Splint

Learn Splint interactively: Naming the Shapes on artifact.land.

Named data types for SwiftUI apps. Splint is a small library that gives every data shape in your app a name — so that an agent (or anyone else) reaching for a value reaches for the right kind of value. It is corrective, not prescriptive: it immobilizes the data skeleton so it heals correctly.

Agents building SwiftUI apps consistently produce a god-object @Observable class with 15+ properties that every view observes, causing cascading re-renders across the view hierarchy. The root cause is not data flow — it's data modeling. Splint names the types.

Install

dependencies: [
    .package(url: "https://github.com/searlsco/splint", from: "0.1.0")
]

Then drop the agent rules file into your project:

mkdir -p .claude/rules
curl -fsSL https://raw.githubusercontent.com/searlsco/splint/main/claude/rules/splint.md \
  -o .claude/rules/splint.md

Commit the resulting .claude/rules/splint.md. Re-run the curl to pick up rule changes when you bump the Splint package version.

Quick start

import Splint

struct Book: Resource {
    let id: String
    let title: String
    let author: String
}

struct BookCriteria: Equatable, Sendable {
    let libraryID: String
}

let catalog = Catalog<Book, BookCriteria> { criteria in
    try await api.fetchBooks(in: criteria.libraryID)
}

// Or seed from a disk cache so the UI renders immediately on cold launch.
// The seed stays visible through the first load() until the fetch lands.
//   Catalog<Book, BookCriteria>(initialItems: cachedBooks) { criteria in … }

catalog.load(BookCriteria(libraryID: "main"))

let favorites = Lens<Book>(source: catalog, filter: { $0.isFavorite })

The type inventory

Type	What it holds	Observation	Persistence	Mutability
Resource	Decoded remote data (channels, programs, episodes)	None — value type	None	Immutable after decode
Catalog	Ordered collection of Resources loaded by criteria, plus fetch lifecycle	@Observable	None (in-memory cache)	Collection mutates on load/refresh
Lens	Filtered/sorted/grouped view over a Catalog	@Observable	None	Criteria mutate; data is derived
Job	Async operation lifecycle (idle → running → completed/failed)	@Observable	None	Phase mutates as work progresses
Selection	Currently selected item identifier	@Observable	None	Mutates on user tap
Setting	Single typed user preference	@Observable	UserDefaults	Mutates, persists automatically
Credential	Keychain-backed secret	None — read on demand	Keychain	Mutates via explicit save/delete

There is no ViewState type. If a value doesn't fit one of the types above, it's @State on the view that owns it. If two views need the same value, make a small purpose-built @Observable class with 1–2 fields. If that class grows past 3 fields, you've created a new category — name it and split it. General-purpose "view state" containers become god-objects.

This is not a state management framework

No reducers, no action enums, no effect types, no stores. The architecture is SwiftUI's own — @Query, @Environment, @State, @Observable — Splint just names the things you put in those slots.

How observation boundaries work

@Observable tracks at the stored-property level, per instance. Each view's body runs in its own withObservationTracking scope. A child view receiving data via init creates an independent observation boundary — it only tracks properties it reads in its own body.

• Value types (Resource structs): no observation tracking. A child view receiving let book: Book avoids re-evaluation unless the parent rebuilds and passes a structurally different value.
• @Observable class instances: the child view tracks only the specific properties it reads. A detail view and a row view can hold the same instance — when one mutates a property the other view sees the change through its own observation. No sync, no events.

The anti-pattern: reading properties from an @Observable object inside a ForEach closure. This registers the dependency on the parent's observation scope, widening it across the whole list.

The fix: always extract the row body into a named child view. Pass data via init.

// ❌ Widens parent's observation scope across all rows.
ForEach(catalog.items) { item in Text(item.name) }

// ✅ Each ItemRowView has its own observation scope.
ForEach(catalog.items) { item in ItemRowView(item: item) }

One store per entity, not one store per view. When a list row and a detail view display the same entity, they hold the same instance. Do not create ItemRowStore and ItemDetailStore.

These are architectural guidelines based on how Observation is designed to work, not guaranteed rendering outcomes. Use the SwiftUI Instruments template to verify performance in your specific app.

SwiftData entities are already correct

@Model includes @Observable. Pass instances directly to child views. Do not wrap them in ViewModels or Splint types. Use @Query in views.

@Model types are not Sendable — they're bound to a ModelContext and cannot cross actor boundaries, so they cannot be held in a Catalog. When a view needs both @Model data and Catalog data, join them at the view level (e.g. @Query for favorites, catalog lookup by ID for the corresponding resource).

When you don't need Splint

If your feature is just @Query → List → detail, write plain SwiftUI. Splint earns its keep when you have remote API resources, async lifecycles, preferences, or derived collections.

Choosing the right type

• Decoded remote data → Resource
• Collection of Resources → Catalog
• Filtered/sorted view of a Catalog → Lens
• Async fetch lifecycle → Job
• Selected item identifier → Selection
• User preference → Setting
• Secret → Credential
• SwiftData entity → @Model + @Query (not a Splint type)
• Presentation state (sheet, alert, popover) → @State on the presenter
• Transition/animation state → @State on the animating view
• Draft/form input → @State on the form view
• Player/media state → dedicated @Observable with 1–2 fields
• Anything else shared across views → small purpose-built @Observable with 1–2 fields. If it grows past 3, name the new category and split.

There is no general-purpose "view state" container. General containers become god-objects.

Catalog.load(_:) vs refresh()

The most important API distinction in Splint:

• catalog.load(newCriteria) — criteria changed; old items are wrong (channel A's programs when you asked for channel B). Clears items immediately so the view shows loading, not wrong data. The one exception is the very first load(): the prior criteria was nil, so there are no "wrong" items to wipe — any seeded items from initialItems: stay visible until the fetch lands.
• catalog.refresh() — same criteria; items stay visible during fetch. This is pull-to-refresh, periodic polling, "show stale and update in place." No-op if load() has never been called.
• catalog.retry() — alias for refresh(); reads better after failure.

Pair with SwiftUI:

.task(id: channel.id) {
    catalog.load(ProgramCriteria(channelID: channel.id))
}

SwiftUI cancels and relaunches the task when the id changes — free cancellation of the old fetch.

Catalog lifecycle

Scope the catalog to the narrowest view that fully contains its usage. Every catalog is @State on the owning view, distributed via .environment(). The catalog lives exactly as long as that view.

• Session-scoped (channels for the logged-in provider): @State on the authenticated root view — not the app root. Logout destroys the authenticated root, deallocates the catalog, and cancels in-flight fetches via deinit. Logging into a different provider creates a fresh catalog with no stale data.
• Navigation-scoped (programs for a specific channel): @State on the detail view. Created on push, destroyed on pop.
• Persistent-detail (iPad sidebar/detail, macOS split view): @State on the detail view that survives selection changes. Driven by .task(id: selectedItem.id) { catalog.load(newCriteria) }. The instance persists across selections — this is where the criteria-clearing branch in load() earns its keep, preventing channel A's programs from showing while channel B's load.

The catalog doesn't decide its own lifecycle — SwiftUI does, based on where you store the instance.

The fetch closure captures its dependencies at construction time. DI happens before the Catalog exists, not after:

init(client: BookClient) {
    self._catalog = State(initialValue: Catalog(fetch: client.fetchBooks))
}

No singletons, no late-binding, no service locators. The same rule applies to Lens filter/sort closures — see "Lens closures capture once" below.

Seeding from a cache

Cold launch shouldn't have to flash an empty view for the 500ms–2s the first fetch takes. If you have a disk-cached snapshot of the last successful fetch, pass it to Catalog via initialItems::

let cached: [Book] = cache.load(key: "books", as: [Book].self) ?? []
let catalog = Catalog<Book, BookCriteria>(initialItems: cached) { criteria in
    try await api.fetchBooks(in: criteria.libraryID)
}

Semantics:

• catalog.items is non-empty from moment zero — any Lens or GroupedLens built on top sees the seed immediately.
• catalog.phase stays .idle until a real load() completes. Seeding is not a completed fetch.
• The first load() preserves the seed until the fetch lands, so the view doesn't flash empty between "we showed the cache" and "the network answered."

Narrowly-scoped catalogs over client-side filtering

Large datasets should be scoped via criteria, not fetched in bulk and filtered client-side with Lens:

struct ProgramGuideView: View {
    let channel: Channel
    @State private var programs: Catalog<Program, ScheduleCriteria>

    init(channel: Channel, api: IPTVClient) {
        self.channel = channel
        self._programs = State(initialValue: Catalog { criteria in
            try await api.fetchPrograms(for: criteria.channelID, on: criteria.date)
        })
    }

    var body: some View {
        ProgramList(programs: programs)
            .task {
                programs.load(ScheduleCriteria(channelID: channel.id, date: .now))
            }
    }
}

The API returns programs for one channel on one date. The catalog holds hundreds, not thousands. If you need filtering within that result (e.g. by genre), Lens is appropriate at this scale.

Lens performance

Lens recomputes the full filtered array when the source catalog's items change — O(n) where n is the source size. Under ~1k items this is negligible. For larger datasets, consider whether filtering belongs in the fetch closure (server-side filtering via Criteria) rather than in a client-side Lens.

Lens closures capture once

Lens captures filter and sort at init. They re-run only on updateFilter / updateSort — not when variables they reference change. Capturing mutable view state at init compiles and looks correct, and fails silently:

// ❌ minRating capture goes stale — Lens never sees changes
@State private var minRating = 0
@State private var lens = Lens(source: catalog, filter: { $0.rating >= minRating })

// ✅ Drive updates explicitly
@State private var minRating = 0
@State private var lens = Lens(source: catalog)

var body: some View {
    BookListView(lens: lens)
        .onChange(of: minRating) { _, new in
            lens.updateFilter { $0.rating >= new }
        }
}

Same rule as Catalog's fetch closure: Splint closures capture at construction; mutable inputs flow through update methods.

Job closures and isolation

Job.run's task closure runs in its own Task. Because task: is sending (SE-0430), the closure can capture non-Sendable values at the call site — including self of a SwiftUI View, whose property wrappers usually make the enclosing struct non-Sendable. Region-based isolation accepts the closure's disconnected copy of the captured values.

Capture whatever's needed — a Sendable service, self's init-time let properties, a value-type input. What to avoid inside the closure body:

• Reading @State, @Query results, or @Environment values from inside the closure body. Even when the capture compiles, you get a frozen snapshot taken when the closure was created — later wrapper updates are invisible. Capture the specific IDs or scalars you need at the call site instead.
• Mutating @MainActor state directly. The Task runs off the main actor, so the compiler usually blocks this. When you genuinely need to mutate main-actor state after the await, hop back explicitly:

metadataJob.run {
    let fresh = try await client.fetchMetadata(book.id)
    await MainActor.run { cache[book.id] = fresh }
    return fresh
}

// ✅ Capture self; read stable init-time `let` properties.
.task {
    metadataJob.run { try await client.fetchMetadata(book.id) }
}

// ✅ Equivalent; explicit capture for readability.
.task {
    let client = client
    let id = book.id
    metadataJob.run { try await client.fetchMetadata(id) }
}

// ❌ Reads @Query results inside the Task. `sending` may allow the
// capture, but `favorites` is a snapshot frozen at closure creation —
// later @Query updates never reach this closure. Pull what you need
// out at the call site and pass it in as a scalar.
.task {
    metadataJob.run {
        let fav = favorites.contains { $0.bookID == book.id }
        return try await client.fetchMetadata(for: book.id, favorited: fav)
    }
}

Session coordination

Login, logout, and reauthentication require atomic coordination across multiple fields (token write + provider switch + player teardown). This is not a Splint type — it's a plain @Observable class with a login() method that performs the coordination. It should be @State on the authenticated root view and destroyed on logout. Splint types (Credential, Setting, Catalog) are the fields it coordinates, not a replacement for the coordinator itself.

Settings sync across instances and processes

Setting observes its key via UserDefaults key-value observation, so:

• Multiple instances on the same key stay in sync. Two Settings bound to the same key/store see each other's writes automatically. No "single owner per key" convention required — instance count is a perf footnote, not a correctness concern.
• App Group suites sync across processes. A Setting backed by UserDefaults(suiteName: "group.example.shared") stays in sync between the host app and its extensions (widgets, intents, share extensions). This is Apple's userdefaultsd behavior for entitlement-granted App Groups, surfaced through standard KVO — Splint adds no code of its own for cross-process notification.
• Dotted keys sync same-process only. KVO treats a dot in a keyPath as nested key-path traversal, so it never fires for a defaults key like "learner.targetLanguage". Setting detects the dot and falls back to UserDefaults.didChangeNotification, which keeps instances in sync within one process but does not observe other processes. If you need App Group cross-process sync, use a dot-free key.

What Splint won't fix

Splint addresses data structure. It does not address SwiftUI rendering performance (symbol effects, navigation transitions, column layout behavior), view lifecycle timing, or platform-specific layout bugs. If your performance problem is in the render layer, Instruments' SwiftUI template is the right tool — not a data modeling library.

What agents get wrong

Agent mistake	Splint type that prevents it
God-object ViewModel with 15+ properties	Named types split data by kind
isLoading / error / data as separate booleans	Job<Value> with phase + value
Duplicate arrays (source + filtered copy, manually synced)	Lens derives from Catalog
Showing stale wrong data after parameter change	Catalog.load() clears items when criteria change
selectedItem on a 15-field observable	Selection<ID> — one value, one observation point
Credential stored in an observable property	Credential is a struct, read on demand
UserDefaults scattered across the app	Setting<Value> — one key, one observation point
Wrapping SwiftData models in ViewModels	Documentation: use @Model directly
Reading child properties in ForEach closure	Documentation + example: extract to child view
Catch-all "view state" objects that grow over time	No ViewState type exists — forces decomposition

Agent rules

The canonical agent guidance lives at claude/rules/splint.md in this repo. See the install section above for how to drop it into a consuming project.

Coverage

script/test enforces 100% line coverage on the Splint target. This is a forcing function, not a metric: if a line can't be covered by a meaningful behavioural test, the line should be deleted, restructured to be testable, or tagged with an inline exclusion marker whose rationale (≥10 characters) names the specific reason:

foo()  // coverage:ignore — <why this line can't be exercised>

Blocks use // coverage:ignore-start — <rationale> and // coverage:ignore-end. Padding coverage with tests that exercise a line without verifying behaviour defeats the purpose.

License

MIT. See LICENSE.