--- name: api-design description: Design and review predictable public APIs for TypeScript, JavaScript, React, and React Native libraries. Use when shaping exported functions, classes, hooks, options objects, event and listener APIs, error behavior, naming, cross-platform abstractions, or JS-only packages. Pair with build-nitro-modules when the library is backed by Nitro. ---

# API Design

Use this skill before implementation or when reviewing a public API surface. Target explicit types, stable semantics, and no irrelevant internals in the public contract.

If the library is a Nitro Module, use this skill for the public TypeScript and React API shape first, then use `build-nitro-modules` for Nitro-specific spec, native-state, and binding constraints. If the library is JS-only, React-only, or React Native JS-only, stay in this skill.

## Workflow

1. Sketch the user-facing TypeScript API before implementing internals. 2. Write 2-3 realistic call-site examples, including error and cleanup paths. 3. Check the surface against the rules below. 4. Verify the exported TypeScript with the repo's typecheck/lint/docs tooling before treating the API as done. 5. Implement only after the public shape is coherent and verified.

## API Freshness

Before choosing public API shape, dependency APIs, platform capabilities, or implementation strategy, verify current official sources instead of relying on trained memory. Library, React, React Native, platform, and tooling APIs evolve quickly.

- Prefer official docs, source repositories, release notes, changelogs, package READMEs, and current package metadata. - Look for `llms.txt` or `llms-full.txt` on official docs sites when available, and use those as compact current context. - Treat remembered API details as a starting hypothesis only. If current docs or source disagree, follow the current docs/source and mention the change when relevant. - Avoid designing against stale blog posts, old snippets, or outdated trained assumptions when an official current source is available.

## API Shape Rules

- Prefer a single source of truth. Do not split related state across booleans and dependent values when one typed value can express the state. Prefer `timeoutMs?: number` over `enableTimeout: boolean` plus `timeoutMs?: number`. - Use option objects or named structs once a function has 3 or more parameters, parameters of the same primitive type, or values that are likely to grow. - Keep APIs specific instead of accepting every possible input shape. A millisecond timeout should be a `number`, not `number | string | bigint | object | null`. - Avoid giant "does everything" objects. Split by domain or lifecycle when responsibilities differ. - Before simplifying or redesigning an API, inventory the workflows the feature is supposed to support. Do not silently drop a workflow, such as a live session API, because a one-shot path is easier to implement. Split workflows into separate APIs when needed. - Prefer literal unions, discriminated unions, interfaces, and typed option groups when the valid states are known. In TypeScript libraries, prefer string literal unions over runtime `enum`s unless consumers need a runtime value. - Avoid untyped dictionaries, boolean clusters, stringly typed commands, and loosely shaped events when the valid states are known. - Do not model a binary option as an optional two-case string union such as `'enabled' | 'disabled'`. If omitted means "use the default" and provided means true/false, use an optional boolean and document the default. - Do not represent multiple object states as one interface full of optional fields. Use a discriminated union, inheritance, or separate variant interfaces so impossible field combinations are unrepresentable. - Keep related fields together on the variant where they are required. If `barcode` and `barcodeType` only make sense together, both should be nonoptional on `ScannedBarcode`, not optional on a generic `ScannedData`. - If you create variant interfaces, use them in the actual public type. Do not define `RecognizedTextDataType` and `RecognizedBarcodeDataType` but keep `recognizedDataTypes: RecognizedDataType[]` where `RecognizedDataType` still contains every variant field as optional. - Use `undefined` or optional fields for absence. Use `null` only when "explicit none" means something different from "not provided". - Prefer discriminated unions for state machines, loading states, and result variants. - Model user intent separately from resolved state when negotiation is involved. For example, an ordered array of constraint objects can express priorities, while a resolved config object reports what the platform actually selected. - For complex negotiation, prefer ranked constraints or preference objects over exposing a combinatorial support matrix. Let callers describe intent, resolve the closest working configuration internally, and expose the resolved configuration through a return value, callback, or explicit resolver method. - Expose common presets as `as const satisfies Record<string, Type>` objects. Keep the accepted type structural so users can provide their own values. - Do not expose ambient facts the caller already knows, such as a `platform` field that only repeats `Platform.OS`, unless the API can return data produced by a different platform than the current runtime. - Do not freeze today's platform support matrix into the type shape. Prefer runtime capability fields such as `availableTextTypes: []` or `supportedFormats: []` over separate platform-specific types or static exclusions. This lets newer native capabilities become available without redesigning the JS API. - Prefer one unified options object. Avoid `ios`/`android` option bags and platform-prefixed methods unless the concepts are genuinely platform-only and cannot be described as a cross-platform capability or no-op. - Classify configuration fields as either requirements or preferences. Throw when a requirement cannot be met. Treat preferences as best-effort when the feature can still perform its core job, such as quality, guidance UI, high-frame-rate tracking, auto-zoom, or a wider scan area. Document best-effort fields and expose capabilities or resolved configuration when callers need to know what was applied. - Split by stable semantic capability when capabilities have different options, results, or futures. For example, barcode scanning and text scanning may deserve separate `BarcodeScanner` and `TextScanner` APIs even when one native API happens to implement both today. A platform that lacks text scanning should fail `createTextScanner()` or report `isTextScannerAvailable: false`, not force text-specific fields to be nullable on a generic scanner used for barcode scanning. - Runtime availability should describe whether a stable capability exists today, not permanently restrict the API shape to the current platform matrix. If Android gains text scanning later, the existing `TextScanner` capability should become available without changing barcode APIs or broadening nullable result types. - Keep capability discovery separate from object contracts. Use capability fields to decide whether a workflow can be created or which optional preferences may apply. Once a factory returns a specific session/resource object, its baseline methods should be guaranteed by construction; otherwise return a narrower type or fail creation instead of making callers check `can*` before every normal method. - Splitting workflows should reduce runtime capability checks. Do not keep broad capability flags only to compensate for one oversized object. For example, a one-shot scanner and an app-owned live scanner can be separate APIs; if the live scanner implementation guarantees zoom, photo capture, or region control, those can be part of the live scanner contract instead of nullable properties plus `can*` checks inherited from a one-shot backend. - Encode lifecycle transitions in the API object graph. If commands are valid only after `configure`, `connect`, `start`, or another lifecycle transition, expose those commands on a handle returned by that transition, not on the parent object with "maybe active" checks. For example, `session.configure(device, outputs)` can return a `Controller` that owns `setZoom(...)` and `focusTo(...)`; the `Session` owns graph configuration and start/stop. - Avoid stale-state APIs. If reconfiguration changes the native resource a command targets, return a new handle and invalidate or dispose the old one. Callers should not be able to accidentally call a command on a parent object that no longer knows which configured device/output it applies to. - Use callbacks or returned resolved objects for post-negotiation facts. If an output, controller, or session config becomes meaningful only after connection, provide `onConfigured`, a returned controller/config, or an explicit `resolve...(...)` method rather than forcing callers to poll nullable properties. - For Nitro-backed imperative APIs, treat the public HybridObject as the API by default. Do not add JS wrappers that pre-parse inputs, translate enum/string shapes, normalize one public format into another, or otherwise make JS call a different API than the generated Nitro spec. Put the intended public shape directly in the Nitro spec and native implementation. - When "all" is a meaningful requested value, model it explicitly instead of using `undefined` as a hidden command. For example, prefer `targetFormats: 'all' | BarcodeFormat[]` or `TargetBarcodeFormat = BarcodeFormat | 'all'` when the implementation benefits from a concrete value. - Do not return half-initialized objects that require a separate `prepare()`, `initialize()`, or `load()` call before normal use. If setup is required, make the factory async and resolve with a ready object. Keep lifecycle methods for real repeatable transitions such as `start()`/`stop()`, not construction readiness. - For larger libraries, expose one small public root or factory that creates stateful domain objects. Keep object construction, async setup, I/O, and validation behind factory methods instead of forcing callers through static functions or half-ready instances. - Return `undefined` only for normal domain absence, and document exactly when it occurs. Do not use optional returns as an unstated error path; throw or reject when an operation fails. - Decide whether returned data is a plain value or a resource. Use plain structs/interfaces for small immutable data whose fields are cheap and semantically complete. Use classes/objects/resources when the value owns native state, needs lazy expensive access, can grow behavior, or should expose methods later. - Choose data representations by semantics first, then performance. Use `string` for decoded text payloads. Use `ArrayBuffer` or byte-oriented objects for raw binary, opaque bytes, media, or large data where zero-copy access is part of the contract.

### Variant Example

Avoid nullable clusters when an object can be in several distinct states:

```typescript interface ScannedData { position: Point text?: string barcode?: string barcodeType?: BarcodeType face?: Rect } ```

Prefer a base type plus variants with nonoptional state-specific fields:

```typescript interface ScannedData { position: Point }

interface ScannedText extends ScannedData { text: string }

interface ScannedBarcode extends ScannedData { barcode: string barcodeType: BarcodeType }

interface ScannedFace extends ScannedData { face: Rect }

type ScannedResult = ScannedText | ScannedBarcode | ScannedFace ```

## Public API Organization

- Split public surfaces into focused files or modules and re-export them from a clear package entry point. Do not create catch-all files that contain a feature's main object plus every enum, option, result, event, and helper type. - Default to one exported public type per file. Group multiple exported types in one file only when they form one tightly coupled log