + onboarding bypass enforced server-side). POST /appeals submits an outcome appeal. GET /snapshots pulls monthly outcome snapshots. GET /export?format=csv|json returns the dry-run submission bundle the clinic billing team rehearses against. POST /lead receives a marketplace partner referral and emails a signed magic link to the patient.
• Per-org bearer keys (rxpulse_pk_{prefix}{secret}, 55 chars, SHA-256 hash + constant-time compare), Idempotency-Key cache with in-flight sentinel for concurrent-retry races, per-endpoint rate limit buckets, partner.* audit namespace. /admin/partners super-admin UI generates + lists keys (raw key shown once in a click-to-dismiss modal), shows per-org integration health: active keys, pending leads, queued webhooks, last key use.
• Webhook outbox with HMAC-SHA256 signing (X-RxPulse-Signature: t=…,v1=…, Stripe-style envelope), exponential backoff (1m → 5m → 15m → 1h → 6h → dead_letter), Vercel Cron drain every minute. Events: attribution.attested, snapshot.created, snapshot.delta, prom.high_acuity, appeal.resolved, access.patient.self-enrolled, lead.received/enrolled/dropped/activated. Orgs without a webhook configured are marked delivered with a "skipped" snippet so /admin/partners stops showing them as pending.
• Pack A foundation: PROM persistence endpoint (POST /api/proms/{instrument}) scores server-side via clinical-core (HIPAA safety property: client never supplies totalScore); PROM-to-target bridge resolver maps patient_prom_responses into the vital-signs shape outcomes.ts already reads, no rewrite. Co-Management billable helpers (4-month spacing + $100/yr annual cap + onboarding bypass). Performance-year threshold ramp via the new access_year_thresholds config table; realizeWithheldAmount accepts an optional thresholds parameter so the snapshot job can resolve per (track, year) once CMS publishes the ramp.
• Audit-trigger extension to seven previously-uncovered PHI tables (patient_prom_responses, patient_outcome_snapshots, outcome_appeals, beneficiary_attributions, patient_care_team_members, patient_access via a patient_id-aware adapter, plus the new patient_co_management_visits). Closes the gap the v0.30 multi-expert compliance audit surfaced. ACCESS audit retention is documented as 10 years; an immutable-archival job lands in v0.31.
• Portal-gate (partner_api_config.portal_enabled, default true) wired into all five /api/clinic/{slug}/* routes. /clinic + /practice surfaces from v0.29 stay live as the demo + sales fallback until a partner moves to API-only; flipping the flag returns 404 from those routes for that org without touching anyone else.

Version 0.29.2

Email invitations + mobile UX papercuts. Provider and patient invitation flows go email-first by default: the clinic admin enters name + email + role, Resend ships an accept link, and the recipient sets their own password at /accept-invitation. The v0.29 admin-types-temporary-password path is preserved as the alternative for the in-clinic helper scenario where the clinic admin is sitting next to the patient with an iPad. The Allie chat input no longer triggers iOS Safari's auto-zoom on focus, the bottom nav's Allie button shadow stops getting clipped at the viewport edge, and the banner-tile copy on the marketing pages reads cleanly across the full width.

• Provider + patient invite forms gain a two-option toggle. "Send invitation email" is the default and runs through Better Auth's createInvitation; the recipient gets a Resend-shipped magic link, lands on /accept-invitation, sets their own password, and joins the clinic. Patient invitations carry the full clinical context (DOB, sex, location, language, primary clinician, care team) through invitation.metadata so the accept handler creates patient + patient_access + careTeam in a single transaction without re-prompting. Specialty + NPI travel through the same metadata for providers.
• Invitation management lives on /clinic/{slug}/providers above the active members list. Each pending row shows invitee name + email + role + expiry; one-click Resend refreshes expiry and re-fires the email (same id so any link the recipient already has stays valid), one-click Cancel kills the link with a 410 on subsequent attempts. Per-IP rate limit on /api/invitation/accept (20/min), single-use tokens via status flip in the same transaction as the member/patient inserts, 48-hour expiry default. HIPAA audit-log rows on create + accept + cancel + resend.
• iOS chat input zoom fixed. The Allie composer (a third-party assistant-ui ComposerPrimitive.Input) was overriding Tailwind's text-base font size below 16px, which makes iOS Safari auto-zoom on focus and pushes the input off-screen. Inline fontSize: 16px on the input wins over the third-party styles, viewport meta pinned at initial-scale=1 + width=device-width so the page never renders at a surprising scale on first navigation.
• Onboarding loop fix. The onboardingCompletedAt user-record write now runs FIRST, before the patients + clinical inserts, outside the catch-all try-catch. Earlier the flag write sat at the end of the same try-catch as the clinical writes; if any of those threw (an RLS edge case after the v0.29 multi-tenant policy rewrite, schema drift, etc.) the flag was silently skipped and the patient saw onboarding on every subsequent sign-in. The flag is now decoupled and persists atomically before any clinical write is attempted.

Version 0.29.1

Onboarding focus-step refresh and a small home surface for joined community events. The goals + cohort pickers were doing busywork (goals are obvious from conditions and the rest of the data; cohort gets derived downstream from conditions plus demographics). Step 2 now keeps the condition cards (positive framing) and adds a single optional freeform context field for what forms cannot predict. Separately, when a patient taps Join on a community event, the event now surfaces in their daily plan if it is today and on a new "you are going to" home module otherwise.

• Onboarding focus step drops the goals and cohort pickers. Condition cards stay with positive labels ("Heart concerns" rather than "Heart failure"). A single optional textarea takes whatever the patient wants to share about what they have tried, family history, or what helps. That gets stored on patient.notes so the health-book adapter can read it.
• "You are going to" home module groups joined community events into Today / This week / Later, sourced from the existing localStorage joined-activities-store with no new schema. The module auto-hides when the joined set is empty so it stays out of the way for patients who have not joined anything.
• Today's joined community events drop into the daily plan card's morning / afternoon / evening clusters as read-only event rows (calendar icon + start time + location + View CTA, no completion checkbox: showing up is not something the app verifies).

Version 0.29.0

Multi-tenant admin. Partner clinics get their own scoped surface at /clinic/{slug} where the clinic admin invites practitioners, creates patient logins, and assigns a primary referring clinician plus a care team. Each provider sees only their panel at /practice/{slug}; cross-doctor coverage flows through time-boxed grants on patient_access. The Postgres RLS predicate now resolves org membership and honors the new expires_at, so the safety property holds at the database layer rather than only at the API.

• /admin gets organization management. Super-admins create orgs, assign owners, and add members with specialty + NPI. The user-create form carries an optional org + member role + practitioner specialty/NPI so a clinic admin can be provisioned in one step.
• /clinic/{slug} is the new clinic-admin surface. Overview, patients, providers, and settings; provider list shows specialty + NPI, the patients tab shows the full org panel for clinic admins and the doctor's own panel for regular members.
• /practice/{slug} is the practitioner panel. Lands a single-org provider straight on their panel; many-org providers get a switcher. Today shows the primary-clinician + care-team filter; cover grants flow through patient_access with an expires_at so weekend coverage auto-expires.
• RLS predicate rewrite. A new org-scoped branch joins through patients, member, and patient_care_team_members so the gate fires the same way for direct DB access and for app reads. SECURITY DEFINER lets the function resolve org membership without recursing through its own policies. Demo robert-mitchell stays publicly visible via the existing open-patient bypass.

Version 0.28.0

Onboarding rebuild and public sign-up closed. New patients land in a three-step welcome → focus → ready flow that finishes server-side; the session cookie refreshes once the loop closes so the first screen they see is the home tab they just configured. The public POST /api/auth/sign-up/email endpoint is closed; admins create accounts from /admin and pre-approve them, ahead of the partner-clinic onboarding work landing in 0.29.

• Three-step onboarding (welcome → focus → ready) replaces the single-screen welcome that shipped earlier. Each step writes server-side, so a reload mid-flow keeps the patient where they left off. The session cookie refreshes when the loop closes so the new home preferences (focus areas, ready state) are live on the first render.
• Public sign-up is closed. POST /api/auth/sign-up/email returns EMAIL_PASSWORD_SIGN_UP_DISABLED for non-admin callers. New patient accounts are created by an admin from /admin, pre-approved at creation, so the new patient lands straight in onboarding on first sign-in instead of bouncing to a pending screen. Sets up the partner-clinic onboarding flow in 0.29.
• Existing approved users get backfilled as already-onboarded so the new flow does not retroactively block them on next sign-in. Run-once at startup; idempotent.
• Shared see-more affordance. A new SeeMoreLink primitive replaces three different patterns across the pulse surface (the loud coral pill on SectionHeader, the chevron+text link inside ribbons, the bare hover-underline on care-plan-today). Steadier at rest, single hover treatment, respects prefers-reduced-motion.

Version 0.27.2

Modular home gets ordering. Each row on /me/settings/home gains a Move Up / Move Down arrow pair, and Allie can do the same conversationally ("move recipes up", "X should be higher"). Pin order still wins, then your chosen order, then the score-based default.

• Each row on /me/settings/home now has a chevron-up / chevron-down pair next to the segment toggle. Tap to shift a card up or down by one position within its section. Disabled at the top or bottom edge so you cannot move past the boundary.
• Allie picks up the same control. Say "move recipes up", "move news to the top", or "put labs above vitals" and she shifts the card. Each call is one position; for a long jump just say "pin it" instead and it goes straight to the top.
• Layout engine sorts in three tiers now: pinned modules first in pin order, then non-pinned modules in your chosen order, then everything else by score descending. Pin order still wins over a manual reorder, so a pin always lands at the top regardless of where the card was in your order.

Version 0.27.1

Modular home: each patient picks which cards show on their home tab. New /me/settings/home page with a Hidden/Visible/Pinned toggle per card, plus Allie can do the same conversationally ("hide the news", "pin my care plan to the top"). Required cards stay locked on so the home never breaks.

• New /me/settings/home page lists every card on the home tab grouped by section (primary, wellness, tracking, discovery). A 3-state segment toggle per row: Hidden, Visible, or Pinned. Pinned cards float to the top, hidden cards disappear from the home until you bring them back, and a reset link at the bottom clears all per-patient overrides.
• Allie gains the same control conversationally. Say "hide the news", "pin my care plan to the top", "show me the recipes again", or "what cards are on my home right now" and she handles it. The chat-side and the settings page write to the same store, so a change made in one shows up in the other on next render.
• Required cards (welcome banner, today checklist, nutrition) stay locked on so the home never breaks. The toggle disables the Hidden state on those rows and shows a small "Always shown" hint. Tier-locked cards (the ones that unlock at week 2 or week 3 of using the app) display an "Unlocks at week N" badge and disable the row until the patient gets there.

Finishing a breathing session, the round-trip ends with a small celebration card.

Active care log: finishing a breathing or movement session now round-trips into today's care plan in the right time-of-day cluster, with a +XP toast and a small celebration card on first-time unlocks. Same surface on /home (compact "just logged" ribbon) and inside Allie's chat widget so the round-trip stays visible wherever the patient looks.

• Finish a breathing or movement session, get a +15 or +20 XP toast, and the entry threads into today's care plan in the right morning, afternoon, or evening cluster. The Done button now reads "See it on your care plan" and routes you straight there.
• A first-time completion shows an inline achievement-unlock card below the CTA: badge icon, localized title, HP gained. Movement gets its own first-session, 7-day-streak, and 30-day-master badges modeled on the breathing trio. Detection runs through the existing engagement-engine filter so HP totals and earned-badge state stay coherent across surfaces.
• New "All today / By time of day" toggle on /care-plan with state persisted via URL param. Allie's show_care_plan widget mirrors the toggle inside the chat bubble. /home picks up a compact "just logged" ribbon above today's checklist that lists up to three sessions logged today with timestamps and a "see care plan" link, returning null on quiet days so the home stays clean.

Home tab. Just-Logged ribbon above today.New All / By time toggle on /care-plan.Time-of-day clusters route logged sessions to the right slot.Movement gets its own first-session badge for +50 HP.

Version 0.26.0

Clinical safety substrate goes in. Allie now routes crisis phrases to per-locale emergency numbers and defers dose-change questions to the care team. Citation contract shipped at the schema layer; chat-route enforcement on Allie's actual output lands next release.

• Crisis routing live: chest pain, suicidal ideation (active and passive), severe shortness of breath, and stroke symptoms now trigger templated, review-gated escalation messages tied to the patient's locale. Localized for the markets we cover, with locale-specific emergency and crisis numbers.
• Dose-change and diagnosis-interpretation questions now defer to the care team rather than letting Allie decide. Templated language, never model-generated, so the words are guaranteed.
• Citation contract shipped at the schema layer. Every clinical claim must cite a source the server can verify against the evidence library or an FDA label. Enforcement on the chat route lands next release; substrate is in.

Version 0.25.0

Version 0.24.0

• Versioning and changelog now run on rails. Every change ships with a small "what changed" note; a bot bundles them into a release MR when they land on dev, and merging that MR cuts the version, tags it, and prepends the entry you are reading right now. The on-site changelog stays the source of truth, just no longer hand-authored each release.
• The daily inspiration card grew from 100 quotes to 130, with new sets focused on nutrition, sleep, mindfulness, hope and small wins, and caregiver voices. Two existing quotes that contained em-dashes were rephrased with semicolon and comma so the card stays consistent with the writing style.
• Allie repeats herself less. The pool of canned greetings, thanks, acknowledgments, and farewells grew by 30 to 50 percent across English, Swedish, and Bosnian, so saying "hi" or "thanks" three times in a row no longer surfaces the same reply twice. Two new conversational patterns were added too: "how are you" and short positives like "cool" and "nice" now route through the cheap Haiku-class model instead of falling through to a more expensive tier.

Version 0.23.0

Generative UI lands across the rest of the chat. Labs, vitals, and care plan now render as live cards inside the conversation the same way medications did in v0.22, so most things you ask Allie to "show" you surface as a tappable widget instead of a navigation. Allie's instant pre-written replies got a wider voice so she no longer sounds like a phrase book. And we ran the longest partner-feedback list in a while as a single sweep through the patient app: Discover redesigned, dark mode actually toggles, Add to Calendar and Share work on desktop, the medication popup scrolls, badges look like pills again, and a dozen more small things. Tests at 1233 of 1233.

• Three more chat widgets shipped, completing the inline pattern medications got in v0.22. Asking "show my labs" returns the recent results with reference ranges and out-of-range pills inside the conversation. "What are my vitals" returns the latest readings with directional cues per metric. "What's on my care plan" lists pending action items with a one-tap mark-done that writes through the same protected endpoint the standalone page uses. Same component, same data path, so chat and page never disagree.
• Allie's instant replies got a voice. The zero-cost tier-0 router that catches "hi", "thanks", "ok", "no", "good morning", "bye" and their Swedish and Bosnian equivalents now pulls from a small per-language library of variations rather than a single hard-coded reply per phrase. Generic "hi" is also time-of-day aware (morning, afternoon, evening get distinctly flavored greetings), and when the patient is signed in their first name is woven into roughly half the variants ("Hi Robert, what is on your mind?"). Same speed (about a tenth of a second), still zero token cost, just less robotic.
• A 19-item partner-feedback list closed in one pass through the patient app. Discover tile restyled to read like the Weekly Meal Plan tile (coral hero with a movement-themed background image instead of the small white card it was). Add Medication popup scrolls on small screens. Status badges actually look like pills again. Add to Calendar downloads on desktop. Share opens a real menu (Copy link, Email, Post on X) instead of silently copying. The Begin Session button on Tai Chi was hidden behind the bottom nav and is visible again. Full list in the Fixed and Improved sections below.
• Dark mode actually toggles. The dark tokens were already defined and the provider was wired correctly, but a hydration ordering issue meant the first click did not always propagate, and several shell components were painting hardcoded white that ignored the theme attribute. Both issues fixed: the toggle now takes effect on the first click and the patient-app shell (page background, card surfaces, header pill) reads from theme tokens, so the switch is visible across home, settings, vitals, medications, achievements, and community.
• Accessibility text-size slider scales the rest of the patient app, not just the article reader. The default 16px stays the default; the four other steps shift the html base font-size so any rem-sized utility (which is most of the UI) follows. The slider also stopped showing "Larger" as a label in the middle position because the partner found that confusing.

Version 0.22.0

Allie is the headline. She talks back faster, she now shows interactive cards inside the chat instead of pushing you to a separate page, and the chat sheet itself was rebuilt on a battle-tested foundation so the typing flow feels natural again. Behind the scenes we put the AI cost on a sustainable footing for scale. Plus the original design polish, a cleaner repo, and the AWS deployment path documented alongside Vercel. Tests green at 1203 of 1203. Deeper engineering notes live in GitLab for anyone who wants the wiring detail.

• Allie is faster and smarter at picking the right brain for the job. Simple greetings like "hi", "thanks", or "ok" get an instant pre-written reply with no AI call at all (a tenth of a second, instead of the five seconds we were burning before). Day-to-day questions go to a balanced, cached model. Heavier work like drafting letters or reasoning across multiple conditions still goes to the most capable model, but only when it is genuinely warranted.
• AI cost on a sustainable footing for scale. Average cost per chat is now around 0.6 cents, down from roughly 1.8 cents, which puts us on track for under $3 per patient per year on AI. That matters because the ACCESS program economics only work if we can keep the per-patient envelope tight at thousands of patients. We are still tuning this as we learn from real usage.
• Generative UI has started landing. When you ask "show me my meds", Allie now renders the live medication list right inside the chat as an interactive card (tap a med to see what it is for, why it was prescribed, or mark it as no longer taking) instead of throwing you to the medications page. We are working on the same pattern for labs, vitals, and care plan in the next drop, so most things you ask Allie to "show" you will surface as a live widget in the conversation.
• Streaming feels instant and natural again. The "three dots show, then disappear, then the message pops in" bug is gone. Replies now stream in word by word from the moment Allie starts thinking. The chat sheet was rebuilt on the standard foundation the rest of the AI ecosystem converges on, which is what unlocked the streaming fix and makes the rest of the chat work easier from here on.
• Cleaner repo and the AWS path documented. Single CONTRIBUTING.md as the dev guide, locale config in one place, dead files pruned, the test suite back to green at 1203 of 1203. The AWS ECS deployment running at dev.rx-pulse.com is documented alongside the Vercel path so the next contributor can pick either.

Version 0.21.0

Cross-device profile sync, a long Notion-feedback list closed in one pass, and a more comfortable rhythm on every patient-app surface. Profile data now follows the account, not the device. The Today page reads cleaner, news and meal cards open on first tap, and the Health Score arc sits where the eye expects it.

• Profile follows the account, not the device. The same logged-in user now sees the same display name, date of birth, height, weight, conditions, and avatar on every device they sign in to. Edits made on a phone show up on a desktop after a refresh, and vice versa.
• A long list of small things on the Today page that had been adding friction (afternoon accordion blue tint, off-axis nutrition cards, empty community avatars, mismatched section spacing) all closed in one pass. Plus a Discover tile so /explore is reachable from Today without needing it in the nav.
• Tap-to-open works again on the home page carousels, and Back from a news/meal/article detail now returns to the page you came from instead of stepping through the canonical listing.

Version 0.20.0

Wider accessibility coverage across the marketing site, a /brand surface that documents itself, and a few pages that had been accidentally hidden from search engines are public again. Every email a visitor sees now reads rx-pulse.com.

• Every marketing page on rx-pulse.com now passes WCAG 2 AA accessibility checks. The two recurring offenders (low-contrast captions and a too-bright coral on white) are fixed everywhere they appeared.
• The /brand pages on the site read live from the code: tokens, components, every page on the site, and a public health page that surfaces inconsistencies as we find them. What is documented is what ships.
• The localized product pages and the legacy landing are public again. They had been accidentally gated behind login, so search engines and direct partner links were hitting redirects.

Version 0.19.0

Accessibility, performance, and a leaner AI stack. A live /accessibility statement, axe-core wired into CI, a roughly 12x reduction in marketing image weight, and OpenAI removed from the chat and embedding paths.

• A live /accessibility statement on rx-pulse.com with the standards we hold ourselves to and the channels for raising blockers, plus axe-core wired into CI on every PR.
• Site-wide WebP migration brought the marketing surface from 438 MB of source imagery down to 36 MB, with every photo served through next/image at the right size and fetch priority.
• OpenAI is no longer in the chat routing chain or in the embedding pipeline. Anthropic leads, Google falls through, and the offline demo covers the rest.

Version 0.18.0

Design system foundation. Every dimension across the app, text sizes, corner radii, heights, widths, is now editable from one CSS file instead of search-and-replace across hundreds. Plus more space between sections on the Discover, Health and News surfaces so the patient app feels editorial rather than stacked.

• Discover, Health, News and article pages got 50% more space between sections, they feel separated and editorial now, not crammed together.
• Every text size, corner radius and dimension across the app now uses Tailwind's native utility names instead of hardcoded pixel values. Changing the typographic scale is now a one-line edit instead of search-and-replace across a thousand files.
• A small migration script ships in the repo so any future drift gets caught, anyone can run it before opening a PR.

Version 0.17.0

Allie grows up. Real conversations that survive a refresh, a voice that finally feels alive, and a chat surface that stops jumping around when you type on your phone.

• Allie remembers. Close the app, come back tomorrow, and your conversation is right where you left it, securely, on our servers, never on the device.
• Voice gets a real choice. Pick from four providers in Settings, Hume for emotional warmth, Gemini for fast answers, ElevenLabs for natural speech, or a built-in offline mode for demos, with automatic fallback if the first one is unavailable.
• The keyboard no longer eats the chat. We rebuilt the chat sheet for mobile so the input box stays visible, the textarea grows with what you write, and pressing Enter sends, with Shift+Enter for a new line.

Version 0.16.0

Password recovery, a redesigned authentication surface, the RxPulse Journal, and invite-based onboarding for the measured first cohort.

• Signing in should feel like opening a quiet journal, not filling out a form.
• Password recovery is now a first-class flow, with a transport layer ready for real transactional email.
• Onboarding is intentional, we meet new families in conversation before we hand them a product.

Version 0.15.0

The patient app home becomes a modular dashboard, the B2B site gets a credentialed content hub at /insights, and the full discovery layer (search plus AI answer engines) is in place.

• The home dashboard assembles itself from a shared module library, so personalizing what a family sees is a data change, not an engineering cycle.
• Every B2B article carries a named clinician byline with verifiable credentials and a real conflict-of-interest line, readers can check the person, not just the company.
• The site is now legible to ChatGPT, Claude, Perplexity, and Google AI Overviews alongside traditional search.

Version 0.14.0

The clinical engine under the patient app, modular condition protocols, regulator-ready outcome reports, and the data plumbing that keeps everything accountable.

• Adding a new condition, kidney disease, heart disease, depression, osteoarthritis, is now a one-file change rather than a refactor of the app.
• Every measurable outcome, a lab value or a symptom score, can be exported as a standard FHIR report that payers and regulators already accept.
• Medicare identifiers, California Medicaid eligibility, and audit trails are handled as core primitives, not as afterthoughts.

Version 0.13.2

Code quality hardening, test suite synchronization, and import refactoring.

Version 0.13.1

Bottom nav polish, Allie chat reliability, smoother mobile keyboard behavior.

Version 0.13.0

Editorial UI overhaul, tinted task CTAs, Discover rebrand, instagrammable Daily Reflection, real community avatars, Health Records, Mode toggle, and a profile redesign.

Version 0.12.0

Allie Intelligence, smart navigation, and voice-first companion experience

Version 0.11.5

Community ecosystem, personalized nutrition, and visual brand refinement

Version 0.11.4

Infrastructure hardening, scroll performance, and layout stability

Version 0.11.0

Production hardening, device integration layer, and health progress visualization

Version 0.10.3

Chat performance, parallel DB calls, and Next.js 16.2

Version 0.10.2

Streaming chat with markdown and Google-first AI routing

Version 0.10.1

Gemini Live API voice, real-time conversational AI with ephemeral tokens

Version 0.10.0

Postgres Row-Level Security, AI safety tools, and production database layer

Version 0.9.5

Interactive meal tracking, multi-tenant security, and recipe UX

Version 0.9.4

44 recipes with AI-generated step images and health insights per dish

Version 0.9.3

Neon PostgreSQL persistence, patient data, AI memory, and context caching

Version 0.9.2

4-phase AI personalization engine, nutrition autopilot, and navigation redesign

Version 0.9.1

Multi-expert UX audit, dynamic patient ID, and navigation overhaul

Version 0.9.0

Condition packs, health insights pipeline, and modular ICD-10 clinical engine

Version 0.8.3

Community system, daily plan with gamification, and AI-generated recipe images

Version 0.8.2

Admin panel, responsive layout, and Better Auth deployment

Version 0.8.1

Better Auth, self-hosted authentication replacing Clerk

Version 0.8.0

Pulse V2, complete patient app rebuild from scratch

Version 0.7.0

AI cost optimization, HIPAA compliance, and 931 tests, board-ready engineering

Version 0.6.1

Health book redesign, multi-patient architecture, and Allie AI rename

Version 0.6.0

Next.js 16 upgrade, React 19 Compiler, and Universal Data Adapter foundation

Version 0.5.3

VITAL design system migration and community features

Version 0.5.2

Allie AI companion, voice system, and task redesign

Version 0.5.1

Unified AI entry point, squircle cards, and Rx-Diet branding

Version 0.5.0

Patient app UX overhaul, 30 items across 5 sprints from multi-expert audit

Version 0.4.11

Version 0.4.10

Version 0.4.9

Version 0.4.8

Version 0.4.7

Version 0.4.6

Version 0.4.5

Version 0.4.4

Version 0.4.3

Version 0.4.2

Version 0.4.1

Version 0.4.0

Dashboard overhaul with CRM interaction patterns

Version 0.3.6

Version 0.3.5

Version 0.3.4

10-chapter health book written for the patient rather than the clinician

Version 0.3.3

Version 0.3.2

Version 0.3.1

Version 0.3.0

Internationalization, 3 locales

Version 0.2.4

Version 0.2.3

Version 0.2.2

Version 0.2.1

Version 0.2.0

Complete 6-page clinical dashboard

Version 0.1.4

Version 0.1.3

Complete patient data migration, 2,600 lines of structured clinical TypeScript

Version 0.1.2

Version 0.1.1

Version 0.1.0

Project initialized