Domain Driven Design

Domain-Driven Design (DDD)

This skill applies only to projects that have opted in to DDD. Do not apply these patterns to projects that use a different approach.

For hexagonal architecture (ports and adapters), load the hexagonal-architecture skill. DDD and hexagonal architecture are complementary but independent — a project may use one without the other.

Use the structure-codebase skill when designing or changing physical placement. It keeps bounded context, feature/use-case cohesion, package enforcement, and any opted-in hexagonal inside/outside boundary as separate structural axes. If physical restructuring is not requested, preserve the repo's existing layout while keeping domain logic isolated.

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

For authoritative sources, see resources/references.md.

When to Use DDD

DDD adds value for complex domains with rich business rules. Not every project needs it.

Use DDD when:

• Domain has complex business rules and invariants
• Multiple stakeholders with domain expertise
• Business logic is the core differentiator
• Terms have specific, important meanings

Don't use DDD when:

• Simple CRUD with no business rules
• Technical/infrastructure-focused projects
• No domain expert to consult

Start simple and evolve: Begin with ubiquitous language (glossary) and value objects. Add aggregates, domain events, and bounded contexts only when the domain demands it. Your first model will be wrong — that's fine. The goal is to learn quickly and refactor toward deeper insight.

Core Principle

The code must speak the language of the domain. The repository's declared glossary is the naming authority for each bounded context. When naming a declared concept, use its canonical spelling and capitalization in types, functions, variables, and tests. A concept absent from the glossary is not automatically approved or forbidden; it is not yet agreed. Plans, research, comments, conversations, and existing code may reveal candidate vocabulary, but they cannot silently canonize it.

Domain models evolve. The first model is never the final model. As understanding deepens through conversations with domain experts and building working software, the model should change — types get renamed, aggregates get split or merged, new concepts emerge. This is expected and ideal. A model that never changes is either perfect (unlikely) or stagnant (the team stopped learning). TDD and behavioral tests make this evolution safe — rename a concept, update the glossary, and the tests tell you what needs to change.

Where Does This Code Belong?

This is the most common decision in DDD. When unsure, use this framework:

The roles below are logical placement guidance. Use structure-codebase to choose their physical folders and packages.

The purity test is necessary but not sufficient. A pure function that formats a date for display does not belong in domain policy just because it's pure. The question is always: "Is this a business rule?"

// ❌ Pure but NOT domain — formats for human display
export const formatEventDate = (date: string | null) =>
  date ? format(parseISO(date), "MMMM d, yyyy") : undefined;
// → Belongs in presentation code; a hexagonal app may place it at the driving edge

// ✅ Pure AND domain — business rule that affects behavior
// The application supplies current time as data; domain code does not read a clock.
export const isPastEvent = (eventDate: Date | null, now: Date) =>
  eventDate ? eventDate.getTime() < now.getTime() : false;
// → Belongs in domain policy for events

// ✅ Pure AND domain — business calculation
export const calculateCommittedTotal = (items: readonly GiftItem[]) =>
  items.filter(i => i.status !== "idea").reduce((sum, i) => sum + i.pricePence, 0);
// → Belongs in domain policy for budgets

Why placement matters: Domain-policy code typically has strict coverage requirements and zero infrastructure imports. Putting code in the wrong role creates unnecessary testing obligations and architectural violations.

When structure-codebase has been applied, enforce domain and any visible hexagonal boundary mechanically with its package/import rules.

When an architecture uses ports, ownership follows the innermost consumer that needs the abstraction. Repository and gateway ports are therefore normally application-owned because use cases consume them. A port belongs in domain only when the domain model itself genuinely consumes that abstraction. "Inside-owned" does not mean "domain-owned."

Keep domain policy free of frameworks, transport and persistence types, clocks, UUID generators, vendor SDKs, and UI concerns. Parse external input at the edge; pass current time, identifiers, and other external facts into domain functions as typed values. Application use cases coordinate those functions and any required ports.

Ubiquitous Language & Glossary

DDD projects must maintain a repository-declared glossary that defines agreed domain terms. It is the naming authority for its bounded context. The glossary evolves as the model evolves — when the team discovers a better name or splits a concept, decide and record the change before code follows.

The Glossary File

For projects with multiple bounded contexts, organize by context. The same term may have different definitions in different contexts — this is correct, not a bug.

## Gifting Context

| Term | Definition | Examples |
|------|-----------|----------|
| Occasion | A gift-giving event (birthday, holiday) | "Mum's Birthday", "Christmas 2026" |
| Gift Idea | A potential gift for an occasion | "Cookbook", "Scarf" |
| Contribution | Money pledged toward a gift | "£25 from Dad" |

## Notifications Context

| Term | Definition | Examples |
|------|-----------|----------|
| Occasion | An upcoming event that may trigger reminders | (same events, different concern) |
| Recipient | The person being gifted — target of reminder scheduling | "Mum" |

Enforcement Rules

• Use the canonical spelling and capitalization whenever naming a declared concept — including when the request, ticket, or user used a rejected alias. The request's wording is input to translate, not a naming authority: a request that says "reservation" where the glossary declares Booking and rejects Reservation is Booking in the code, Reservation never appears as a type, field, or function name, and the reply says which words were translated.
• Treat an absent term as a candidate requiring agreement, not as automatically approved or forbidden.
• Rejected or deprecated aliases must point to the owning bounded context and canonical replacement.
• Apply glossary checks to type and interface names, function names, and test descriptions within the check's declared scope.
• Treat a green language-lint command as evidence for its targeted ratchet only, never as proof that every prose passage and identifier uses the ubiquitous language correctly.

// ✅ Uses domain language
type GiftIdea = {
  readonly id: GiftIdeaId;
  readonly description: string;
  readonly occasion: OccasionId;
  readonly estimatedCost: Money;
  readonly status: 'proposed' | 'selected' | 'purchased';
};

// ❌ Technical jargon
type Item = { readonly id: string; readonly text: string; readonly parentId: string; };

Building Blocks

Value Objects

Immutable, identity-less, compared by their attributes (not by reference). Represent domain concepts defined by their attributes. Two Money values with the same minor units and currency are equal — value objects have no identity.

type Currency = 'GBP' | 'USD' | 'EUR';
type Money = { readonly minorUnits: number; readonly currency: Currency };

const createMoney = (minorUnits: number, currency: Currency): Money => {
  if (!Number.isSafeInteger(minorUnits)) throw new Error('Money minor units must be a safe integer');
  if (minorUnits < 0) throw new Error('Money cannot be negative');
  return { minorUnits, currency };
};
// Factory throws = invariant violation (a bug in calling code).
// Schemas catch invalid user input at trust boundaries BEFORE
// the factory is called. If the factory throws, something
// bypassed the schema.

Store and calculate money in integer minor units, never binary floating-point major units. A boundary that accepts decimal text must use the currency's exponent and an explicit policy for excess precision (for example, reject it or round half-even); never convert with Math.round(rawNumber * 100).

For value objects crossing trust boundaries (API input, form data), use Zod schemas. For domain-internal value objects, plain types + factory functions suffice. See the typescript-strict skill for schema-first patterns.

Zod-to-branded-type bridging — parse raw input into branded domain types at trust boundaries:

// Schema at trust boundary — parses raw strings into branded types
const PledgeInputSchema = z.object({
  occasionId: z.string().min(1).transform(createOccasionId),
  contributorId: z.string().min(1).transform(createContributorId),
  amount: z.object({ minorUnits: z.number().int().safe().positive(), currency: CurrencySchema }),
});

// Reconstitution from persistence — same pattern, used at the integration boundary
// (a driven adapter in hexagonal architecture loads the gift ideas alongside the row)
const toOccasion = (row: OccasionRow, giftIdeas: ReadonlyArray<GiftIdea>): Occasion => ({
  id: createOccasionId(row.id),
  name: row.name,
  giftIdeas,
  budget: createMoney(row.budgetMinorUnits, parseCurrency(row.budgetCurrency)),
  totalPledged: createMoney(row.pledgedMinorUnits, parseCurrency(row.budgetCurrency)),
  isFundingClosed: row.isFundingClosed,
});

Reconstitution (rebuilding domain objects from DB rows) uses the same factory functions as creation. The factory validates, so invalid persisted data is caught on read rather than silently corrupting the domain.

Branded Types

The branded type pattern itself is covered by the typescript-strict skill — load it for the general rules. The DDD-specific application:

• Give every entity its own branded ID (OccasionId, GiftIdeaId) so the compiler rejects cross-entity ID mixups — passing a GiftIdeaId where an OccasionId is expected is a compile error, not a runtime bug.
• Brand once, through one validating factory per branded type — raw strings become branded values only after validation. The single as inside that factory is the only assertion the brand gets: validate, then brand. Budget: exactly one as per branded type, zero elsewhere. A value derived from a branded value — a date computed from another date, an id read back from a row — is re-branded by calling the factory again, never by writing another as.
• Use branded types for entity IDs and single-value value objects (EmailAddress).

type OccasionId = string & { readonly __brand: 'OccasionId' };
type GiftIdeaId = string & { readonly __brand: 'GiftIdeaId' };

const createOccasionId = (raw: string): OccasionId => {
  if (!raw.trim()) throw new Error('OccasionId cannot be empty');
  return raw as OccasionId; // justified: factory validates, then brands
};

Entities

Have identity and a lifecycle. Always valid after construction or state transition.

type Occasion = {
  readonly id: OccasionId;
  readonly name: string;
  readonly giftIdeas: ReadonlyArray<GiftIdea>;
  readonly budget: Money;
  readonly totalPledged: Money;
  readonly isFundingClosed: boolean;
};

// Immutable update — returns new valid state
const renameOccasion = (occasion: Occasion, newName: string): Occasion => ({
  ...occasion,
  name: newName,
});

Always-valid principle: An entity must satisfy its invariants at all times. Validate on construction (factory functions or schema parsing) and on every state transition. Never allow an entity to exist in an invalid state, even temporarily.

Make Illegal States Unrepresentable

General type-safety patterns (contained any only at unavoidable interop boundaries, discriminated unions, schema-first validation) live in the typescript-strict skill. The DDD-specific application: model entity lifecycles as discriminated unions where each variant carries only the data valid for that state — never boolean flags plus optional fields, which allow contradictory combinations like { isVerified: true, verifiedAt: undefined }:

type Order =
  | { readonly status: 'draft'; readonly items: ReadonlyArray<OrderItem> }
  | { readonly status: 'placed'; readonly items: ReadonlyArray<OrderItem>; readonly placedAt: Date }
  | { readonly status: 'shipped'; readonly items: ReadonlyArray<OrderItem>; readonly placedAt: Date; readonly shippedAt: Date; readonly trackingNumber: string };

Always handle all lifecycle variants exhaustively. The never type ensures the compiler catches unhandled states when you add a new variant:

const describeOrder = (order: Order): string => {
  switch (order.status) {
    case 'draft': return `Draft with ${order.items.length} items`;
    case 'placed': return `Placed at ${order.placedAt.toISOString()}`;
    case 'shipped': return `Shipped: ${order.trackingNumber}`;
    default: { const _exhaustive: never = order; return _exhaustive; }
  }
};

Aggregates

Clusters of entities and value objects with a single root. All modifications go through the root.

1. One aggregate root per transaction
2. Reference other aggregates by ID — never embed
3. All invariants enforced by the root
4. Keep aggregates small — only what's needed for consistency

For detailed aggregate design guidance, see resources/aggregate-design.md.

Specifications (Predicate Functions)

Complex business rules for filtering, eligibility, or validation are expressed as predicate functions in the domain layer. Evans calls these "specifications."

// Specification: "can this eligible contributor pledge to this occasion?"
const canPledge = (occasion: Occasion, eligibility: ContributorEligibility, amount: Money): boolean => {
  const values = [amount.minorUnits, occasion.totalPledged.minorUnits, occasion.budget.minorUnits];
  if (!values.every(Number.isSafeInteger)) return false;
  if (occasion.totalPledged.minorUnits > occasion.budget.minorUnits) return false;
  return eligibility.mayPledge &&
    !occasion.isFundingClosed &&
    amount.minorUnits > 0 &&
    amount.currency === occasion.totalPledged.currency &&
    occasion.totalPledged.currency === occasion.budget.currency &&
    amount.minorUnits <= occasion.budget.minorUnits - occasion.totalPledged.minorUnits;
};

// Compose specifications for complex eligibility
const isGiftReady = (occasion: Occasion): boolean =>
  occasion.totalPledged.minorUnits >= occasion.budget.minorUnits &&
  occasion.giftIdeas.some(idea => idea.status === 'selected');

Specifications are pure predicate functions — they return boolean and have no side effects. Use them in domain services, use cases, and query filters. Name them with is, can, or has prefixes.

Domain Events

Domain events represent something meaningful that happened in the domain ("OrderPlaced", "ContributionPledged"). They coordinate side effects across aggregates and bounded contexts.

Domain events earn their complexity when:

• Side effects cross aggregate boundaries
• Other bounded contexts need to react to changes
• You need an audit trail or event-driven workflows

Don't add domain events when:

• All logic is within a single aggregate
• Side effects are within the same transaction
• Explicit return values from domain functions suffice

For most projects, start without domain events and add them when the domain demands coordination. See resources/domain-events.md for the Decider pattern and detailed guidance. When events become the source of truth — persisted to an append-only log and replayed to rebuild state — that is event sourcing; load the event-sourcing skill (it builds directly on this Decider). Where domain events already exist, observability rides for free as another subscriber — an observability adapter on the existing publisher port, not a second announcement channel (see the hexagonal-architecture skill's four-tier observability model and the observability skill).

Domain Services

When business logic doesn't belong to a single entity, it belongs in a domain service — a stateless function in the domain layer that combines domain values or read-only policy facts. A domain service does not make multi-aggregate writes atomic: redraw the aggregate when one invariant requires synchronous change, or coordinate one aggregate write plus a durable event in the application layer.

// ❌ WRONG — cramming an external eligibility policy into one entity
const addContribution = (occasion: Occasion, contribution: Contribution): Occasion => {
  // This cannot know whether the contributor is currently eligible.
};

// ✅ CORRECT — pure domain service consumes a read-only policy fact
const pledgeContribution = (
  occasion: Occasion,
  eligibility: ContributorEligibility,
  pledge: { readonly id: PledgeId; readonly amount: Money },
): PledgeDecision => {
  if (!eligibility.mayPledge) {
    return { success: false, reason: 'contributor-ineligible' };
  }
  if (occasion.isFundingClosed) {
    return { success: false, reason: 'funding-closed' };
  }
  if (
    pledge.amount.currency !== occasion.totalPledged.currency ||
    occasion.totalPledged.currency !== occasion.budget.currency
  ) {
    return { success: false, reason: 'currency-mismatch' };
  }
  const values = [
    occasion.totalPledged.minorUnits,
    occasion.budget.minorUnits,
    pledge.amount.minorUnits,
  ];
  if (!values.every(Number.isSafeInteger)) throw new Error('Invalid Money invariant');
  if (pledge.amount.minorUnits <= 0) {
    return { success: false, reason: 'non-positive-amount' };
  }
  if (occasion.totalPledged.minorUnits > occasion.budget.minorUnits) {
    throw new Error('Invalid Occasion funding invariant');
  }
  if (pledge.amount.minorUnits > occasion.budget.minorUnits - occasion.totalPledged.minorUnits) {
    return { success: false, reason: 'exceeds-budget' };
  }
  const totalPledged = occasion.totalPledged.minorUnits + pledge.amount.minorUnits;
  if (!Number.isSafeInteger(totalPledged)) throw new Error('Money addition overflowed');
  return {
    success: true,
    occasion: { ...occasion, totalPledged: createMoney(totalPledged, pledge.amount.currency) },
    events: [{
      type: 'PledgeRecorded',
      id: pledge.id,
      occasionId: occasion.id,
      contributorId: eligibility.contributorId,
      amount: pledge.amount,
    }],
  };
};

Domain service vs use case (application service):

Keep domain business decisions pure by default. A domain-owned driven port is a rare, explicit exception only when the model itself—not application orchestration—owns the conversation. The consuming domain service is then effectful but still provider-free: isolate its pure decision logic and test the port interaction with a fake. Do not use this exception for repositories or gateways needed only by application orchestration.

Physical placement depends on the selected architecture. Organize by bounded context or business capability before technical role. In a visible hexagonal backend both roles are inside, commonly in separate domain/application packages when that dependency boundary earns its cost, or in cohesive feature/use-case modules inside one capability package. DDD without hexagonal architecture may use another context-first arrangement. Follow structure-codebase rather than assuming domain/ contains all inside code.

For detailed guidance, see resources/domain-services.md.

Error Modeling

Use discriminated union result types for expected business outcomes. Reserve exceptions for programmer mistakes and infrastructure failures.

type PledgeDecision =
  | { readonly success: true; readonly occasion: Occasion; readonly events: readonly PledgeRecorded[] }
  | { readonly success: false; readonly reason: 'contributor-ineligible' | 'non-positive-amount' | 'currency-mismatch' | 'exceeds-budget' | 'funding-closed' };

type PledgeResult =
  | PledgeDecision
  | { readonly success: false; readonly reason: 'not-found' | 'concurrent-change' }; // application outcomes

The test: Could a user's action legitimately cause this outcome? If yes → result type. If no (it would mean a bug) → exception.

For detailed error modeling patterns and how errors propagate through layers, see resources/error-modeling.md.

Repository Pattern

Repositories provide collection-like access to aggregates. Put the interface with the innermost policy that consumes it and the concrete implementation with infrastructure or integration code. In an explicitly hexagonal system, repository ports are normally application-owned because use cases load and save aggregates; place them beside that application policy and let a driven adapter implement them. Put one in domain only when the domain model itself genuinely owns and consumes the abstraction. DDD without ports and adapters does not gain adapter layers by implication; follow the architecture selected by structure-codebase. Repositories use interface (not type) because they define behavior contracts, and their methods use domain language.

// Application-owned repository contract — an inside-owned port when hexagonal architecture is used
interface OccasionRepository {
  readonly findById: (id: OccasionId) => Promise<Occasion | undefined>;
  readonly save: (occasion: Occasion) => Promise<void>;
}

// Concrete implementation belongs with infrastructure/integration;
// in hexagonal architecture it is a driven adapter.

Repositories handle writes and single-aggregate reads. For reads that need to JOIN across aggregates (dashboard views, detail pages combining data from multiple entities), repositories are the wrong tool — they enforce aggregate boundaries that reads need to cross. Use query functions that JOIN freely instead. This is the CQRS-lite pattern: writes go through repositories (consistency), reads go through query functions (flexibility). See the hexagonal-architecture skill's CQRS-lite section and its resources/cqrs-lite.md for details.

For simple domains where reads map cleanly to a single aggregate, repository reads are fine. Don't separate prematurely.

DDD + TDD Integration

Test by Domain Concept, Not Implementation File

Follow the physical shape selected for the project and keep focused tests beside the behavior they describe; DDD does not require a root tests/ directory.

Title describe and it blocks with the business behaviour in glossary terms, never with the name of the function under test:

// ❌ Titled after helpers — proves nothing a domain expert would recognise
describe('addDays', () => { it('rolls over a month boundary', ...) });
describe('createMoney', () => { it('rejects a negative amount', ...) });

// ✅ Titled after the rule
describe('Pledging to an Occasion', () => { it('is refused once funding is closed', ...) });
describe('Committed total for an Occasion', () => { it('excludes gift ideas that are only proposed', ...) });

A date or arithmetic helper gets no describe of its own; it is proved through the rule that uses it. One test file per implementation file is the smell that produces helper-named tests — group tests by the concept they exercise instead.

<selected-context-or-feature>/
  occasions/
    create-occasion.test.ts       # Behavior: creating occasions
    add-gift-idea.test.ts         # Behavior: managing gift ideas
    occasion-budget.test.ts       # Behavior: budget constraints

Select the Primary Behavioral Boundary

Test through the broadest stable public behavior that the chosen architecture actually has:

• For a domain-only model or library, call aggregate operations, domain services, or Deciders directly; these tests are primary, not a fallback.
• When application policy exists, call the use case with simple fakes or stubs for its collaboration contracts. This exercises domain behavior and orchestration together without inventing ports.
• In an explicitly hexagonal system, exercise the driving port with driven actors replaced by outside test interactors; load hexagonal-architecture for that testing strategy.
• Add integration and delivery tests according to real infrastructure and user risk, not because DDD mandates a layer pyramid.

Focused tests for complex pure business rules complement application tests when both exist. They may be the complete primary suite for a provider-free domain package.

See resources/testing-by-layer.md for the architecture-neutral decision. For a DDD system that also adopts hexagonal architecture, the hexagonal skill's resources/worked-example.md traces one feature across ports, adapters, and tests.

Test Factories Use Domain Language

const getTestOccasion = (overrides?: Partial<Occasion>): Occasion =>
  OccasionSchema.parse({
    id: createOccasionId('occasion-1'),
    name: "Mum's Birthday",
    giftIdeas: [],
    budget: createMoney(10_000, 'GBP'),
    totalPledged: createMoney(0, 'GBP'),
    isFundingClosed: false,
    ...overrides,
  });

Bounded Contexts

A bounded context is a linguistic boundary within which a particular domain model and glossary apply. The same word (e.g., "User") can mean different things in different contexts — and that's correct.

1. Each context owns its own model and glossary — User in billing differs from User in shipping
2. Communicate between contexts via events or explicit contracts — never share mutable state
3. Anti-Corruption Layer (ACL) — when integrating with external systems or other contexts whose model doesn't fit yours, translate at the boundary rather than letting their types leak in
4. Shared kernel should be minimal — only truly universal value objects (Money, Email). If the shared kernel grows, boundaries are unclear
5. Each context has its own glossary section

For context mapping patterns, monorepo structure, and ACL examples, see resources/bounded-contexts.md.

Anti-Patterns

Anemic Domain Model

Entities are data bags with no behavior. All logic in "services." Fix: put behavior as pure functions next to the types they operate on.

Generic Technical Names

Using Item, Entity, Record, Data, Info instead of domain terms. Always use the glossary.

Presentation Logic in Domain

Display formatting does not belong in domain policy. The test: "make this look right for a human" = presentation. "Enforce a business rule" = domain. Purity is not sufficient — a pure formatting function is still presentation.

Leaking Domain Logic

Business logic in route handlers, database queries, or adapters. Keep it in domain-policy modules under the physical shape selected for the project.

Relationship-Driven Aggregates

Designing aggregates by mapping entity hierarchies ("Order contains OrderLines, OrderLine contains Product") instead of discovering invariants. The tell: aggregate methods only add, remove, or attach children — no business rules are enforced. If an aggregate's only job is managing parent-child associations, those relationships likely belong as database constraints or simple data models, not an aggregate hierarchy.

The fix: Ask "What must remain true when state changes?" If the answer is only structural (one-to-many, one-to-one), you don't need an aggregate — you need a foreign key. Aggregates earn their complexity when they enforce behavioral invariants during commands.

Over-Engineering

Not every project needs aggregates, domain events, or bounded contexts. Start with:

1. Ubiquitous language (glossary)
2. Value objects and entities
3. Add complexity only when the domain demands it

Resisting Model Evolution

Treating the initial model as sacred — refusing to rename types, split aggregates, or restructure bounded contexts as understanding deepens. The model should evolve continuously. If a refactoring reveals that "Occasion" should really be "GiftEvent" and "SavingsGoal", do it. The glossary changes, the types change, the tests guide the migration. Evans calls these "breakthroughs" — moments where the model fundamentally improves because the team learned something new about the domain.

Checklist

• Glossary file exists and is up to date
• Declared concepts use the glossary's canonical spelling and capitalization
• Missing terms are treated as candidates requiring agreement, not silently adopted
• Rejected aliases name the owning bounded context and canonical replacement
• Any green language-lint result is reported with its actual targeted scope
• Value objects are immutable and identity-less
• Entities are always valid (invariants enforced on construction and transitions)
• Entities have branded IDs; primitive value objects use branded types
• Aggregate roots enforce all invariants
• Aggregate boundaries justified by invariants, not entity relationships
• Aggregates contain no read-only properties that exist solely for query convenience
• Other aggregates referenced by ID, not embedded
• Cross-aggregate logic in domain services, not crammed into one entity
• Use cases and orchestration live in application policy, not domain policy
• Repository and gateway ports live with the innermost consumer—normally application policy; domain owns only ports it genuinely consumes
• Concrete driven adapters implement inside-owned ports and live with infrastructure/integration
• Discriminated unions have exhaustive switch handling
• Expected business outcomes use result types, not exceptions
• Domain logic has zero infrastructure dependencies
• If structure-codebase has been applied, its domain/context and any visible inside/outside rules are present and passing
• Presentation logic is NOT in domain policy (even if pure)
• Tests organized by domain concept, not implementation file
• Each logical role has behavioral tests at the appropriate level