agentleFS
zakariafzakariaf/Flutter-Skills/skills/state-management-riverpod/SKILL.md

state-management-riverpod

Skill32 starsChanged 4 months agoMITView on GitHub

Enforces feature state as one Notifier/AsyncNotifier/StreamNotifier ViewModel over an immutable state value with value equality, private mutable state mutated only through intent methods, derive-don't-store, a single write path through a repository, and unidirectional data flow; Riverpod 3.x is DI (providers-as-collaborators, ProviderScope overrides per flavor, throwing seams), reads split ref.watch/.select for display vs ref.read(p.notifier) in callbacks vs ref.listen for side effects, async modeled as AsyncValue, per-entity state family-keyed + autoDispose, and stale-closure captures / legacy StateProvider-StateNotifierProvider-ChangeNotifierProvider / get_it / package:provider are banned. Use when adding state to a screen, writing a feature controller/ViewModel, wiring providers or DI, deriving a read model from a stream, or reviewing rebuild/state-leak/disposal/write-path issues.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read the reference for the task at hand:

    line 10
  2. Very specific:

    Run scripts/ban-legacy-providers.sh before a PR.

    line 16
  3. Very specific:

    State is private; mutate only through intent methods: No widget reaches into the state to poke a field.

    line 23
  4. Distinctive:

    Single write path: Every durable mutation is one repository method that persists transactionally and returns only after commit.

    line 26
  5. Very specific:

    Default to hand-written providers (StreamNotifierProvider/AsyncNotifierProvider(...), .autoDispose/.family modifiers).

    line 45
  6. Very specific:

    Before adding a provider, ask whether a constructor argument would do.

    line 107
  7. Very specific:

    The stale-closure hole: never capture a ref.watched value into an onTap closure — pass a stable key and resolve at tap time via ref.read.

    line 128

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

State management (Riverpod)

State and dependency injection are one mechanism: Riverpod 3.x. Each feature has one Notifier/AsyncNotifier/StreamNotifier ViewModel that exposes an immutable state value; dumb ConsumerWidgets read it. Widgets hold only ephemeral UI state; every durable mutation routes through one repository (the single write path). This skill covers the state-agnostic core first, then the Riverpod "how", then a Provider/ChangeNotifier appendix for the official Flutter-guide stack.

Read the reference for the task at hand:

  • references/ownership-and-lifecycle.md — the ownership table (which provider shape), family/autoDispose/onDispose rules, composition-root DI.
  • references/reads-and-side-effects.md — watch vs read vs listen vs select, the stale-closure hole, void action methods, ref.mounted/BuildContext guards.
  • references/riverpod3-api-and-testing.md — Riverpod 3.x API shifts (legacy moves, overrideWithValue, retry, ProviderContainer.test), what 2023 tutorials get wrong, and testing seams.

Run scripts/ban-legacy-providers.sh before a PR.

Non-negotiable rules

These hold regardless of the state library.

  1. One ViewModel per feature over one immutable state value. The state is a value type with value equality (freezed/sealed + copyWith), never a mutable field bag. Cross-feature/shared state lives in a repository, not a ViewModel — two ViewModels owning the same fact is two facts that will disagree.
  2. State is private; mutate only through intent methods. No widget reaches into the state to poke a field. A view calls notifier.rename(id, name); it never sets state. This is the one door through which state changes, so every change is reviewable in one place.
  3. Every transition assigns a new value. Value-equality listeners diff by value — a mutated-in-place instance reassigned to the same reference is missed and the UI silently stales. Always copyWith into a new instance.
  4. Derive, don't store. A value computable from existing state (a total, a filtered count, a streak) is a getter or a stream projection, never a second stored field — a stored derivation is a second source of truth that drifts out of sync.
  5. Single write path. Every durable mutation is one repository method that persists transactionally and returns only after commit; the UI updates because a stream re-emits, not because you optimistically republished. Persist-before-publish makes crash-safety structural, not a matter of discipline.
  6. Unidirectional data flow. Intent → ViewModel → repository/service → new state → view. The view never mutates data it reads and never short-circuits back up the chain. Data flows down; events flow up.
  7. Depend on abstractions, injected — never constructed. A ViewModel receives its repositories/services/clock; it never news a live collaborator and never holds a BuildContext. This is what makes it testable with fakes and swappable per flavor.
  8. Model async as an explicit state, not loose flags. One AsyncValue (or a status enum + data + error) rendered exhaustively — never scattered isLoading/hasError booleans that can encode a half-set, impossible state.
  9. Cascade-clean references on delete. Deleting an entity must drop every reference to it (assignments, foreign keys, selection) — an item left pointing at a deleted parent is a latent crash. Never silently drop the orphan; re-home or unassign it explicitly.
  10. No DateTime.now() in state logic. "Now" enters through an injected Clock (from package:clock) exposed as clockProvider, so time-dependent behaviour is deterministic in tests (clockProvider.overrideWithValue(Clock.fixed(t))). Never DateTime.now(), never a bespoke ClockService. service-boundary-and-native owns the clockProvider seam.

Riverpod: pick the right shape

Match the provider shape to what it owns (full table in references/ownership-and-lifecycle.md):

OwnsShape
Per-feature presentation state over a live queryStreamNotifier whose build() returns the repository stream
Per-feature presentation state, non-streamNotifier / AsyncNotifier over one immutable value
A reactive projection of the DBStreamProvider over a repository/DAO stream (never stored)
An injected collaborator (repo, service, Clock)plain Provider (DI)
Per-entity/per-session stateNotifier/AsyncNotifier/StreamNotifier .family + .autoDispose

Default to hand-written providers (StreamNotifierProvider/AsyncNotifierProvider(...), .autoDispose/.family modifiers). @riverpod codegen is an optional convenience, not the default (see the reference). Riverpod 3.0 API only — base classes Notifier/AsyncNotifier/StreamNotifier; autoDispose/family are provider modifiers, never AutoDispose*/*Family base classes (removed in 3.0).

// features/task — one StreamNotifier ViewModel: build() returns the LIVE repo stream,
// so a committed write re-emits and the UI updates with no manual republish (D6).
class TaskListNotifier extends StreamNotifier<TaskListState> {
  @override
  Stream<TaskListState> build() {
    final filter = ref.watch(taskFilterProvider);          // in-session filter, its own Notifier
    return ref.watch(taskRepositoryProvider).watchAll()    // the live source of truth
        .map((tasks) => TaskListState(tasks: tasks, filter: filter));
  }

  // Durable act routes through the single write path. void so the call site never drops
  // the Future; no `state = ...` — the committed write makes watchAll() re-emit.
  void complete(TaskId id) =>
      unawaited(ref.read(taskRepositoryProvider).markComplete(id).catchError(_report));

  void _report(Object error, StackTrace stack) {/* surface via a logger — never swallow */}
}

final taskListNotifierProvider =
    StreamNotifierProvider.autoDispose<TaskListNotifier, TaskListState>(
        TaskListNotifier.new);

// The reversible in-session filter is its own tiny Notifier (manual NotifierProvider).
class TaskFilterNotifier extends Notifier<TaskFilter> {
  @override
  TaskFilter build() => TaskFilter.all;
  void set(TaskFilter f) => state = f; // watched by build() above; re-projects the stream
}

final taskFilterProvider =
    NotifierProvider<TaskFilterNotifier, TaskFilter>(TaskFilterNotifier.new);

Riverpod: providers are DI (throwing seams)

A ViewModel reads collaborators from providers; the composition root wires the live impls once. A placeholder seam throws until overridden, so an un-wired dependency fails loudly at first read instead of silently constructing a real service inside a test.

// A seam: throws until a flavor main() overrides it. One live impl per flavor.
final databaseProvider = Provider<AppDatabase>(
  (ref) => throw UnimplementedError('override databaseProvider in main()'),
);

final taskRepositoryProvider = Provider<TaskRepository>(
  (ref) => TaskRepository(ref.watch(databaseProvider)), // inject, never `new` a live DB here
);
// main.dart — the ONLY place live collaborators are constructed.
void main() {
  final db = AppDatabase.open();
  runApp(ProviderScope(
    overrides: [databaseProvider.overrideWithValue(db)],
    child: const App(),
  ));
}

Before adding a provider, ask whether a constructor argument would do. Provider count going up is a smell, not progress — reach for family/scoping/codegen only when a plain provider genuinely cannot express the need.

Riverpod: reads — watch / read / listen

Full rationale in references/reads-and-side-effects.md.

  • ref.watch(p) / ref.watch(p.select((s) => s.field)) — in build(), for display. .select narrows the rebuild to one field. Never watch a whole controller at the top of a large widget.
  • ref.read(p.notifier) — in callbacks (onTap, onPressed). A read of a value in build() freezes on stale data; a watch in a callback rebuilds unexpectedly.
  • ref.listen(p, ...) — in build(), for side effects (navigate, snackbar) on change. Never fire a side effect directly from build().
class TaskCounter extends ConsumerWidget {
  const TaskCounter({super.key});
  @override
  Widget build(BuildContext context, WidgetRef ref) {
    final n = ref.watch(taskListNotifierProvider.select((s) => s.value?.openCount ?? 0));
    return Text('$n');
  }
}

The stale-closure hole: never capture a ref.watched value into an onTap closure — pass a stable key and resolve at tap time via ref.read. A captured value silently acts on the previous entity after a fast re-tap, and no lint catches it.

// WRONG — acts on a stale item after a re-render.
onTap: () => ref.read(p.notifier).complete(task),
// RIGHT — id is stable; resolve now.
onTap: () => ref.read(p.notifier).complete(task.id),

Action-path methods return void. An arrow closure onTap: () => notifier.doThing() that "returns" a Future satisfies neither discarded_futures nor unawaited_futures (target type is VoidCallback), so the Future and its error are dropped. A void intent method makes the hole unreachable; kick off async inside it with unawaited(_run().catchError(_report)). See async-safety.

Riverpod: lifecycle & disposal

  • Per-entity/per-session state is .family-keyed by a stable equatable value (an id or a @freezed args record), never a mutable object, and .autoDispose so heavy session state dies on unmount.
  • App-scope singletons (database, engine, services) are plain Provider — never autoDispose.
  • Guard writes after an await with ref.mounted (the Riverpod analogue of use_build_context_synchronously).
  • Release owned resources in ref.onDispose (streams, controllers, native handles). Riverpod owns a StreamProvider's subscription, so don't hand-manage it.

Riverpod: accessibility state does NOT go through providers

Read MediaQuery.boldTextOf(context) / .textScalerOf(context) / .highContrastOf(context) at build time in the widget, never via a provider. MediaQuery is already an InheritedWidget with correct-by-construction invalidation; routing it through a provider means either a BuildContext in a provider body or a one-frame-stale hand-sync. App/domain state via Riverpod; platform/a11y state via BuildContext. See accessibility-as-code.

Anti-patterns

  • Legacy providers — StateProvider / StateNotifierProvider / ChangeNotifierProvider (moved to flutter_riverpod/legacy.dart), or adding get_it / package:provider / Bloc alongside Riverpod. Two DI/state mechanisms is strictly worse than one.
  • Business logic in a provider body. A provider is a thin wire; logic lives in the ViewModel, engine, and repository.
  • Mutating the state value in place and reassigning the same instance — value-equality listeners won't see it.
  • ref.watch at the top of a big widget rebuilding the whole subtree on any field change. Use .select or a leaf ConsumerWidget.
  • App/domain state in setState(). setState/local State is for ephemeral UI only (a toggle, an in-flight animation flag).
  • Optimistic republish before commit — a crash between republish and commit shows a fact the disk never held.
  • A ViewModel that news its collaborators, holds a BuildContext, navigates, or shows snackbars. Inject; publish state; let the view (or a router redirect) react.
  • family-keying on a mutable object, or autoDispose on an app-scope singleton, or keepAlive() to undo an autoDispose you shouldn't have added.

Definition of done

  • Feature state is one Notifier/AsyncNotifier/StreamNotifier over an immutable value; widgets hold only ephemeral UI state.
  • State is private; every mutation goes through an intent method that assigns a new value; no in-place mutation.
  • Derived values are getters/stream projections, not stored fields.
  • Async is one AsyncValue/status rendered exhaustively; no loose loading/error booleans.
  • Every durable mutation is a repository method that commits before returning; no optimistic pre-commit republish.
  • Reads use ref.watch/.select for display, ref.read(p.notifier) in callbacks, ref.listen for side effects; no broad top-level watch; no captured-value closures.
  • Collaborators injected via providers; seams throw until overridden; live impls wired once per flavor in main; no BuildContext/DateTime.now() in the ViewModel.
  • Per-entity providers are family-keyed by a stable value and autoDisposed; app singletons are plain providers; resources released in ref.onDispose.
  • No legacy.dart, get_it, package:provider, or Bloc import (scripts/ban-legacy-providers.sh passes).

Related skills

  • flutter-architecture — where ViewModels, repositories, and the write path sit in the feature-first DAG.
  • error-handling-typed-results — the Result/Failure spine a repository method returns instead of throwing.
  • async-safety — the Future-drop hole, mounted guards, subscription/timer disposal.
  • persistence-drift — the Drift streams the StreamProviders project and the one-transaction-per-mutation write path.
  • service-boundary-and-native — the throwing-seam provider pattern for every side effect and native channel; owns the clockProvider (package:clock) seam these ViewModels read instead of DateTime.now().
  • widget-composition — the dumb ConsumerWidget Views that read this state.
  • accessibility-as-code — the split this skill co-owns: app/domain state via Riverpod, platform/a11y state read from MediaQuery/BuildContext, never a provider.
  • value-objects-money-and-units — value types that take a Clock param for deterministic time.
  • app-startup-and-bootstrap — the composition root where ProviderScope overrides are wired.

References

Provider / ChangeNotifier appendix

For the official Flutter-guide stack (package:provider + ChangeNotifier), the same ten rules hold; only the mechanism changes.

  • ViewModel = ChangeNotifier; new state = notifyListeners() after all fields are set (exactly once per logical change, never mid-update or in a loop).
  • Private state, exposed immutably. Fields are _private; expose read-only getters; return List.unmodifiable(...) — never hand out a mutable collection.
  • DI = constructor injection of abstractions, wired with Provider/ProxyProvider above the screen (auto-disposed on pop). Do not use get_it as the container.
  • Reads = context.select/Consumer/Selector for display, context.read in callbacks. Read in didChangeDependencies or callbacks, never initState.
  • Guards = check a _disposed flag before notifyListeners() after an await; cancel subscriptions and dispose owned controllers in dispose().
enum ViewStatus { idle, loading, ready, error }

class TaskListViewModel extends ChangeNotifier {
  TaskListViewModel(this._repo);            // injected abstraction
  final TaskRepository _repo;

  ViewStatus _status = ViewStatus.idle;
  ViewStatus get status => _status;
  List<Task> _tasks = const [];
  List<Task> get tasks => List.unmodifiable(_tasks); // immutable view
  int get openCount => _tasks.where((t) => !t.done).length; // derive, don't store

  Future<void> load() async {
    _status = ViewStatus.loading; notifyListeners();
    try {
      _tasks = await _repo.load();
      _status = ViewStatus.ready;
    } catch (_) {
      _status = ViewStatus.error;
    }
    notifyListeners();                        // exactly one, after all fields set
  }

  Future<void> complete(TaskId id) async {
    await _repo.markComplete(id);             // single write path
    _tasks = [for (final t in _tasks) t.id == id ? t.copyWith(done: true) : t];
    notifyListeners();
  }
}

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.