hexagonal-architecture

Hexagonal Architecture + DDD

When to Use

• You are adding or refactoring features and need to place code in the correct layer (domain/service/port/adapter/controller).
• You are defining a new port or adapter and want to ensure dependency direction is correct.
• You are creating or modifying adapters, services, or controllers and need layer-specific conventions.
• You are reviewing imports/structure to enforce architectural boundaries.

Quick Reference

Controllers (Thin Layer)

Controllers are the entry points that connect external inputs to the application. Examples: a FastAPI app, a Click CLI, a Chainlit chat app. Keep them thin:

• Parse/validate inputs
• Call shared services
• Map errors to appropriate responses (e.g., HTTP status codes, CLI exit codes, UI messages)
• Serialize outputs via model_dump(mode="json") or equivalent
• No business rules in controllers

For FastAPI-specific conventions, see the fastapi skill. For Click-specific conventions, see the cli-conventions skill.

Domain Models

• Put business rules/invariants in domain models (shared/src/shared/domain/**)
• Raise ValueError with clear messages for domain violations
• Pydantic v2 is the sole allowed third-party dependency in domain models

For DDD building blocks (entities, value objects, aggregates), file organization, and the persistence contract, see the domain-models skill.

Services (Orchestration)

Services are use-case orchestrators in shared/src/shared/services/**. For system health checks (technical, non-business concerns):

• Put domain models in shared/src/shared/domain/platform_ops/health/
• Services should aggregate adapter results into a generic SystemHealth (no per-adapter probe logic in services)

Inputs

• Fixed internal knobs (model IDs, temperatures, limits) live in a nested:
  • @dataclass(frozen=True, slots=True) class Hyperparams: ...
  • with _HP: Final[Hyperparams] = Hyperparams()
• Controller-controlled parameters should be Pydantic v2 models named *Params (typically in shared/src/shared/domain/**) and passed as params=...
• Direct inputs (what to act on) are primitive args alongside params (e.g., source_url: str, article_id: UUID)

Outputs

• Return domain models (Pydantic); let controllers serialize via model_dump(mode="json")
• If a workflow needs a composite result, prefer a dedicated domain response model over a raw dict

Validation

• Validate service-level inputs early (bounds, URL format) using shared validators (e.g., validate_url)
• Domain invariants live in domain models and raise ValueError
• Do not duplicate the same validation across service + domain layers

Resource Handling

• If a service receives closeable dependencies, the wiring layer (controller-side DI / setup) must own lifecycle and close them in finally

Ports (Interfaces)

• Define boundaries as ports (shared/src/shared/ports/**) using typing.Protocol
• Minimal method surface; no side effects; no external library imports; return domain types
• Ports MUST be technology-agnostic: folder names describe business capabilities (e.g., ports/database/, ports/llm/), NOT technologies
• Repository naming conventions:
  • Single item: get_{model}(...) -> Model | None (missing → None)
  • Lists: list_{model}s(..., limit: int | None = None, offset: int = 0) -> list[Model] (missing → [])
  • Prefer one list_* with multiple optional filters over specialized *_by_* methods

Adapters (Implementations)

Adapters live in shared/src/shared/adapters/** and are technology-specific (e.g., adapters/database/postgres/, adapters/llm/). For adapters integrating external components (DB/LLM/HTTP/etc.), expose:

• async def health_check(self) -> ComponentHealth
• Use generic health VO models (ComponentHealth, SystemHealth) and put component diagnostics in details (e.g. db-version, base-url)

Validation

• Validate adapter inputs early and fail fast with ValueError (non-empty strings, bounds, required IDs)
• Avoid magic numbers: define timeouts/limits/constants at module top

Error Handling

• Wrap integration failures as RuntimeError with actionable messages
• Keep original cause via raise ... from e
• Log failures at error level (no secrets in logs or exceptions)

Resource Cleanup

• If an adapter owns network/IO resources, implement close() (async if needed)
• close() should be best-effort: log and continue on cleanup errors
• The wiring layer (controller-side DI / setup) must always close adapters in finally

Database (Repository) Adapters

Repository adapters follow the thin wrapper pattern:

• Delegate CRUD operations to generic helpers in {db_engine}_aggregate_crud.py
• Define TABLE_NAME and STORED_MODEL_NAME constants at module top
• Validate db connection in __init__
• Wrap asyncpg.PostgresError as RuntimeError (done in CRUD helpers, not repeated in adapters)
• Return validated domain models (deserialized via AggregateRoot.from_payload())

Dependency Direction

• shared/src/shared/domain/** → no dependencies on project code (standard library + Pydantic only)
• shared/src/shared/services/** and shared/src/shared/ports/** → may depend only on shared/src/shared/domain/**
• shared/src/shared/adapters/** → may depend on shared/src/shared/domain/**, shared/src/shared/services/**, and shared/src/shared/ports/**
• Controller directories (e.g., api/, app/, cli/) → may depend on shared/ (domain, services, ports, adapters), but shared/ must not depend on controllers
• shared/src/shared/adapters/database/{db-engine}/migrations/** and shared/src/shared/config/** → may depend on shared code but must not be imported from domain, services, or ports

Change Workflow Checklists

Adding New External Capability

1. Add a new Port in shared/src/shared/ports/
2. Implement an Adapter in shared/src/shared/adapters/.../
3. Update/create a Service method in shared/src/shared/services/ to orchestrate via ports
4. Wire the adapter in the controller layer as needed (do not wire inside shared)

Adding Database Read Model / Cross-Aggregate Lookup

1. Add a QueryPort in shared/src/shared/ports/database/queries/
2. Implement the adapter in shared/src/shared/adapters/database/.../queries/
3. Use services to orchestrate; do not expose adapters directly to controllers

Adding a New Aggregate Repository

1. Create a migration
2. Add a repository port in shared/src/shared/ports/database/repositories/
3. Implement a repository adapter — thin wrapper delegating to {db_engine}_aggregate_crud helpers
4. Wire the adapter in the controller layer

Adding New Domain Rule

1. Enforce in the domain model (entity/value object) with clear ValueError messages
2. Add unit tests in shared/tests/ for both valid and invalid cases

Changing Error Message or Validation

1. Update/extend tests to match the new behavior and message (tests commonly use match=...)

Cross-References

• Project structure: See project-structure.md for canonical paths, directory tree, and layer responsibilities
• Domain models: See domain-models skill for DDD building blocks
• FastAPI: See fastapi skill for API-specific conventions
• CLI: See cli-conventions skill for Click-specific conventions
• Python patterns: See python skill for type hints and error handling