What’s new with Murph

Murph now asks who the group newsletter is for, what belongs in it, and what to call it rather than inventing a name and creating it on the spot.

The weekly note from Murph now carries what shipped every other week, and in the off weeks it surfaces a feature you have not tried yet.

The hero phone on the homepage now morphs from a one-on-one thread into a group chat mid-demo, and the page explains challenges and the group newsletter for anyone who does not watch it.

Asking Murph to text you a PDF from your vault used to end with Murph saying it sent one that never showed up. Attachments deliver over iMessage now, and a failed send is reported instead of claimed.

When the connection to the model goes silent mid-reply, Murph gives up and reconnects after 90 seconds instead of sitting there for five minutes.

Text Murph while it's mid background task, a scheduled scan or a big import, and it now stops and answers you right away instead of finishing the chore first. The background work retries on its own later.

When your AI allowance runs out, Murph finishes the reply in flight and then tells you once, instead of going quiet until you text in days later and get turned away.

On iMessage the typing indicator used to vanish a second or two before Murph's message arrived. It now holds until the reply actually sends, and replies land a touch sooner too.

Inviting someone to your family plan is now one screen: name first, a single contact field that switches between Messages, email, and Telegram, and far less fine print. The privacy line always matches what you actually entered.

After a wearable links up, the confirmation screen now reliably shows the connected source and a Text Murph button, even when the browser replays the callback or a fresh sign-in hasn't caught up yet.

Murph got a dedicated playbook for writing music prompts, so the songs it generates for reminders, challenges, and celebrations land closer to the genre and mood it's going for.

Scheduled nudges now rotate their angle: a plain cue one day, a question or a tiny fallback version the next, sometimes a callback, a light challenge, or a song. Same frequency, more reasons to actually read them.

A group's Murph now stays alive as long as anyone in the room has an active Murph, not just the person who set it up, and keeps replying when members are added or removed instead of wedging into a silent retry loop.

Trimmed the wake path end to end: a leaner cold start, a reused warm connection to the runtime, a direct fast-path wake for texts, and quick follow-ups now get picked up by the still-warm runtime instead of waiting minutes.

Murph now durably marks a text answered the moment the reply is delivered, so a runtime restart or an overlapping wake can't replay the same answer twice.

When a wearable syncs after a day or more offline, Murph no longer narrates the days-old sleep and workouts in the backfill. Anything that ended more than 24 hours before the sync stays quiet.

Murph no longer re-pushes its contact card to people who already saved it, so the name and photo you chose stop getting overwritten. Settings gains a row to swap Murph's character and re-download the card whenever you want.

Invites, group joins, experiments, and biomarker pages now unfurl with their own preview images in iMessage and social. An invite says you're invited, an experiment shows its actual title.

During a slow turn, a big lab upload or a long lookup, Murph's mid-turn progress updates now actually reach your chat instead of silently disappearing while you wait.

Access checks now understand family sponsorship on every path, so a sponsored member's group-chat message gets answered instead of silently dropped.

After you create a family invite, the dialog stays open, says plainly that Murph did not text the invitee, and gives you the link to share yourself.

If your browser redelivers the connect callback after a wearable links up, you now land back in the app showing the connected state instead of a bare connection-failed page.

Text Murph seconds after signing up and the queued welcome message now steps aside once real conversation has started. Silent signups still get the welcome as before.

Inviting someone to a full plan now adds the $7/mo seat in the same step with the cost shown up front, the seat count waits for billing to reconcile instead of showing a stale number, and empty seats can be removed.

One broken optional endpoint at our device-data partner had been failing entire hourly syncs for Garmin connections. Murph now skips just the broken resource, so sleep, activity, and daily timeseries land again.

Searching Advanced Antioxidant now finds Advanced Antioxidants, creatin monohydrate finds Creatine Monohydrate, and a brand query that matches nothing falls back to normal search instead of coming back empty.

Fixed the compact sleep-cycle import from Junction and stopped pulling unrelated provider records alongside it, so Garmin sleep stages land correctly in the vault.

A scheduled background run no longer delays the reply to a message you just sent. Foreground conversation preempts cron work instead of queueing behind it.

Ask about sleep, recovery, activity, or workouts and Murph now sees which connected source needs reconnect, so it flags the gap and points you to the fix instead of treating the silence as the truth.

Slow saves during onboarding — supplement regimens and lab panels — now run in a fresh background subagent so the chat keeps moving instead of stalling while writes finish.

Rebuilt the public homepage with a new clock-in hero, an asymmetric editorial layout for the asks-Murph-handles section, and a flat trust band rewritten to match what the security FAQ actually claims.

Closed a state bug where the hosted phone-number sign-in could land on a blank step right after the SMS code was accepted.

Restored the read indicator on inbound Linq messages from active members. Closes a regression where receipts had stopped firing on the active-member webhook path.

A consumed scheduled-reminder wake no longer replays when the cron loop reprocesses the same window. Closed a small path that could deliver the same reminder ping twice.

Murph's Sunday health digest now skips itself when there's no real wearable or experiment movement. If your wearable's quietly broken, it sends one short reconnect note instead of a hollow recap.

The unknown-sender classifier now lets through anyone who mentions Murph or asks a real Murph question, with a per-contact cap of 4 attempts so transient outages don't permanently silence one person.

If you keep replying on the wrong Linq chat, Murph tells you to switch to your home line once per chat — not on every inbound until you finally move.

When a vault-file delivery is missing approval state, the outbox now parks the intent durably and resumes when you approve, instead of churning through pre-provider retries.

Retell now receives call metadata in the fields its API expects, and delayed analysis callbacks can't resurrect a phone call Murph already failed before it started.

Finish a browser task that started over text and Murph replies on text. Telegram-origin handoffs return on Telegram. Email-origin handoffs stay in the same email thread instead of opening a fresh compose.

Closed a class of bugs where WHOOP or Junction records could drift into the next calendar day when the provider timestamp crossed UTC midnight — including WHOOP records flowing through Junction. Past records repair on replay.

At most two non-required progress updates per turn now, spaced apart and only when silence would hurt. Required system notices still bypass the budget.

When Murph can't send yet because approval is pending or a config gap exists, the outbox parks the intent before claiming a delivery attempt — no wasted retries, no auto-reply replay.

Approvals are now consumed atomically against the exact approved media in the model's reply, so stale URLs and replayed retries can't deliver a file you didn't approve.

Rewrote the bank of system replies — welcome, invite/signup, daily quota, home-redirect, AI usage notices — for clearer voice and steadier deliverability when many people see the same message.

Linq typing stays alive across an assistant turn without a 2-second keepalive loop. You still see Murph working — just without the high-frequency churn.

When a Linq voice memo's audio hasn't fully landed yet, Murph retries briefly then degrades cleanly. No provider URLs, storage paths, or routing labels leak into the model prompt.

Multiple matching activities in one wearable sync now each queue the configured reminder. The durable listener stays put for next time, instead of being archived after one fire.

Newly created weekly managed automations now seed at vault-deterministic times across a daytime window, instead of every account receiving the same exact send minute.

The 'All set' screen after a browser handoff now offers a single Reply to Murph button instead of stacking every channel CTA at the bottom of the page.

Strangers texting the Murph line now pass a fail-closed classifier before any invite, member record, or reply gets created.

When Murph needs you to look at a paused browser, you get the live, passkey-gated handoff used everywhere else — not a static screenshot.

A simple "yes" or "go ahead" lets Murph place the order. The handoff link is still there if you want to take over.

Before scheduling a Linq reminder for 11pm–5am local time, Murph flags the spam risk, suggests a nearby waking-hour time, and asks.

Hosted auto-compaction now waits until 164k tokens, so long browser/computer-use loops get more room to finish in one piece.

When Murph plans multiple replies in one turn, all of them dispatch instead of only the first reaching you.

When you accept a habit, ramp, or non-experiment plan, it now lands on a canonical goal-and-regimen surface — and Murph reads that surface back at the start of each turn, so baselines, ladders, and targets don't get stranded in chat.

When Murph introduces movements you haven't seen before, it now picks 2-4, attaches catalog images, gives only the immediate safety stop rule, and asks whether to walk you through them — instead of dumping a long numbered plan.

Murph now compacts its working memory partway through long browser sessions, instead of waiting until the end. Multi-step tasks stay quick and cost less, even when they grow.

Closed a class of races where a long-running turn could overwrite its own next-wake while you sent a follow-up, so reminders, device-sync continuations, and retries no longer slip behind.

Photos, audio, and video you send Murph are now purged from the inbox after 14 days. Murph still remembers what they were — only the raw bytes go.

Bytes are hash-verified before deletion, so nothing tampered with disappears silently, and anything Murph is still mid-reply about stays protected until the reply lands.

In-container tools that run during a turn — research scout, route estimates, supplement lookups — authorize through the same container-identity fence the rest of Murph uses, instead of failing with a silent 401.

Closed a delivery gap where Murph's generated and reference images would sometimes fail to attach on Telegram. They now come through cleanly the same way iMessage does.

Murph can now use italic and underline alongside bold and strikethrough on supported chat channels — emphasis renders natively, not as raw markdown.

If you bounced out of checkout, signing back in through /join now drops you right back into the same checkout step instead of restarting the flow.

When Murph auto-replies to a message that came in while you were away, the reply now sees context from your earlier sessions instead of starting from a blank slate.

Dropped the artificial OS-control typing delay, so Murph fills forms and types into web apps without the slow per-keystroke pause.

Closed a forwarding gap that silently dropped the connected-apps capability before it reached the Codex turn, so Gmail, Calendar, and the rest stay usable end-to-end.

Search now drops weak supplement tokens and uses the correct trigram operator, so brand and ingredient lookups land the product you meant.

If you signed up but didn't finish onboarding, the dashboard sends you back to the hosted onboarding flow instead of dropping you into an empty home.

First-run setup questions got shorter, voice-memo answers are honored, and the onboarding follow-up automation reconciles its own state instead of leaving legacy reminders behind.

Sharing the changelog link on iMessage now renders a clean five-row digest in Murph's own palette, instead of a generic link preview.

Authentication, biomarkers, experiments, connections, and settings received a focused visual polish pass.

Backfilled strict gram serving sizes across 1,710,438 foods and 28,415 supplements using structured evidence only — no volume, count, or container fallbacks.

Murph now recovers more reliably when a new message arrives while the hosted assistant is already active.

Murph's Telegram text replies no longer quote your last message back to you — reactions still target the specific inbound message when they fit.

Progress cards and follow-up decisions now read from the query metric projection, so secondary analysis metrics with data no longer fall through to a misleading "no wearable data" status.

Replies arrive in order and stop dropping when Murph is processing several messages at once.

Trial redemption and billing recovery paths were tightened so interrupted signup flows are less likely to leave confusing account state.

The welcome message you get right after signing up now arrives reliably and exactly as written.

The Wednesday insight automation now rejects tautological or unsupported findings — it stays silent rather than shipping a hollow note.

Replaced the generic spinner on /join and the green-bar provisioning notice with Murph's own dot-constellation mark animated as a sonar ripple, honoring prefers-reduced-motion.

Distance estimates were silently failing with "not configured" — the hosted Codex shell now passes MAPBOX_ACCESS_TOKEN through, so workout and travel questions return real routes.

Triggers the moment your usage crosses the limit instead of waiting for your next message, and releases its claim cleanly if a send fails.

Metrics that were captured but invisible to queries — caffeine, water, mindfulness, heart-rate recovery, AFib burden, lowest/highest glucose, basal body temp, height, period and cycle length — now surface through the same generic path as measurements and lab samples.

Queryability is now a property of the canonical data instead of a per-metric registry, so future observation metrics gain query access with zero new wiring.

Hosted Workers AI transcription writes a per-run usage record, so voice-memo cost shows up in the same ledger as Codex turns and image generation.

When you message Murph after a long quiet stretch, the first reply comes through noticeably sooner.

Asking about WHOOP now pulls from your live sync and any historical ZIP exports together, instead of splitting them.

Hosted transcription now passes verified original audio straight through to Workers AI when it is safe, and routes blocked or oversized formats through ffmpeg into 64 kbps mono MP3 — practical support reaches 15+ minute clips.

Long conversations now keep themselves trim in the background, so Murph stays faster and cheaper to use the longer you chat.

Murph's internal tools start faster between turns, so replies that need them feel quicker.

After autofilling a verification code from Messages, iOS no longer leaves blue selection handles floating over the sign-in OTP boxes.

Brief connection hiccups no longer leave your next message hanging for minutes.

Hosted voice-memo transcription moved off the local whisper.cpp path to Workers AI. A 65-second clip now transcribes in seconds instead of roughly nine minutes.

When Murph ships an update in the middle of a conversation, you no longer get the same reply twice.

Scheduled reminder agents now read your full vault, so a nightly protein nudge won't pretend you ate nothing after you already logged lunch.

Recurring reminders (the PT reset, the nightly meditation prompt) now thread back into the iMessage conversation they were scheduled from instead of starting a fresh session.

Inbound iMessage now requires explicit direct-chat attestation before Murph will rebind your home thread, closing a potential leak where a missing is_group flag could route 1:1 replies into a family group.

Junction-backed wearables — WHOOP, Oura, Garmin and friends — now reconcile every hour instead of every six. Morning sleep and recovery show up sooner.

Changing your email is now a single dialog with an inline auto-submitting OTP step, replacing the previous nested modal stack.

Device-sync passes were silently overwriting the runtime's next wake — confirmed misses last week included a 15:00 PT reminder delayed 179 minutes. Wake selection now merges in the armed assistant cron, earliest wins.

Searching by brand now scopes results to that brand's catalog, so Thorne or Pure Encapsulations queries land on the right product.

Multi-step exercise instructions are paced one move at a time instead of dumped into a single reply.

Settings now offers a single typed-DELETE confirm dialog that cancels any Stripe subscription before the local wipe (fail-closed) and best-effort removes the Stripe customer and Privy user after.

Murph syncs your wearable data in lighter daily summaries, so device connections stay quick instead of slowing under heavy data.

Scheduled reminders that miss their occurrence by more than 30 minutes are now skipped instead of pinging hours later.

Murph stays warm between your messages, so a quick follow-up comes through almost instantly instead of starting from scratch.

Murph's internal tools start up faster after a quiet stretch, so the first reply of the day comes through sooner.