Alby96andClaude Opus 5
06ab816ebe
A saved place is a named point the user keeps: it frames the map now, and once the worker publishes station data it is what the nearest-station readings and, later, the rain notifications will hang off. Free points rather than stations, because people think in terms of home and work, not in terms of which weather station happens to represent them. Everything stays on the device. Places live in SharedPreferences under a versioned key, and there is no account to attach them to and no server that would accept them. That is what keeps the Data safety declaration able to say no location is collected, and it must survive the notification work: the device will subscribe to the topic for the cell containing a place, so the link between a person and a place never leaves their phone. Location is coarse only, and that took enforcing. geolocator declares ACCESS_FINE_LOCATION in its own manifest and the merger pulls it in, so the system dialog offered "Precise" despite the app asking for nothing of the sort; the manifest now removes it with tools:node="remove", and the dialog reads "approximate location" with no choice offered. Requests also go through the platform LocationManager rather than the Play Services fused provider, which prompts about Location Accuracy and, when declined, returns no fix at all — an absurd outcome for an app that only ever wanted an approximate one, and one that also tied location to Play Services being present. The prominent disclosure comes before the system dialog, as Play requires, and is repeated in Settings so someone who already answered can still read what the permission is for. Declining leaves the app fully usable. Two more defects found by running it and by a test: - MapLibreMap leaves cameraPosition null unless trackCameraPosition is set, so "save the map centre" silently saved the region default rather than what the user was looking at. - Place ids came straight from the microsecond clock, so two places saved in the same microsecond shared an id and rename, remove and the duplicate-name check all acted on the wrong one. A test caught it on a fast machine. Saving refuses points outside the region rather than accepting them: a place in Rome would look like it worked and then show nothing forever. Also corrects CLAUDE.md, which still said ARPA states no licence, and records the ARPA realtime API there with the property that governs how it may be used — it lags about 4.5 hours, so it is an observation archive and must never sit next to 5-minute radar looking current. Verified: analyze clean, 158 tests passing, and on the emulator the disclosure precedes the system dialog, the dialog asks only for approximate location, a place survives restart and reinstall, and tapping one moves the map onto it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
9.5 KiB
Roadmap
Each milestone ends with tool/verify.ps1 green and one commit. A milestone is not
done until its acceptance criteria hold, and until it has been seen running on the
emulator — twice now that step has caught defects the tests did not.
Legend: ✅ done · 🔨 in progress · ⛔ blocked on something the project owner must supply
Scope: radar on a map, saved places, ground-station observations, official alerts, rain notifications. Forecasts, lightning, a home-screen widget and advertising are all out — see CLAUDE.md.
M0 — Toolchain and repository hygiene ✅
Flutter 3.47.3 stable, Android SDK (platform-tools, platform 36, build-tools 36.0.0), JDK 21, environment variables. Commit the removal of the old Xamarin skeleton.
Done. flutter doctor -v reports "No issues found!". Installed: Flutter 3.47.3 at
C:\src\flutter (telemetry disabled), Android SDK at C:\Android\Sdk with platform 36,
build-tools 36.0.0, platform-tools and all licences accepted, and JAVA_HOME pointing at
the JDK 21 a previous Visual Studio install had already left on the machine.
M1 — Scaffold, Italian l10n, region config, CI ✅
flutter create with applicationId it.nuvolari.app; flutter_localizations + intl
with app_it.arb as template and no hardcoded UI strings; RegionConfig loaded from
assets/regions/piemonte.json with parsing tests; .gitignore, env.example.json,
tool/verify.ps1, .gitea/workflows/ci.yml.
Done. All four stages green; the AAB carries applicationId it.nuvolari.app,
minSdk 24, targetSdk 36. Tests assert against the shipped Piemonte asset and a captured
copy of the live ARPA CAP feed.
The Gitea instance may have no Actions runner. The workflow file is written to be GitHub-Actions compatible, but
tool/verify.ps1is the verification that must pass.
M2 — Map, attribution, Sources screen ✅
MapLibre on OpenStreetMap tiles, permanent OSM/OpenMapTiles attribution, and a Sources / Licenses / Disclaimer screen with the explicit "not an official app" disclaimer.
Done and verified on the emulator. The attribution bar lists only the sources actually rendered — crediting OpenStreetMap while showing the offline fallback would be a false attribution — and sits below the map, inside a SafeArea, so neither a map control nor the system gesture pill can cover a credit the licences require to be visible.
Running it caught two defects the unit tests could not: the configured initial zoom put the viewport entirely inside the region so the map read as a blank expanse, and the attribution bar sat behind the gesture pill. The camera now fits the region bounds at runtime, which works at any screen size.
M3 — Animation and timeline ✅
RadarSource with MockRadarSource (synthetic frames in assets), DpcRadarSource
(reads our CDN) and ArpaRadarSource (disabled stub). Timeline scrubber, play/pause,
adjacent-frame prefetch, FrameCache LRU, animation suspended in background, graceful
degradation with a data-age banner.
Done and verified on the emulator. The overlay is double buffered, the timeline lands on the newest frame, playback advances and wraps, scrubbing takes over from playback, and the legend is drawn from the manifest rather than a constant.
Running it caught three more defects:
- The notifier wrote to
statefrom insidebuild(), which Riverpod rejects as an uninitialised provider. This broke startup, not just tests. - Eight-month-old demo frames rendered as "Aggiornato 342535 minuti fa". The age formatter now steps up to hours and days.
- Demo mode sat permanently behind a stale-data warning, so it never showed the
working state it exists to demonstrate.
MockRadarSourcenow shifts the bundled timestamps onto the present, leaving images, order and spacing untouched.
Frame caching is in memory only for now: mock frames are already in the asset bundle and a disk layer belongs with the network adapter, where it would save a real request.
M4 — Saved places and device location ✅
A named point the user keeps, used to frame the map and — once notifications exist — to
anchor them. Stored in SharedPreferences on the device and nowhere else.
Done and verified on the emulator. The prominent disclosure appears before the system dialog, the permission is optional throughout, places survive an app restart and a reinstall, and tapping one moves the map onto it.
Running it caught three defects the tests could not:
geolocatorinjectsACCESS_FINE_LOCATIONinto the merged manifest, so the system dialog offered "Precise" despite the app declaring only coarse. Removed withtools:node="remove"; the dialog now reads "approximate location" and offers no choice.- The Play Services fused provider prompts about Location Accuracy, and declining it
yields no fix at all. Switched to the platform
LocationManager, which also drops the Play Services dependency. MapLibreMapleavescameraPositionnull unlesstrackCameraPositionis set, so "save the map centre" silently saved the region default instead of what the user was looking at.
A unit test also caught an id collision: ids came straight from the microsecond clock, so two places saved in the same microsecond shared an id and rename, remove and the duplicate check all acted on the wrong one.
M5 — Backend worker ⛔
The critical path. Until this exists, DpcRadarSource has nothing to read and the app
can only show demo frames.
backend/nuvolari_worker/: dpc_client (with the origin header), crop (bbox +
reproject to EPSG:3857 — the source CRS is read from each file, never assumed),
palette (dBZ colormap, legend exported into the manifest), render (RGBA PNG,
transparent below threshold), manifest, and a publisher/ with LocalPublisher and
S3Publisher. Also publishes, so the app never polls ARPA directly:
alerts.jsonfrom the ARPA CAP bulletin;stations.jsonfrom the ARPA realtime API — rain accumulations (1/3/6/12/24 h) and 72 hours of hourly temperature for the 374 stations, joined to their coordinates from/pie_anag. The feed lags about 4.5 hours, so every reading carries its own timestamp and the UI must present it as an observation, never as the current conditions.
WebSocket trigger on wss://radar-wss.protezionecivile.it, with a 5-minute cron as
fallback.
Accepts when: python -m pytest passes and one full run produces PNGs plus a
manifest.json that the app consumes from a local server, showing real Piedmont
precipitation on the emulator.
Blocked on (publishing only): VPS / object storage endpoint and credentials.
Development proceeds against LocalPublisher and a LAN python -m http.server.
M6 — Official alerts
ArpaCapAlertSource reading alerts.json from our CDN. Zones Piem-A…Piem-M with
levels shown verbatim and a link to the official bulletin next to every one.
Six level values, not four: VERDE, GIALLO, ARANCIONE, ROSSO, plus BIANCO
(avalanche scale, out of season) and - (not published). BIANCO and - are their own
state and must never be collapsed into VERDE — that would report "no alert" where the
bulletin reports "not assessed".
Accepts when: the eleven zones render with the levels the live feed carries, the captured fixture parses, and every alert view links the official bulletin.
M7 — Rain notifications ⛔
FCM topics per geographic cell, subscribed from the device, so no user location ever reaches a server. The worker publishes per-cell rain state.
Accepts when: subscribing and unsubscribing works, and a published cell state produces a notification on the emulator.
Blocked on: Firebase project and google-services.json.
M8 — Play Store release preparation ⛔
Signing config reading key.properties, release AAB, target API 36, privacy policy,
store listing copy, Data safety declaration, prominent disclosure for location.
Data safety is unusually simple here: no advertising, no analytics, no accounts, and no location leaving the device.
Blocked on: upload keystore and Play Console account.
Pending inputs from the project owner
| Needed for | Item |
|---|---|
| M5 | VPS / object storage endpoint and credentials |
| M7 | Firebase project and google-services.json |
| M8 | Play Console account and upload keystore |
| ARPA radar adapter | The email to info.meteo@arpa.piemonte.it asking for the real-time access link and the reuse licence. Free of charge is not a licence, and without stated terms the frames cannot be republished. |
Nothing is needed for the base map: OpenFreeMap requires no key and no account.
Verifying on the emulator
An AVD named nuvolari is set up on this machine: API 36.1, google_apis, x86_64,
GPU passthrough to the host. It uses a 2 GB data partition and no SD card because disk
is tight, and quickboot snapshots are disabled — they cost 2.6 GB to save a few seconds
of boot, which is the wrong trade here. Expect a cold boot of a minute or so.
In VS Code, just press F5: every emulator debug configuration boots it first. From a shell:
# Boot it (idempotent — returns at once if a device is attached)
powershell -File tool/start_emulator.ps1
# Build, install, launch
cd app
flutter build apk --debug
adb install -r build/app/outputs/flutter-apk/app-debug.apk
adb shell am start -n it.nuvolari.app/.MainActivity
# Capture what it looks like
adb exec-out screencap -p > screen.png
flutter run -d emulator-5554 works too and gives hot reload; the steps above are what
a non-interactive session uses.