main
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
| e2fe1b1771 |
A white page, and cards that still read on it
The canvas was #F1F0EC — a warm grey that made every screen look slightly dull. It is white. ── You cannot have both ── That canvas was not an accident: it was deepened from #F7F6F4 on 2026-09-23 precisely so a white card would sit *on* something and its 1px edge could go quiet. A white card on a white page has nothing to sit on. Measured rather than judged by eye: after the canvas change, a pixel scan across a card boundary on the Orders golden read #FFFFFF from end to end. The shadow contributed nothing findable. The cards had not become subtle, they had disappeared. So the edge comes back. `DmCard` drew its border only for tones that are not white — the warm canvas was doing that job for the white ones — and now draws it for every tone, with the lift still underneath. Border plus a soft shadow is the iOS and Revolut shape rather than the borderless Airbnb one, and that is the price of a white page. Paid deliberately. ── What moved with it ── `surfaceAlt` lifts to #F4F2EF. With a white page it is the only token that makes a surface read as recessed — filled inputs, tiles, skeletons, the sunken card — so it can no longer sit within a point of the canvas. `DmCardTone.sunken` filled itself with the canvas colour, which on a white page is not sunken but invisible. It takes `surfaceAlt`. |
|||
| 8757b16cf5 |
PIN sign-in, because the code could never arrive
── What was actually broken ── The SMS gateway was switched off, and `POST /auth/otp/request` does not fail when that happens: it still answers `sent: true`, still issues a valid 4-digit code, and writes it to the **server log**. So the phone path walked customers to a code screen for a code that could not arrive, and every digit they eventually typed was wrong. The failure read to them as "I entered it wrong". Phone sign-in is now a PIN, which needs no gateway. ── Email still sends codes, so email is untouched ── Email OTP goes over SMTP and works. Deleting a working way in to tidy up a broken one is a net loss for anyone with an email on their account, so "Use email instead" and the code screen stay exactly as they were. `login_otp_guard_test` moves to that path — the `sent: false` guard still matters there, and that is now the only place it can fire. ── One screen, three entrances ── `POST /auth/login` says which of them a number is before anything is asked, so the app never guesses. Guessing is not cosmetic: offer "create a PIN" to a returning customer and the server answers `pin_already_set` on a screen that cannot succeed; offer "enter your PIN" to somebody who has never set one and every attempt is wrong. The separate sign-up screen is deleted rather than hidden. It asked for a name and then sent an SMS code — a second entrance asking the same questions and posting a letter that never lands. A new number now gives its name and PIN on the same screen. ── A second sign-in path found a latent bug ── `AppState.signIn` only started `refreshOrders`, and the OTP screen called `detectPickupLocation` itself afterwards to make up the difference. That held exactly as long as there was one sign-in screen. PIN sign-in did not know about the extra call, so Home opened with no pickup and no serviceable cities. The work belongs to signing in, not to whichever screen happened to be last, so it moved into `signIn` and the OTP screen's copy is gone. ── What the screen deliberately does not do ── It does not greet by name. `POST /auth/login` returns the account holder's name, which tells anybody who types a number who owns it; the field is read but never displayed, so it disappears quietly when the backend drops it. It does not say whether the number or the PIN was wrong — the server answers identically for both on purpose, and narrowing it here would turn sign-in into a way of testing whether a number has an account. "Forgot your PIN?" renders only when a support contact is configured. There is no reset endpoint, so it can only point at a human — and telling somebody locked out that help exists without saying where is worse than silence. ── The handover note does not reach the Miler ── The app said "we pass this to your Miler as a note". It does not: `remarks` reaches the admin console and stops, because the rider app reads a `notes` field per stop that the backend never sends. A customer could hand their parcel to a neighbour believing the Miler had been told. Both screens now say it is recorded on the booking, and that the Miler still calls the account's number. ── Also ── DmTextField gains `obscure`, and PinScreen carries a back button — without one the only correction for a mistyped digit was killing the app. |
|||
| c3e25feaea |
Five payload bugs, four pages behind dead rows, and one sheet
── The full-address path was reaching the Miler empty ──
`DestinationGroup.toBookingJson` spread its details FLAT across the
destination. The contract nests them under `details{}`, and a destination
carrying keys the server does not recognise is accepted without a word — so
every building number, street, landmark, recipient name, recipient phone and
pin a customer typed was written, answered 201, and thrown away. The Miler
arrived with a district.
Four more on the same call. The destination pin spelled `latitude`/`longitude`
— the same spelling that answered 422 unserviceable for months on the pickup
before it was fixed there and missed here. A PATCH that sent `null` to clear a
field, with a comment saying so, when the server writes only non-nil values, so
a landmark could be added and never removed. Per-destination `instructions`
folded into the visit's one `remarks` line on the belief the contract had no
per-destination note; it has one. And `contactName`/`contactPhone` on the
pickup object, which the create contract has no room for and drops.
The fix ships unverified, deliberately. If `details{}` is also the wrong shape
the fields drop exactly as they do today — it cannot be worse, and holding it
costs every full-address booking in the meantime. docs/BACKEND_CHANGES.md asks
for the confirmation; tool/verify_booking.sh runs it in one command.
── Who the Miler rings ──
One number reaches the rider and it is the account's: `GET /miler/bookings`
returns a single `customerphone`, verified against production and written down
in the rider app's own stop_contact.dart. So "Someone else is handing it over?"
was collecting a number that reached nobody.
Review now shows the number that will actually be dialled, and the handover
person travels in `remarks` with a name, labelled for whoever reads it. Both
screens say plainly that the rider's call button still dials the account —
better than letting somebody hand their parcel to a neighbour believing
otherwise.
── Account's rows led nowhere ──
Two had no `onTap` at all — a chevron pointing at a page that did not exist —
and three answered with a toast. Five rows making a promise, one keeping it.
Notifications, Payment, Help and About are real screens now, written to one
rule: say only what is true of this app today. There is no notification
endpoint, no stored payment instrument and no push SDK wired in, so none of
them pretends to manage any of that. Support shows no contact block at all
rather than a number that rings nowhere — AppConfig carries the fields empty
until somebody fills them in.
── ONE TOUCH is one sheet ──
It was two in sequence with a dismissal between them, and the destination step
made you open a state to see any city — two levels of navigation for something
its own search already flattened. One flat list headed by state, which is also
the answer to "where do you deliver?", and one surface that changes its
question instead of closing so another can open.
Home says the reach in a line, and it needed two fixes to appear at all:
`cachedCities` walked closed states looking for districts that are only fetched
for open ones, and `loadCities` filled two caches while notifying nobody.
── Sending a second parcel ──
`maxDestinations` is 1 in production, so two parcels for two places means
booking twice — and that cost the whole flow twice, re-answering a door the
customer had not moved from. `startBookingFrom` carries the door, carries the
destination only when asked, and never carries the window: a slot fills up, and
a second booking pinned to one that is now full is refused at confirm with
nothing the customer can act on.
Review also says why there is no "add another destination", so a cap reads as a
limit rather than a missing button.
── Bundle ──
pubspec named its images one by one. Declaring `assets/images/` as a folder
shipped a 974 KB launcher-icon master to every customer for a file no code
opens.
|
