hexagonal-architecture

name: hexagonal-architecture description: Hexagonal (ports and adapters) architecture patterns for TypeScript. Use when implementing ports, adapters, dependency inversion, or domain isolation. Only applies to projects that explicitly use hexagonal architecture. Do NOT use for projects without ports/adapters structure.

Hexagonal Architecture (Ports & Adapters)

This skill applies only to projects that have opted in to hexagonal architecture. Do not apply these patterns to projects that use a different architecture. For introducing hex arch into an existing codebase incrementally, see resources/incremental-adoption.md.

For domain modeling (entities, value objects, aggregates, ubiquitous language), load the domain-driven-design skill. Hex arch and DDD are complementary but independent — hex arch provides structural isolation (how the outside connects), DDD provides the domain model (what lives in the center). A project may use one without the other.

If the folder-structure skill has been explicitly applied to this project, follow its protected-domain-core layout and lint/import-boundary rules for DDD/hex contexts. Do not introduce those physical folder or lint rules into projects that have not applied folder-structure; in that case, preserve the repo's existing structure while enforcing the dependency direction described here.

Deep-dive resources are in the resources/ directory. Load them on demand:

For authoritative sources and naming rationale, see resources/references.md.

Core Concept

Business logic lives in the center. External systems connect through ports (interfaces) and adapters (implementations). Dependencies point inward — the domain never knows about the outside world.

         Driving (left)                    Driven (right)
    ┌──────────────────┐            ┌──────────────────┐
    │  Route handlers  │            │  Repositories    │
    │  CLI commands    │──────┐┌────│  API clients     │
    │  Event listeners │      ││    │  Email services  │
    └──────────────────┘      ││    └──────────────────┘
         call into ──────►  ┌────┐  ◄────── implement
                            │    │
                            │ DO │
                            │ MA │
                            │ IN │
                            │    │
         call into ──────►  └────┘  ◄────── implement
    ┌──────────────────┐      ││    ┌──────────────────┐
    │  Cron triggers   │──────┘└────│  File storage    │
    │  Message queues  │            │  Payment gateway │
    └──────────────────┘            └──────────────────┘

Driving adapters (left): initiate actions on the application. They call use cases. Driven adapters (right): the application reaches out to them. They implement port interfaces.

This asymmetry is fundamental. Driving adapters depend on driving port interfaces exposed by the application. Driven adapters implement repository/gateway interfaces defined by the domain.

Ports = Public Contracts

A port is a named, purposeful conversation at the application boundary. In TypeScript, represent ports as explicit interface types when the boundary is public or architectural: driving adapters call driving port interfaces, and driven adapters/fakes implement driven port interfaces.

Use case implementations satisfy the driving port. The driving adapter should depend on the port's role-shaped interface, not a framework-specific controller, handler class, or concrete adapter.

Naming Ports and Public Interfaces

Name every port from the application's point of view, using the domain language of the conversation:

Public interface naming rules:

• Use interface for behavior contracts, not data shapes that are better modeled as type/schemas.
• Do not prefix interfaces with I or suffix them with Interface; name the role (PaymentGateway, not IPaymentGateway or PaymentGatewayInterface).
• Avoid Port in type names. A name like PaymentPort says "architecture" instead of "conversation."
• Avoid Impl in implementation names. Name the concrete technology or strategy: createStripePaymentGateway, createFakePaymentGateway, createDrizzleOrderRepository.
• Prefer role interfaces over header interfaces: include only the methods the use case needs, not every method an adapter happens to expose.
• Make names stable under adapter swaps. If moving from Stripe to PayPal, SQL to DynamoDB, or HTTP to a queue forces a port rename, the port name leaked infrastructure.

// Driving port — exposed by the application, called by driving adapters
interface ForPlacingOrders {
  readonly placeOrder: (command: PlaceOrderCommand) => Promise<PlaceOrderResult>;
}

// Driven port — defined in domain, implemented by adapters
interface UserRepository {
  readonly findById: (id: UserId) => Promise<User | undefined>;
  readonly save: (user: User) => Promise<void>;
}

// Driven port — defined in domain, implemented by adapters
interface PaymentGateway {
  readonly charge: (amount: Money, paymentInfo: PaymentInfo) => Promise<ChargeResult>;
}

// Driven port — event publishing (outbound to message brokers)
interface OrderEventPublisher {
  readonly publish: (event: OrderEvent) => Promise<void>;
}

Port design principles:

• Name ports by business purpose, not technology (UserRepository, not DatabasePort)
• Keep ports focused — one per aggregate or capability, not one god port
• Port methods use domain types, never infrastructure types (no SqlRow, no HttpResponse)
• Creation param schemas co-locate with the repository port they describe

Adapters = Implementations

Adapters implement ports for specific technologies. A good adapter is simple — it translates between the port's domain types and the technology's native types. No business logic.

// Driven adapter — implements the repository port using Drizzle/D1
const createDrizzleUserRepository = (db: D1Database): UserRepository => ({
  findById: async (id) => {
    const row = await db.select().from(users).where(eq(users.id, id)).get();
    return row ? toUser(row) : undefined;
  },
  save: async (user) => {
    await db.insert(users).values(toRow(user)).onConflictDoUpdate({ ... });
  },
});

// Driven adapter — implements the same port for tests
const createFakeUserRepository = (initial: readonly User[] = []): UserRepository => {
  const store = new Map(initial.map(u => [u.id, u]));
  return {
    findById: async (id) => store.get(id),
    save: async (user) => { store.set(user.id, user); },
  };
};

Adapter error handling: Infrastructure errors (connection lost, timeout, constraint violation) should either propagate as exceptions to a top-level handler or be translated into domain-appropriate results at the adapter boundary. The domain never catches infrastructure errors — it doesn't know infrastructure exists.

// Driven adapter: translate expected infrastructure errors into domain-specific errors
const createDrizzleUserRepository = (db: Database): UserRepository => ({
  save: async (user) => {
    try {
      await db.insert(users).values(toRow(user)).onConflictDoUpdate({ ... });
    } catch (e) {
      if (isUniqueConstraintError(e)) throw new UserAlreadyExistsError(user.id);
      throw e; // unexpected errors (connection lost, disk full) propagate
    }
  },
});

Expected constraint violations become domain-specific errors (caught by the use case or driving adapter). Unexpected infrastructure errors propagate to the top-level handler.

Key principle: If swapping an adapter requires changing domain code, the boundary is wrong.

Reads vs Writes (CQRS-lite)

Not all reads need to go through repositories. The repository pattern enforces aggregate boundaries — essential for writes, but reads often need to JOIN across aggregates for display.

Query functions are driven adapters too — they live in the adapter layer (e.g., db/queries/) and return read-optimized DTOs. They bypass the repository pattern intentionally.

// Query function — JOINs across aggregates for display
// Lives in db/queries/, NOT in domain/
const getParticipantEventView = async (db: Database, eventId: string) => {
  return db.select({ ... })
    .from(events)
    .innerJoin(occasions, ...)
    .leftJoin(giftClaims, ...)
    .where(eq(events.id, eventId))
    .all();
};

Domain-layer pure functions can transform query results into display types — these encode business rules about what data means (e.g., "is this item claimed by the current user?"). The query fetches; the domain function interprets.

For detailed CQRS-lite guidance, see resources/cqrs-lite.md.

Dependency Injection

Inject all dependencies via function parameters. No DI container needed. The driving adapter gathers impure dependencies, passes them to the use case, and acts on the result — Seemann's "impureim sandwich" (impure/pure/impure).

// WRONG — creates dependencies internally (untestable, tightly coupled)
const createOrder = async (order: NewOrder) => {
  const repo = new DrizzleOrderRepo(getDb());          // hardcoded
  const gateway = new StripeGateway(process.env.KEY);   // hardcoded
  // ...
};

// RIGHT — dependencies injected into a use case implementation
interface ForPlacingOrders {
  readonly placeOrder: (order: NewOrder) => Promise<OrderResult>;
}

const createOrderPlacement = (
  repo: OrderRepository,
  gateway: PaymentGateway,
): ForPlacingOrders => ({
  placeOrder: async (order) => {
    const charge = await gateway.charge(order.total, order.payment);
    if (!charge.success) return { success: false, reason: charge.error };
    const saved = await repo.save({ ...order, chargeId: charge.id });
    return { success: true, order: saved };
  },
});

Composition root: Wiring happens at the application entry point — where adapters are created from environment/config and injected into use cases. This is the only place that knows about concrete implementations.

// Route handler = composition root + driving adapter
export async function POST(request: Request) {
  const { env } = getCloudflareContext();
  const db = createDb(env.DB);

  // Wire adapters
  const repo = createDrizzleOrderRepository(db);
  const gateway = createStripeGateway(env.STRIPE_KEY);
  const orderPlacement: ForPlacingOrders = createOrderPlacement(repo, gateway);

  // Call use case
  const body = CreateOrderSchema.parse(await request.json());
  const result = await orderPlacement.placeOrder(body);
  return NextResponse.json(result);
}

The route handler is thin glue: parse input → wire adapters → call use case → return response. No business logic.

Non-HTTP driving adapters follow the same pattern — parse, wire, delegate:

// Queue consumer = driving adapter (same structure as route handler)
const handlePledgeMessage = async (message: SQSMessage, env: Env) => {
  const db = createDb(env.DB);
  const occasionRepo = createDrizzleOccasionRepository(db);
  const contributorRepo = createDrizzleContributorRepository(db);
  const pledging: ForPledgingToOccasions = createPledgingToOccasions(occasionRepo, contributorRepo);

  const dto = PledgeSchema.parse(JSON.parse(message.body));
  await pledging.pledgeToOccasion(dto);
};

The use case doesn't know or care whether it was triggered by an HTTP request, a queue message, a cron job, or a CLI command. Every driving adapter is thin glue.

Naming: Use case interfaces and implementations are named after the business capability — ForPlacingOrders, createOrderPlacement, pledgeToOccasion. Never CreateOrderUseCase or PlaceOrderHandler. Pattern suffixes are technical jargon, not domain language. You can tell a use case implementation from a domain function by its dependencies — use case implementations take driven ports (repositories, gateways) as parameters; domain functions take only domain types.

File Organization

The locations below describe the logical layers — physical layout is governed by the folder-structure note at the top of this skill. Use cases live in domain/ by default; when the folder-structure skill's layout is applied, they live in a sibling use-cases/ folder instead — required there, because its lint rules forbid domain/ from importing use-cases/.

Key rules:

• Domain has zero external dependencies (no framework, database, or HTTP imports)
• Driven port interfaces live in domain alongside the entity or capability they serve
• Schemas co-locate with their entity in domain
• Adapters import from domain, never the reverse
• Route handlers are thin — parse, wire, delegate, respond
• When folder-structure has been applied, enforce these rules with its recommended import-boundary lint configuration.

Testing Strategy

Hex arch's primary benefit is testability. The primary test boundary is the use case — call it with driven ports replaced by in-memory fakes (not mocks). This proves the feature works as a whole.

Fakes over mocks: Fakes implement the real interface and maintain state. Mocks verify call sequences and break on refactoring. See resources/testing-hex-arch.md for detailed patterns.

For a complete worked example showing one feature traced through every layer (glossary → types → domain → use case → adapters → tests → file locations), see resources/worked-example.md.

Cross-Cutting Concerns

The domain never imports a logger, catches HTTP errors, or manages transactions. It returns results; adapters handle the rest. See resources/cross-cutting-concerns.md for detailed patterns.

Anti-Patterns

Domain Depending on Infrastructure

The most common hex arch violation. Domain code imports from frameworks, databases, or external services.

// ❌ Domain imports Drizzle
import { eq } from 'drizzle-orm';
export const findActiveUsers = async (db) => db.select()...

// ✅ Domain defines the contract; adapter implements it
interface UserRepository {
  readonly findActive: () => Promise<readonly User[]>;
}

Business Logic in Adapters

Route handlers or repositories contain business rules instead of delegating to domain.

// ❌ Business rule in route handler
export async function POST(request: Request) {
  const order = await orderRepo.findById(id);
  if (order.total > 1000) { await requireManagerApproval(order); } // business rule!
  ...
}

// ✅ Business rule in domain
const placeOrder = (order: Order): PlaceOrderResult => {
  if (order.total > 1000) return { success: false, reason: 'requires-approval' };
  ...
};

Bypass Adapters

Route handler accesses the database directly instead of going through a port.

// ❌ Route handler hits DB directly
export async function GET(request: Request) {
  const users = await db.select().from(users).where(eq(users.active, true));
  ...
}

// ✅ Route handler calls use case, which uses a port
const result = await getActiveUsers(userRepo);

Port Proliferation

Creating a port for every tiny abstraction. Ports should represent meaningful boundaries — one per aggregate (repositories) or per external capability (payment, email, auth).

Technology-Shaped Ports

Port methods that expose technology details. Port methods should use domain language.

// ❌ Technology leaks into port
interface UserRepository {
  readonly findBySqlQuery: (sql: string) => Promise<User[]>;
  readonly getFromRedisCache: (key: string) => Promise<User>;
}

// ✅ Business language
interface UserRepository {
  readonly findActive: () => Promise<readonly User[]>;
  readonly findById: (id: UserId) => Promise<User | undefined>;
}

Pattern-Shaped Public Names

Names that advertise the architecture pattern instead of the business role make code harder to read and easier to cargo-cult.

// ❌ Names the pattern or type mechanism
interface IPaymentPort {
  readonly charge: (amount: Money, paymentInfo: PaymentInfo) => Promise<ChargeResult>;
}
const createPaymentGatewayImpl = (): IPaymentPort => ...
const placeOrderUseCase = async (...) => ...

// ✅ Names the role and the concrete adapter
interface PaymentGateway {
  readonly charge: (amount: Money, paymentInfo: PaymentInfo) => Promise<ChargeResult>;
}
const createStripePaymentGateway = (): PaymentGateway => ...
const placeOrder = async (...) => ...

Checklist

• Domain logic has zero framework/infrastructure dependencies
• If folder-structure has been applied, protected-domain-core lint/import rules are present and passing
• All external boundaries use ports/public contracts
• Driving adapters (routes) are thin — parse, wire, delegate, respond
• Driven adapters (repos) implement ports, contain no business logic
• Dependencies injected via parameters, never created internally
• Driven port interfaces live in domain, named by business purpose
• Public interfaces avoid I prefixes, Interface suffixes, Port suffixes, and Impl implementations
• Driving port and use case names use business capability language, not UseCase/Handler pattern names
• Schemas defined in domain, not duplicated in adapters
• Reads that JOIN across aggregates use query functions (CQRS-lite)
• Each layer has behavioral tests at the appropriate level
• Swapping any adapter requires zero domain code changes
• Cross-cutting concerns (auth, logging, transactions) live in adapters, not domain
• Domain returns result types for expected outcomes, never throws for business rules