Skip to content

Clean Architecture

Principle · clean-architecture

Loads automatically whenever the agent touches code this principle governs. Install the skills · View source

The point of layering is not tidiness. It is that business rules should be readable, testable, and changeable without starting a database, a broker, or a web server — and that swapping any of those should not touch a line of business logic.

One rule generates almost everything else: dependencies point inward, toward the domain.

Config Resolution

  1. Read .msskills/config.yaml in the repo root; look up paths.architecture.
  2. A custom document exists at that path → read its frontmatter mode:
  3. override → use it alone; ignore the defaults below.
  4. overlay (default, or no mode key) → read defaults first, then apply the custom document's sections on top, matched by exact heading; new sections append.
  5. A path is configured but no file exists there → say which path is missing, then use the defaults.
  6. No config file or no paths.architecture key → use defaults.
  7. .msskills/stack.md exists → express every rule in that stack's idiom: its module system, its dependency-injection style, its package conventions, its architecture-test tooling.
  8. No project document and no stack profile → use the defaults, and say once that stack-profile would make the guidance concrete.

Self-Validation Checklist

STOP after generating or reviewing each class, module, or file. Verify every check. Fix failures before presenting.

  1. LAYER PLACEMENT: Can you name this file's layer — domain, application, adapter, infrastructure — and does its responsibility match? If it does two layers' jobs (a handler that contains a business rule, an entity that writes SQL) → split it.
  2. DEPENDENCY DIRECTION: Does every import point inward or sideways within the layer? An inner layer importing an outer one → invert it with an interface owned by the inner layer.
  3. FRAMEWORK ISOLATION: Is the domain free of framework and library types — annotations, ORM base classes, HTTP types, broker clients, serialisation attributes? If not → move them to the adapter and map at the edge.
  4. PORTS OWNED INWARD: Is each outbound interface declared in the layer that uses it, not the layer that implements it? A repository interface next to its database implementation is not a port; it is a package split.
  5. SINGLE RESPONSIBILITY: Can you describe this class in one sentence without "and"? If not → extract. Watch for the common pair: business logic and persistence.
  6. OPEN/CLOSED: Would a new variant of an existing concept require editing a switch in several files? If yes → introduce polymorphism, but only for a variation that has actually happened twice.
  7. LISKOV: Does every implementation of an interface honour its contract, or do some throw UnsupportedOperation, tighten preconditions, or return null where others cannot? If so → the abstraction is wrong; split it.
  8. INTERFACE SEGREGATION: Does any caller depend on an interface where it uses one of eight methods? Split by consumer need, not by implementation convenience.
  9. DEPENDENCY INVERSION: Does business logic depend on an abstraction it owns, or on a concrete client? Constructor-injected, never resolved from a static locator or newed inline.
  10. TESTABLE WITHOUT I/O: Can this class's behaviour be tested without a database, network, or clock? If not → push the side effect outward and inject time, IDs, and randomness.
  11. PATTERN JUSTIFIED: If a named design pattern appears, can you state the force that demands it in one sentence? "It is more extensible" is not a force → remove the indirection.

All checks pass → state "Passes clean-architecture. [next step]."

Active Anti-Pattern Scan

Any box you can check is a defect. Fix it before presenting.

  • [ ] Anaemic Domain: entities are field bags and all rules live in services → move the invariants onto the type that owns them.
  • [ ] Framework in the Domain: ORM, HTTP, or serialisation annotations on domain types → map at the adapter boundary instead.
  • [ ] Leaky Repository: a port that returns the ORM's query builder, a DataSet, or a framework page type → return domain types.
  • [ ] Fat Controller: business rules, orchestration, and validation in a handler → the handler parses input, calls one application service, and renders output.
  • [ ] Service Locator: dependencies fetched from a static container or singleton inside a method → constructor injection.
  • [ ] Circular Dependency: two modules importing each other → extract the shared concept, or invert one direction with an interface.
  • [ ] Utility Dumping Ground: a utils, helpers, or common package with unrelated statics → move each function next to the concept it belongs to.
  • [ ] Speculative Abstraction: an interface with exactly one implementation and no second one in sight, or a plugin system nobody plugs into → inline it until a real second case appears.
  • [ ] Pattern Theatre: factories producing factories, a strategy with one strategy, an observer with one observer → delete the indirection.
  • [ ] Leaked External Model: a third-party or upstream DTO used directly as a domain type → add an anti-corruption layer and translate at the edge.
  • [ ] Hidden Global State: static mutable state, ambient singletons, DateTime.now() buried in logic → inject it.
  • [ ] Transaction in the Domain: commit or session handling inside domain code → own the transaction in the application layer.

Ambiguity Signals

Route these through collaborative-judgment. Each has two defensible answers.

  • How many layers. Full four-layer separation is right for a service with real domain complexity and ceremony for a CRUD adapter over one table. Propose the lighter shape and let the user ask for more.
  • Port granularity. One repository per aggregate is conventional; one narrow port per use case is more honest about what each caller needs and produces more interfaces. Both defensible.
  • Mapping cost. Separate domain, persistence, and API models are the clean answer and triple the type count. Sharing one model is pragmatic until the first field that only one of them needs.
  • When an abstraction has earned its place. The rule of three is a guideline, not a law; a second case with the same reason to change can justify it, and three coincidentally similar cases do not.
  • Where a rule lives when two aggregates share it. A domain service, a policy object, or duplication in both — the right answer depends on which one owns the invariant.