Where are we now?

Widgets have grips, edge and corner resize handles, and keyboard layout controls. The first-drag resizing jump and overlapping widgets on a fresh workspace are fixed. Narrow widgets reflow their controls and wrap long labels instead of colliding or forcing the whole page sideways. Local browser checks cover resize handles, first load, reload, phone, and tablet widths; regression tests also cover scrolling, cancelled gestures, and older saved layouts. Verified by hand on an actual iPad.

Widgets stop being one fixed instance per view. Each gains layout options for what it shows — for Today, side-by-side lanes or a single stacked list — plus per-widget filters (a project or a lane) and the freedom to keep more than one copy of a widget, each with its own filters, size, and colour. The add-widget picker no longer hides a widget that is already on the view, and long task titles now wrap instead of overlapping the neighbouring column. Built, locally verified, and confirmed by hand on a real iPad.

A real dopamine break is part of the product, not an afterthought — and its image streams are edited visually now, one section per animal with live previews, suggested sources, and an advanced JSON panel for anyone who liked the old way.

Brown, pink, and white noise work without an account or subscription.

Messy input can land quickly and become a manageable next step.

Drag a task to a new position or a different lane — with a finger on a tablet as readily as a mouse — or move it with the arrow keys and hear where it landed.

Title, project, due date, lane, and effort all change in one visible panel instead of hiding behind a right-click.

Type a whole brain-dump without re-clicking, move a batch of tasks at once, and take back a mis-tap without hunting for it.

Create a list and choose its fields — text, one-of options, tags, dates, numbers, checkboxes, or a link to a task — then drill into any entry for the details and notes.

Two shared lists and a page of their own, so a thing that broke or a wish either of us has takes seconds to record while we are still using the app.

Recording no longer stops itself the moment you go quiet to think. It keeps listening until you press stop, and a pause is treated as thinking rather than as something going wrong.

While recording, a small meter shows sound arriving and says so in words — including when nothing is arriving at all, which is usually a microphone chosen somewhere else rather than anything wrong with e-f. Stopping now says how many words were caught and what to do with them.

Brave deliberately leaves out the service that turns speech into text, but still looks like it supports it — so recording appeared to work and silently produced nothing. e-f now checks for this and says so before you press record, instead of letting you hunt through microphone settings that were never the problem.

The browser's speech recognition gives a website no say in which microphone it listens to, so for now e-f explains where to change the system default rather than offering a picker that would not work. A real choice needs a different recording engine — the same one the phone and desktop apps will need.

Opening a task for editing now turns it into a full-width block, with the grip, tick box and buttons moving into a header above it, instead of squeezing a whole form into a column one word wide.

Each task gets a button for whatever comes next — hand it back to Zola, open the thing it refers to, start focusing on it, or send it somewhere else — with Zola proposing the one that fits when she can tell, and quietly saying so when the button came from a guess.

Track time from a task, stop it from the timer bar, and review totals per project for a chosen date range with CSV export. A four-hour check offers to keep timing or record when you actually stopped. Running time is excluded from totals, and signed-in sessions use a separate personal ledger with one active timer per account and safe retries. The ledger migration is applied, a real signed-in production start and stop succeeded, and two concurrent signed-in sessions on the deployed site proved the cross-device rule: the second session's start was refused while the first device was timing, its refresh showed the first device's running timer, and either session could stop it and see the stopped interval.

Focus audio now has three sources behind one volume slider: the original browser-made noise, Streams — internet-radio and other audio you add yourself by name and https address (there is deliberately no built-in station list: a SomaFM terms check showed curated lists need the provider's written permission) — and Library — your own music played straight off a self-hosted Subsonic-compatible server like Navidrome, with the password stored encrypted and the audio proxied so seeking works without the password ever reaching the browser. Library tracks report to last.fm under its own play-length rules, with failures that can never interrupt the music; streams carry no track metadata and say so. A security pass has since hardened the plumbing: the server address must be a public internet host (private, loopback, and link-local addresses are refused and re-checked through DNS before every fetch), and the last.fm connect flow binds its callback to a short-lived cookie so a cross-site page can't link the account. The pieces are built and unit-verified: what remains is proving live audio against a real connected Navidrome, and Apple Music stays a documented plan until a real MusicKit developer token exists. A Synamp integration is now planned around a proven playback contract; the embedded-player issue remains open until that connection works end to end.

Create Morning, Client work, Deep focus or Evening from a starting layout, or add a blank dashboard. Ordered tabs now sit directly above the canvas with a small add control and a menu on each tab. Rename, widget creation, layout editing, appearance, schedules, location rules, scene design, duplication and deletion are available there; right-click and keyboard canvas menus provide the relevant actions too. The misplaced bottom add-widget bar and hero jump counters are removed. Each dashboard retains independent appearance, scene, widgets and optional local schedules. Model and synthetic browser checks cover creation, rename, duplication, pointer and keyboard ordering, targeted widget additions and reload. Widget scrolling still passes the earlier local wheel checks. Deployment, signed-in behavior and native momentum/touch checks remain.

Private pages share one workspace provider and a persistent icon menu. The logo goes to the dashboard. Choose and order top-bar pages through Customize menu or drag selected links; More lists the remaining pages without duplicates and opens the same customization dialog. Preferences survive the owned snapshot and reload. Compact headers replace introductory copy, repeated titles, dashboard backlinks and header add-widget buttons across the private sections. A small top-right save indicator distinguishes online, local, pending and failed saves. Contract and synthetic browser checks cover selection, ordering, reload, phone layout and recovery from full local storage. Deployment, signed-in navigation and native device checks remain.

A shared catalog connects current widgets with focused private pages. Calendar now combines month, week, three-day, day and list views; older calendar-list links open the canonical Calendar page. Calendar settings choose the shortlist shown as compact toggles, and every calendar view uses the same enabled choices. Hidden calendars stay out of the views without deleting their cached events. The separate task, project and calendar date horizon is named Agenda and supports multiple project filters; Capture inbox is now To sort. Mail, Time and Findings keep glances based on existing records with loading and incomplete state visible. Fresh model and synthetic browser checks cover filters, date ranges, cached records, aliases, preference reload and compact headers. The existing provider event window, richer dedicated editors, deployment and signed-in touch verification remain.

Built locally: the Control panel stays available across private pages with up to eight copies of Tasks, Agenda, Calendar list, Clients, Lists, Notes and Time. Its header is compact, and a full-height edge supports pointer and keyboard resizing with a saved width. Per-copy settings, heights and ordering remain independent of dashboard copies. Desktop opening floats over the page; pinning creates a column, navigation retains the pin, and closing clears it so reopening starts as an overlay. Phones use a dismissible modal with focus return. Snapshot, desktop and phone browser checks cover sizing, pinning, navigation, close/reopen and compact controls. Deployment, signed-in cross-device behavior, native touch/accessibility checks and richer system controls remain.

Next UI slice after the shared shell and Control panel: open two focused workspaces with a saved divider, clear focus and navigation ownership, retained drafts and selected records, and a warning before closing unsaved work. Both panes must share authoritative records through one provider. Phones need a safe switch between retained panes. This is planned; the reference drawer is the first working foundation.

The scene editor opens through a dashboard menu and previews atmosphere, background images, depth, motion and quiet mode before Save or Cancel. PNG, JPEG and WebP files can now be uploaded; the browser resizes them and saves a bounded raster image with that dashboard. HTTPS image links still work. Upload results stay with the dashboard being edited, and cancelling or superseding an upload prevents it from replacing a later choice. Model checks cover saved image validation and duplication; a browser upload and reload verify the actual image flow. Delayed upload races still need dedicated browser coverage. The exposed Customize scene and Pause switching buttons have been removed from the main dashboard while a future focus control is planned. Synamp playback integration, bringing remaining design controls together, deployment and signed-in touch checks remain.

The app already has shared colors, surfaces, type, density, and time-of-day shifts.

A matter-of-fact progress view makes the invisible backend journey followable.

Paper is the first real light theme, Moss and Plum extend the dark family, and a picked accent recolors buttons and highlights while staying readable in every theme. Settings, types, and the stylesheet share one catalog, checked by tests.

The dashboard now has three scene depth choices and theme-colored readable surfaces. A coherent dimensional system across the rest of the product and signed-in visual verification remain.

A quiet setting removes decorative motion and background detail while keeping the same capabilities. The local dashboard scene editor now offers this quiet setting, restores the chosen atmosphere when it is turned off, and suppresses ambient motion through the device’s reduced-motion preference. Existing larger-text controls continue to apply. Product-wide coverage, deployment, and signed-in accessibility checks remain.

Planets, aliens, a toaster, the TARDIS, Cheez-Its, and friends drift by slowly and sparingly.

A first client workspace is built locally with contacts, stable personal-project links, dated progress/waiting/decision/milestone records, last contact, next follow-up, and explicit cadence. Clients can be archived and restored. Client follow-ups and unresolved milestones join the existing task/project/calendar horizon with their source shown. Project links survive renaming and shared household work stays outside client updates. Model, isolated storage, and synthetic browser checks pass. These records still use the personal snapshot: authoritative client/project contracts, per-client horizon lenses, deployment, and signed-in sync/touch checks remain.

Built locally: additional contacts and roles, client goals/audience, approved tailoring context, tone/detail/channel, response windows and independent meeting/update rhythms with project overrides. Private notes stay out of writing prompts. Model, actual-route storage and synthetic browser checks cover save/reload, inheritance and retention through ordinary contact edits. These settings do not book meetings. Deployed signed-in verification and real client checks remain.

A local writing flow saves the original thought, then asks Zola for editable client wording using the chosen client/project and checked facts. A selected client email now has its own owned, saved review with quoted request suggestions, the chosen contact’s recipient/greeting, an editable reply and conflict-aware saves. The original thought and source remain available, duplicate preparation retries reuse the saved result, and provider failure preserves manual wording. Local SQL/provider and synthetic browser checks cover preparation, editing, reload and competing edits. Draft meaning and unsupported claims still need review. Applying the review migration, deployment, signed-in full-process/device checks, real voice examples and authoritative CRM sync remain; no email is sent.

Built locally: exact saved CRM contacts are protected from automatic junk actions, and Client mail reads recent Inbox and Quarantine headers and selected bounded text copies without changing read status. An explicit Zola review rechecks the source, requires a matched client/contact and linked project, quotes proposed requests and saves an editable reply with the original thought and checked facts. Shared addresses require a client choice; an address match and a literal quote do not authenticate the sender or prove intent. Owned review records prevent duplicate AI calls and stale edit overwrites. Local engine, SQL/provider and synthetic browser checks cover these boundaries, recovery and narrow layout. Applying the new review migration, deployment, a signed-in full-process pilot, real-account/provider and native touch checks remain. Full threads, original authentication signals, normalized CRM writes and unattended intake remain ahead.

Built and checked locally: quoted client meeting requests offer saved Google calendar preferences and freshly rechecked times, then an editable invitation with explicitly selected client attendees, a confirmed primary organizer and scoped Zoom host. The user confirms client agreement and approves the exact invitation before booking. Fresh buffered checks precede Zoom and calendar writes. Owned durable operations, revisions and worker leases record provider results; lost responses stay unknown until a matching read, and recovery checks never create meetings or send invitations. Only participant join links are saved. Local migration/SQL, route/provider HTTP and synthetic desktop/narrow browser checks cover ownership, consent, changed contacts/times, concurrency, OAuth, refresh and partial recovery. Applying all three new migrations, HTTPS Zoom setup, deployment, a signed-in full-process pilot, controlled real-provider checks and native touch remain. Inbox delivery, rescheduling, cancellation and meeting-note follow-ups are still ahead.

Built and checked locally: pasted notes become an account-owned saved review before generation. Suggested decisions, actions and questions show literal passages; owner/date mentions are quoted, every personal ownership choice starts unchecked, and calendar dates require user input. Saving has no task effects; explicit approval atomically creates chosen source-linked tasks in Next with durable receipts. Exact duplicate notes reuse their review, stale edits are blocked and later task deletion is respected. Local migration/SQL, production routes, provider HTTP, desktop editing/receipt reload and narrow cross-tab conflict checks pass. Apply the notes migration, deploy and verify the signed-in full process and real wording quality. A separate locally checked fact review now promotes approved decisions, questions and waiting items into update sources; source edits and resolution retain receipt identities. Apply the additional facts migration and verify the signed-in flow. Connected Zoom transcripts and native touch remain ahead.

Deliver reminders even when the dashboard is closed, with lead time, quiet hours, time zones, snooze, cancellation, and a receipt that prevents repeated alerts. Saved Digest schedules are preferences today; dependable notification delivery is not built yet.

Manual drafts have editable greeting, opening, sign-off, and bullets/paragraphs. The new local Zola writing flow also uses client tone/detail preferences and project overrides. Wording stays separate from facts and saved drafts keep their edits. Local model and browser checks pass. Learning from user-approved examples, richer phrasing controls and review of real voice samples remain ahead.

A first manual studio prepares editable updates from dated project records, open waiting items and decisions, milestones, and explicit deadlines. Each prepared fact has a visible source snapshot; missing progress or recipient context becomes a review question. Edited drafts save/reopen and download as text, while changed facts flag older drafts. Reviewed meeting decisions, questions and waiting items now have durable owned source records, with explicit approval and literal passages. Open waiting items carry across windows; agreed choices stay dated, source edits flag saved wording, and stale snapshots cannot restore removed facts. Manual updates and new Zola writing/email reviews read these facts locally. Local model, actual-route storage, and synthetic browser checks pass. Apply the additional facts migration and verify the signed-in process. Task-derived completion history, authoritative draft sync, deployed signed-in verification, and review of real client drafts remain. No email is sent and no progress is inferred from time logs or quiet projects.

Prepare weekly updates or due follow-up drafts in your voice, keep one draft per reporting window, and refresh its facts before review. Draft preparation can run automatically after the manual flow is proven. Sending remains a separate confirmed action with the account and recipient visible.

Build estimates from reusable simple, standard and custom feature blocks. Show meetings, preparation, updates and admin effort using explicit project dates and editable client preferences. Recalculation replaces generated overhead, while accepted scope keeps its version. Existing calculation patterns have been inspected; the e-f quote builder and recurrence checks are still planned.

Prepare design directions with prototypes, palette, typography and clear rationale beside the relevant work. Preserve feedback against stable topics and specific versions; any approval is an explicit event for the reviewed version. Begin with private preparation, then prove scoped access before adding an external client portal. Existing review-room patterns have been inspected; this interface is still planned for e-f.

Type or dictate the all-over-the-place version first.

Zola says what she understood and asks before creating anything.

Retries cannot accidentally create the same task twice.

Restarts, pauses, corrections, asides, and trailing thoughts are part of the test set.

Zola waits through real pauses and keeps the thread during long, nonlinear speech.

Visitors see the product; signed-in people go straight to their own workspace.

An e-f account exists independently from whether Calendar is currently connected.

Two people using the same browser cannot inherit one another’s private workspace cache.

Older local work is offered for review instead of silently adopted or deleted.

Keep Auth.js now; revisit Clerk when native clients need it instead of migrating accounts just for fashion.

Tasks are real database records, not only one giant browser snapshot.

Inbox thoughts keep their lifecycle, source, snooze state, and Zola provenance.

Personal and Google-projected events have their own safe, queryable records.

Stale clients get a conflict instead of silently overwriting a newer calendar edit.

Tasks and inbox thoughts now carry the same revision shield calendar events have.

Both versions stay visible and you choose; nothing is thrown away first, and it never interrupts with a popup.

Shared saves now refuse stale versions, combine independent record edits, and show both versions when the same item changed in two places. Pending drafts survive reload and failed requests, and a choice is checked again if another edit arrives before it saves. Isolated Postgres and two-browser fixtures pass, and two real signed-in accounts have proven it live in production.

A signed-in device can resume from a saved cursor and learn which tasks, captures, events, lists, or list rows changed, including deletions, without relying on an always-on connection. The feed is transaction-bound, ordered, and account-scoped. The migration ran in production, and the deployed site proved the reconnect story end to end: a session that slept through several changes woke, paged through the feed from a cursor, and saw exactly the edits that happened after it — metadata only, no private documents in the feed.

Every data route is checked automatically for a missing sign-in check or a query that forgets whose data it is. A production-database proof also confirms that two accounts stay separate, shared data requires membership, and stale writes cannot erase newer work.

Lists and every row in them are real database records rather than passengers in the workspace snapshot: each carries a revision, a stale save is refused with both versions shown, and deleting a list removes its rows in the same statement. Offline edits queue in a replay-safe outbox, schema rewrites and drag-reorders send only the rows that changed, and moving a list into or out of the shared Home space carries its records across the boundary. Every change lands in the same durable catch-up feed as tasks, captures, and events. Proven against an isolated database, and then live: the migration ran in production, and a signed-in session on the deployed site created, conflict-checked, and deleted a list with its rows through the real routes, snapshot and feed included.

Shared items get equally explicit ownership rules before native devices can change them.

Plan a SwiftUI iPhone app, WidgetKit extension, and Apple Watch companion around the same records, account boundaries, offline queue, and action receipts. The screen scopes and implementation backlog are written down. Native sign-in and authoritative project/view contracts still need to be specified before an Xcode app is created.

Catch the thought before the thought disappears.

See Today, upcoming commitments, inbox decisions, and a project glance with comfortable native controls. The phone shares the same work and shows its refresh state; it does not duplicate the whole desktop dashboard.

Configurable Now/Next, upcoming commitment, due follow-up count, and quick-capture entry points. Widgets use Apple's refresh rules and show when information is old; reminders use a separate notification path. Planned alongside the first iPhone slice.

Bring important-date and client-follow-up reminders to the phone with permissions, quiet hours, snooze, completion, and cancellation tied to the same reminder records as the web workspace.

A Live Activity keeps the current focus step and running timer available without opening the app, using the existing one-timer ledger rather than a second clock.

Continue the same safe proposal and confirmation loop away from the web app.

See Now/Next, the next meaningful date, and the running timer in an Apple Watch companion. Native phone/watch checks must prove freshness and the shared timer rule before this is complete.

Dictate a short thought, complete the current step, or snooze a reminder from the wrist, with offline actions retained until the shared service acknowledges them.

Put the next meaningful item on the watch face and in the Smart Stack through WidgetKit, with visible refresh state and safe handoffs into the companion app.

A compact widget for Now/Next, capture, timer, focus audio, and Zola status.

The Mac app can remember a folder you chose without uploading hundreds of gigs of MP3s.

Prepare a useful overview without turning it into a nagging loop.

Observe view choices, focus sessions, completed work, and reminder feedback only when you opt in. Suggest useful schedules or timing changes with evidence you can inspect, accept, dismiss, or delete. Saved rules remain your choice; a learning engine is still planned.

Remember why something matters and show where that memory came from.

Add mail accounts in the app with encrypted credentials, test their connection, review sender verdicts and unsubscribe approvals, and inspect per-account run counts. Local engine checks and isolated database checks now prove confirmed quarantine, restoration, and purge records, failure and partial-batch handling, recovery of older allowed mail, and matching across folder UID changes. Activity checks and pending-recovery messages expose missing results and failed moves. Hosted work shares a deadline, interrupts slow network calls, and prioritizes accounts waiting longest. The standard one-click header is recognized and redirects are refused as completion. Real read-only sweeps have changed no mail. The ordering migration, deployment, controlled live quarantine/restore persistence, and a recorded scheduled invocation remain. Full-mode deletion, actual unsubscribe delivery, and cryptographic unsubscribe authentication remain unverified.

Choose newest-first or oldest-first, save a default, or run a chosen date window from the app or command line. Parked older work resumes later. Regression checks now prove the engine fetches exactly the selected messages, excluding already-examined gaps and mail outside the date window; backfill stops at its exact cap. Real read-only sweeps checked the corrected path. The CLI settings sync reads the saved order. Quarantine recovery also rotates through bounded batches, resetting its cursor if folder UIDs change. The ordering migration still needs applying to the configured database, followed by deployment and a real persisted sweep before this is complete.

Texting a thought to the household number can land it in the right person's inbox like any other capture: the webhook proves Twilio, checks the owned destination, resolves a verified mobile number to that signed-in person's account, answers provider retries exactly once, and files the text with the same protected write every capture gets. The number, production credentials, saved webhook, per-account linking, database table, settings, documentation, and tests are in place. One account has started linking but has not yet sent the exact one-time LINK message, so no production capture has passed end to end. The settings checker now refreshes that pending state without flashing the panel. A verified real text into each account is still needed before this is complete; A2P registration stays separate unless e-f later sends replies.

Every draft remembers its source account; sending always requires a separate confirmation.

The root site already explains the core promise and routes returning people into /app.

Use real screenshots, working features, and purposeful animation when the product is ready to show off.

Tell the design and engineering story—including the hard invisible parts—in a way you can proudly share.