Ramadhan Sjamsani
368d18a0bf
Verified the 2026-05-15 disconnect() fix end-to-end on emulator-5556: mitra logs in → online → accepts blast → backend force-expires → goodbye composer renders → back-press → lands on Bestie Home with online status preserved, zero flutter:E in logcat. - ts-mitra-3-08-back_press_after_session_expired_no_red_screen.yaml codifies the repro for Maestro. Extends ts-mitra-3-04 with the back-tap + home-assertion + red-screen guard. - mitra_app/CLAUDE.md adds a Pitfall section beneath the existing "no ref in dispose" rule: never mutate notifier state synchronously from deactivate() cleanup — wrap in SchedulerBinding.addPostFrameCallback or Riverpod throws "Tried to modify a provider while the widget tree was building" during the back-nav teardown. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
66 lines
3.8 KiB
Markdown
66 lines
3.8 KiB
Markdown
# Halo Bestie — Mitra App
|
|
|
|
Flutter mobile application for mental health professionals (mitra/partners).
|
|
|
|
> See root `CLAUDE.md` for full project context and architectural decisions.
|
|
|
|
## Stack
|
|
|
|
- **Framework:** Flutter (iOS + Android)
|
|
- **Auth:** Self-managed (Phase 3.4). Phone OTP only — no Google / Apple. Access token lives in memory on an `AuthBridge`; refresh token persists in `flutter_secure_storage`. `firebase_auth` is no longer used; `firebase_messaging` is kept for FCM push.
|
|
- **API:** Calls public Fastify backend (`/api/mitra/` and `/api/shared/` routes). `shared.auth` covers refresh + logout for both apps.
|
|
|
|
## Key Concepts
|
|
|
|
- Users are **mitra** — trained mental health professionals
|
|
- Core flow: phone OTP login → set availability → accept sessions → chat with client → receive payment
|
|
- Mitra accounts require approval from control center before going live (backend returns `ACCOUNT_INACTIVE` 403 on OTP verify when `is_active=false`)
|
|
|
|
## Conventions
|
|
|
|
- Never call `/api/client/` or `/internal/` routes from this app
|
|
- API calls go through `ApiClient`; it auto-attaches the JWT from `AuthBridge` and auto-refreshes on 401
|
|
- WebSocket handshake (`/api/shared/ws`) sends the same access token in the first frame's `{type:"auth", token}` message
|
|
- Mitra role is encoded in the JWT claims (`user_type: "mitra"`) — the backend enforces the role per route; never trust client state alone
|
|
|
|
## Pitfalls (HARD rules — silent failure modes)
|
|
|
|
### Never call `ref.read` / `ref.watch` / `ref.listen` from `State.dispose()`
|
|
|
|
In a `ConsumerStatefulWidget`, Riverpod invalidates `ref` the instant `dispose()` starts. Any `ref.*` call throws `Bad state: Cannot use "ref" after the widget was disposed.`. Flutter catches it inside `BuildOwner.finalizeTree` — **so it does not surface as a red-screen crash**. Instead the widget tree is left half-finalized and the NEXT screen freezes (looks like a hang; the app process is alive). Real case in this app: `mitra_chat_screen.dart` (2026-05-14).
|
|
|
|
**Rule:** any cleanup that needs `ref` goes in `deactivate()`, which runs *before* `dispose()` while `ref` is still valid. Non-Riverpod cleanup (`TextEditingController.dispose()`, `WidgetsBinding.removeObserver`, `StreamSubscription.cancel`) stays in `dispose()`.
|
|
|
|
```dart
|
|
@override
|
|
void deactivate() {
|
|
ref.read(someProvider.notifier).cleanup();
|
|
super.deactivate();
|
|
}
|
|
|
|
@override
|
|
void dispose() {
|
|
WidgetsBinding.instance.removeObserver(this);
|
|
super.dispose();
|
|
}
|
|
```
|
|
|
|
A lint rule (`no_ref_in_dispose` in `halo_lints`) fails `dart run custom_lint` on this pattern. When debugging "screen frozen after navigation", grep the *previous* screen's State for `void dispose()` followed by `ref\.` — that's the first suspect.
|
|
|
|
### Never mutate notifier `state` synchronously from `deactivate()` cleanup
|
|
|
|
`deactivate()` is the safe place to call `ref.read(...).cleanup()`, but if that cleanup does `state = SomeData()`, Riverpod will notify watchers while the widget tree is mid-teardown → `Tried to modify a provider while the widget tree was building.` → red error screen + engine restart (in debug) or silent state loss (in release). Real case: `mitra_chat_notifier.dart::disconnect()` was firing this on back-press from the session-ended chat screen.
|
|
|
|
**Rule:** when a notifier's cleanup-from-`deactivate()` path needs to reset its own state, defer the assignment to the next frame:
|
|
|
|
```dart
|
|
void disconnect() {
|
|
_cleanup(); // closing channels / cancelling subs is fine synchronously
|
|
SchedulerBinding.instance.addPostFrameCallback((_) {
|
|
state = const SomeInitialData();
|
|
});
|
|
}
|
|
```
|
|
|
|
The synchronous side-effects (closing the WS, cancelling timers) still happen immediately. Only the `state =` assignment is deferred, which is a no-op for users — they're navigating away anyway. Regression coverage: `.maestro/flows/ts-mitra-3-08-back_press_after_session_expired_no_red_screen.yaml`.
|