── The splash, in three beats ──
Crimson edge to edge with the truck running across it in white; the invert;
then the mark.
The red starts before Flutter does. Four surfaces painted white before any
Dart runs — launch_background at both API levels, windowSplashScreenBackground
on Android 12+ in light and dark, and the iOS launch storyboard — and leaving
any one of them white makes the launch a white flash followed by a red one.
That flash is the only part of a launch a customer consciously notices.
The invert is one gesture rather than a fade. A white truck on a background
turning white is an invisible truck, so the ground lightens as the truck
darkens, off one controller, and the screen turns itself inside out with the
truck still on it. Fading it would have left the screen empty for the moment
before the mark.
The truck's colour is no longer its own: it is painted through srcIn, so the
file's palette is discarded and only alpha survives. A replacement Lottie now
needs no preparation, and tool/lottie_brand.py is off this path.
`splash.json` is a seamless 3.9s loop — frame 0 is frame 60, the truck never
arrives or departs — so there is no completion to hand over on. _truckBeat is
a decision about how long a launch may hold somebody, not a property of the
file. The comment claiming six seconds was wrong on both counts.
── Three things the splash was getting wrong quietly ──
It showed the wrong logo: doormile-icon.png, the previous mark, to a customer
who had just tapped the new one. tool/icons.py now cuts doormile-mark.png from
the same master alpha as the launcher icon, so they cannot drift again.
The fallback loader was invisible on red. _RoadLoader painted in DmColors.brand
on what used to be a white screen; on crimson that is crimson on crimson, and
it drew nothing at all on exactly the devices that had fallen back to it.
The mark appeared and left in the same frame — _minimum was the sum of the
beats exactly, so the clock ran out as the entrance finished. Hence _markHold.
── The truck was not in the middle ──
Not a layout bug: both beats sit in a Center and always did. The artwork is
drawn low and to the right inside its own 500x500 composition, so a centred
widget rendered an off-centre picture — 30pt right, 36pt down.
splash_centring_test.dart renders a frame at phone size and density, finds the
ink and fails if either beat drifts. It is the only form of test that could
have caught this, and the one that will catch it again when the Lottie is
replaced, which is when the correction goes stale. Two things it taught:
one enormous pump does not let the splash's async start-up chain advance, and
capturing at pixelRatio 1 rasterizes the speed lines too faintly to detect,
which truncates the bounding box and moves the measured centre by 12pt.
── The destination sheet stops resizing ──
Tapping ONE TOUCH opened a tall sheet that snapped shorter a few frames later.
DmAsyncList renders four skeleton rows while it loads — 302pt — and the states
that replace them are nearer 200; the sheet was Flexible, so it was as tall as
whichever state its content happened to be in, and the modal is still
animating up while that swap happens.
The list now lives in a box of one height. That also removes a second resize:
the sheet grew from 48% of the screen to 74% when a state was picked. Both
steps now measure 64% and it never changes size again.
And DmAsyncList takes initialItems, fed by AppState.cachedCities: FutureBuilder
reports `waiting` on its first build even for an already-complete future, so a
warm cache still flashed a skeleton over data it already had.
Three skeleton rows here rather than four — sheet_stability_test caught that
302pt overflows the smallest box the clamp can produce.
── Navigation ──
Pages arrived from the right and left to the right. Right is the direction
the system back gesture drags, so a page leaving under a button press read as
a swipe nobody made — and every push in this app is a layer over the one
before it, which arrives from below. The axis is vertical now: the incoming
page rises and covers, the page underneath stays put and dims, and back drops
the top page off the bottom. Full travel rather than the old 14% nudge, which
is what makes the direction nameable. test/page_transition_test.dart holds it.
── Home ──
The pickup pill is back under the greeting. It was cut on the argument that
the form below already carried the pickup; the form is four hundred points
down the screen, and a location you can only see by scrolling to a form is a
location you do not trust. It shows the locality while the form keeps the
door, so the screen answers one question once, and it now has a resolving
state — it used to print the last place you were, confidently, while the fix
was still in flight.
The sphere's caption was falling off the screen. Adding the pill pushed it out
of the scroll viewport, silently, with no overflow warning — the fourth time
Home growing has clipped something under the sphere. DmBookOrb.field is no
longer a constant: the glow gives way to whatever is left, the sphere never
does.
── 74 points of dead canvas, on every root ──
Home, Orders and Account each cleared the floating tab bar by
`MediaQuery.paddingOf(context).bottom + 74`, on the reasoning that the inset is
the home indicator and 74 is the bar above it. Scaffold with extendBody
replaces the body's bottom padding with the bar's whole laid-out height —
measured, a 34pt inset and this bar hand the body 106. They were clearing the
bar, then clearing it again. One helper now, dmTabBarClear().
── One journey, one vocabulary ──
Review said ORIGIN/DESTINATION; Home's form and tracking say PICKUP/DROP.
Three screens a customer walks in order. The tracking card had already written
down the argument; the booking screen hadn't followed.
The rail's first milestone read "Booking created" directly above "Order
created" — down the rail: created, created, in transit. It is "Pickup booked",
which is what JourneyStage.milestoneLabel has called that stage all along on
Orders and on Home's live row.
── The launcher icon ──
tool/icons.py renders all twenty-eight files from one master, because the
failure mode of exporting by hand is nineteen replaced and nine left on the
old mark, at the density nobody checks.
The master is a bare mark on a field, so it cannot be dropped into Android's
adaptive canvas whole: crimson is the background, the white mark is the
foreground, and it is sized by its reach from the canvas centre so a round
mask cannot take the arrow tip. Measured on the built file: 63.7dp against a
66dp safe zone. A real monochrome layer replaces the colour foreground that
was standing in for one — the system tints by alpha, so that themed as a blob.
ic_launcher_round.png exists at last; minSdk is 24 and android:roundIcon had
nothing to resolve to below API 26.
The artwork is flattened (the source vignettes over ten values of red) and its
horizontal offset is placed deliberately rather than inherited — geometric
centring made the mark lean, because the arrow carries the bounding box right
while contributing almost none of the ink.
── Bundle ──
pubspec declared assets/images/ as a folder, which shipped the 974 KB icon
master to every customer for a file no code opens. Named explicitly now.
The effort pass, end to end. Every screen was run through one test — if I
remove this sentence, does the customer make a worse decision? — and the parts
that failed it are gone.
The flow
Home ▸ BOOK ▸ Where is it going? ▸ When shall we collect? ▸ details ▸ booked
BOOK opens a sheet, not a form. The destination is browsed state-then-district
because a flat list of every serviceable district survives twelve and not
sixty, and search cuts across states because somebody who knows they are
sending to Chennai should not have to know which state it is in. Districts
multi-select, but only where the server allows it: BookingLimits advertises
maxDestinations: 1 until the Miler build keys on consignmentid, and a sheet
that ignored that would sell a booking the network cannot complete.
The pickup window is now a step the customer answers rather than a slot chosen
for them. A pickup window is a promise about somebody's afternoon.
What the screens stopped saying
Home lost the orb caption for returning customers and a four-cell live card.
Send lost the city strip, both address fields, the optional disclosure and
three sentences about charging — the route, the packages and the button are
what is left. Tracking lost a radar with a bike in it, a Milers-in-your-zone
count, a "Step 2 of 7" and a sentence describing the screen you were looking
at. The window sheet lost "Fastest pickup", "4 Milers nearby" and "Relaxed
evening handover".
Type
Poppins, which has no variable release — four static cuts, and the sans styles
set fontWeight alone because fontVariations on a static font is ignored in
silence. Every weight dropped a step and the tracking went deeper: Poppins is
built on near-circles and carries more ink than the humanist faces before it.
Objects
One lit sphere on Home, and the primary button now takes its gradient and rim
because a committing action that is not lit like the hero reads as a different
material. The tracking rail's connector is crimson as far as the parcel has
come, so the line is the progress bar. Confirmation is a white tick on green:
crimson is this app's action colour and that screen has nothing left to do.
Bugs found on the way
The OTP screen dropped digits. Four fields passing focus along lose a keystroke
that arrives mid-transition, so "1234" became "124" and the screen answered
"That code did not match" — blaming the customer for its own race. One field
now, four boxes that only draw.
Nothing ever asked for the customer's location: detectPickupLocation was the
OTP screen's job, so a restored session or an auto-login never triggered the
permission prompt and the pickup map had nothing to centre on.
The launcher icon and both splash screens pointed at a house drawn as two
vector paths — a placeholder that shipped.
The splash clock started when the widget was built rather than when it was
visible, so the truck got 0.45s of a 1.8s beat behind Android's own splash.
It waits on waitUntilFirstFrameRasterized now, raced against a timeout so a
binding that never reports one cannot strand the app.
Also: design/screens/ holds all 19 screens under readable names, tool/ has the
scripts that refresh them and rebrand the Lottie, and DESIGN.md is current.
flutter analyze clean. 88 tests, 1 skipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqVJPB9B4QuieZnBAAKgYQ
Book a pickup, track it to delivery — the rebuilt customer app.
- Design language from doormile-screens.html: brand #8F0F06, Manrope +
Geist Mono (variable fonts), bordered cards instead of shadows, crimson
brand headers, sliding tab indicator, mono for anything read digit by digit.
- lib/data (one live API implementation, plus a debug-only offline fake),
lib/state, lib/ui (tokens, widgets, screens).
- 84 tests, plus a design snapshot harness that renders every screen with the
real fonts: flutter test test/design_snapshot_test.dart --run-skipped
--update-goldens
This replaces the previous app (pubspec 'doormile', app id
com.doormile.customer). That tree remains in history at 6c7d656; note its
android/app/google-services.json is not carried over, and the application id
here is in.doormile.customer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
