Coding Python Clean Architecture

Layered Architecture for Python

Overview

Framework-agnostic, typed Python architecture optimized for change, testability, and clear boundaries. Inner layers never import from outer layers. Domain stays pure.

Primary goal: Keep the cost of change low over the lifetime of the system. Business logic should be easy to understand, test, and evolve independently of infrastructure choices.

Target: Python >=3.10 | Typed | Pure domain | Framework-agnostic core | No compatibility shims

When to Use

• Starting a new Python project that needs long-term maintainability
• Reviewing existing code for architecture violations (imports crossing layers)
• Deciding where a new class/function belongs (domain vs application vs adapter)
• Structuring data flow between external inputs, domain logic, and outputs
• Setting up testing strategy with proper isolation
• Implementing reliability patterns (UoW, outbox, idempotency)
• Refactoring framework-centric code to layered architecture

When NOT to use:

• Projects already using a different established architecture

Single-file scripts and throwaway prototypes stay in this skill - they route to SCRIPT mode (see script-mode.md and the mode-selection table below), one of this skill's four operating modes. Do not drop the skill for them.

Capability check (REVIEW / judgment). Spotting architecture/layer violations and the true-vs-accidental-duplication call is capability-sensitive - a weaker model misjudges it. Run REVIEW mode on a pinned sonnet subagent (opus for a large or high-stakes codebase); pin the tier per dispatch, never inherit the session model. If you keep it inline and the session is on a lesser tier, offer switch-model-or-continue. See bitranox:process-agents-subagent-driven-development ("The session model is fixed" / "Concrete tiers").

Layers & Dependency Rule

Inner layers never import from outer layers. Source code dependencies point inward only, toward more stable and abstract code.

Inner layers change less frequently and for more important reasons. Outer layers (I/O, formatting, frameworks) are plugins to the inner layers.

+------------------------------------------------------------+
|  Composition Root (outermost)                               |
|  compose.py, main entrypoints, framework wiring             |
|  +------------------------------------------------------+  |
|  |  Adapters                                             |  |
|  |  Route handlers, ORM repositories, Pydantic models,   |  |
|  |  message consumers, CLI parsers, HTTP clients          |  |
|  |  +------------------------------------------------+  |  |
|  |  |  Application                                    |  |  |
|  |  |  Use case classes, Protocol ports,               |  |  |
|  |  |  request/response dataclasses                    |  |  |
|  |  |  +------------------------------------------+   |  |  |
|  |  |  |  Domain                                   |   |  |  |
|  |  |  |  @dataclass(frozen=True) entities,         |   |  |  |
|  |  |  |  Value Objects, Domain Services,           |   |  |  |
|  |  |  |  Domain Events                             |   |  |  |
|  |  |  +------------------------------------------+   |  |  |
|  |  +------------------------------------------------+  |  |
|  +------------------------------------------------------+  |
+------------------------------------------------------------+

You may need more or fewer layers depending on complexity. The dependency rule always applies: dependencies point inward, abstraction increases inward.

Data Crossing Boundaries

Data passed across a boundary must always be in the form most convenient for the inner layer. Never pass ORM row structures or Pydantic request models inward -- adapters convert to domain types at the boundary.

Bounded and Sanitized Inputs at the Boundary

Adapters/edges are the only place external data enters, so two rules are enforced there:

Full per-sink escaping rules (SQL parametrization, HTML/XSS, shell, path traversal, deserialization) and the boundary-vs-internal-libs scope live in bitranox:coding-input-sanitization. Sanitize at the app/facing-API edge, NOT in the libraries between boundaries.

• Sanitize and bound every external input. Validate type and shape, enforce length limits (guard overflow/underflow), and handle arbitrary bytes/characters safely - non-ASCII, emoji, CJK, control characters, and binary data are rejected, normalized, or escaped, never trusted raw. The boundary parser (Pydantic at the edge) owns this, and it is covered by tests with adversarial and edge inputs.
• Keep memory bounded on large or unbounded data. When an adapter reads big files, huge database result sets, or huge log files, stream/iterate/chunk/paginate - never load the whole thing into memory or accumulate an unbounded collection. Ports expose iterators/pagination, not fully-materialized lists. Materialize only when the dataset is provably and safely bounded.
• Never trust structured input. Structured data crossing the boundary (a dict, JSON, an API/IPC payload, a deserialized object) is parsed into a typed model at the edge (Pydantic, then a domain type) - never consumed raw on the assumption its keys, types, or shape are correct. Exception: only where the user has deliberately chosen to skip the check.

Domain vs Use Cases

Use case request/response models must be independent -- no framework types, no entity references, no knowledge of delivery mechanism.

Domain-Centric Project Structure

Your top-level directory structure should tell readers about the system, not the framework. When new developers look at the project, their first impression should be "This is a lending platform" -- not "This is a FastAPI app."

Bad -- framework-centric:

src/api/routes/ middleware/ dependencies/ models/ schemas/ crud/

Good -- domain-centric:

src/lending_platform/
  loans/
    domain/         # Loan entity, Money VO, interest rules
    application/
      use_cases/    # create_loan.py, approve_loan.py
      ports/        # loan_repository.py, credit_check.py
    adapters/       # sqlalchemy_loan_repo.py, experian_credit.py
  customers/
    domain/ application/ adapters/
  composition/      # compose.py -- FastAPI only appears here

Design Principles

Brief, practical guidelines for structuring modules in Python layered architecture.

One actor per module. Each use case serves one stakeholder group. If two different teams need changes to the same module, split it. Use a Facade if callers need a single entry point.

Extend by adding, not by modifying. New behavior should mean new adapters implementing existing ports, not editing use cases. New output format? New adapter -- zero changes to the use case.

Honor port contracts. All adapters implementing a port must satisfy the same behavioral contract. Verify via contract tests -- a single parameterized test suite run against every adapter implementation (in-memory, PostgreSQL, Redis). If any adapter needs special handling, the port contract is wrong.

Keep ports narrow. A UserRepository port with find, save, delete, search, bulk_import, export_csv serves too many consumers. Split into focused ports: UserReader(Protocol), UserWriter(Protocol). Each use case depends only on the port slice it needs.

Depend on abstractions for volatile code. Domain and application layers import only Protocols and dataclasses. Stable things (stdlib, well-established libraries) are fine to depend on directly. The composition root is the only place that imports concrete adapter classes.

No concrete paths, URLs, or hostnames in the domain. Filesystem paths, endpoint URLs, hostnames, and external-service strings live ONLY in config passed (as a Pydantic model) from the composition root into adapters. The domain/application layers never read an env var or hardcode a path. This is testable: a unit test that instantiates an adapter with explicit values and reads NO env var proves the layer is pure -- if a test needs an env var or a real path constant, a config value leaked into the wrong layer.

Keep the domain immutable. Use @dataclass(frozen=True, slots=True) for entities and value objects. Push mutable state to adapters where it belongs, protected by the Unit of Work pattern. This is not a preference: "no mutable state in the domain" is a non-negotiable below.

Use Protocol to invert dependencies. When the natural direction of a function call opposes the desired dependency direction, introduce a Protocol in the inner layer and implement it in the outer layer. This is what makes the dependency rule possible.

Data Modeling

External Input -> Pydantic (validate) -> Dataclass (domain) -> Pydantic (serialize) -> Output
     (raw)         (adapter boundary)     (pure business logic)   (adapter boundary)      (JSON/API)

Never: Pass dict[str, Any] between modules or layers. Dicts acceptable only when: truly dynamic data, <3 keys, single function scope, no business logic.

Define types; do not suppress the checker. When pyright/mypy strict flags your code - often a third-party stub gap, not a real defect - DEFINE the missing types: the real annotation, or a typed facade you write (a Protocol plus a cast, or a local .pyi stub on stubPath) so call sites are complete. Never reach first for a per-file reportX = false, an exclude entry, or a bare # type: ignore - they blind the same scope to real errors and rot silently; a narrow, rule-specific # pyright: ignore[rule] with a remove-when comment is the last resort. Worked example (rich-click's partially-typed decorators fixed with a module-cast Protocol facade): bitranox:coding-python-enforce-data-architecture-strict.

Core Patterns

Ports (Application-owned Protocols)

class PaymentPort(Protocol):
    async def charge(self, customer_id: str, amount: Money) -> str: ...

Ports are Protocols in application/ports/. Adapters implement them. The business rules own the interface; adapters are plugins. See port-contracts.md for all standard port definitions (UoW, Outbox, IdempotencyStore, IdProvider, Clock).

Unit of Work

class UnitOfWork(Protocol[D]):
    async def run(self, fn: Callable[[D, RequestContext], Awaitable[T]], *,
                  ctx: RequestContext | None = None, timeout: float | None = None) -> T: ...

• Runs callable inside transaction scope
• Supplies transaction-bound repos/adapters
• Never pass raw connections/locks into core

Reliability Patterns

For resilience/self-healing at the adapter boundary (retry+backoff, health-check/evict/replace, circuit breaker, graceful degradation), see bitranox:coding-resilience.

Boundaries

Plugin Architecture

The database and UI are plugins to the business rules. Core business rules are kept separate from, and independent of, components that are either optional or implemented in many different forms. All dependency arrows point inward.

# Business rules own the interface (port)
class WikiPageRepository(Protocol):
    def find(self, slug: str) -> WikiPage | None: ...
    def save(self, page: WikiPage) -> None: ...

# Adapters are plugins -- each implements the same port
class InMemoryWikiPageRepo: ...     # development/testing plugin
class FileSystemWikiPageRepo: ...   # flat-file plugin
class PostgresWikiPageRepo: ...     # production plugin

The boundary sits at the Protocol. Business rules know nothing about which adapter is plugged in.

Boundary Crossing Modes

Recommended approach: Start at source-level. Push decoupling to where a service could be formed, but keep components in the same process. Escalate to deployment-level or service-level only when development, deployment, or operational issues demand it.

At runtime, boundary crossing is just a function call. The trick is managing source code dependencies so they always point toward the inner layer. Use Protocol to invert dependencies when control flow opposes the desired dependency direction.

Partial Boundaries

Full boundaries are expensive. When the cost is too high, use a partial strategy:

When to draw boundaries: Watch for friction. Implement boundaries at the inflection point where the cost of implementing becomes less than the cost of ignoring. Review boundary decisions as the system evolves.

I/O Boundary Isolation

Split behaviors that are hard to test from behaviors that are easy to test into two modules. The I/O-facing module contains hard-to-test behavior stripped to its barest essence; the logic module contains all extracted testable logic.

Presenter example:

@dataclass
class LoanViewModel:
    principal_display: str    # "$1,234.56"
    rate_display: str         # "4.50%"
    is_overdue: bool
    status_label: str         # "Active" | "Defaulted"

class LoanPresenter:
    def present(self, loan: Loan, today: date) -> LoanViewModel:
        return LoanViewModel(
            principal_display=f"${loan.principal.to_decimal():,.2f}",
            rate_display=f"{loan.rate * 100:.2f}%",
            is_overdue=loan.next_due_date < today,
            status_label="Defaulted" if loan.is_defaulted else "Active",
        )
# View just renders the ViewModel -- zero logic, strings/bools only

Principle: At every architectural boundary, look for the opportunity to isolate I/O. Extract testable behavior into a module that works with simple data structures and protocol ports.

Error Handling

Never let raw exceptions leak to external consumers.

Folder Layout

src/<pkg>/
  <bounded_context>/
    domain/        # entities.py, values.py, events.py, services.py
    application/
      use_cases/
      ports/       # uow.py, outbox.py, idempotency.py, id_provider.py, clock.py, <entity>_repository.py
    adapters/
      memory/      # in-memory implementations
      uow/
    platform/      # config.py, observability.py
  composition/     # compose.py (wiring)
  testing/         # Testing API (NOT deployed to production)

Flatten if single small domain; re-introduce <bounded_context>/ on growth.

Guardrails: domain imports nothing outside domain. application imports domain and its own ports. adapters implement application.ports and map to/from domain DTOs. Composition wires everything.

Each package directory requires an __init__.py (can be empty). Omit only if using implicit namespace packages (PEP 420).

Package Organization

Keep packages cohesive and their dependency graph clean:

• Group by change reason. Modules that change together belong in the same package. A requirement change should ideally touch only one package.
• Split to reduce coupling. If consumers only use a fraction of your package, split it. Don't force dependents to pull in modules they never import. Break fat common/ packages into focused ones: common_types, common_testing, common_infra.
• No cycles in import graphs. Cycles make independent testing and reasoning impossible. Break cycles by introducing a Protocol in the inner layer or extracting shared types into a new package.
• Stable packages should be abstract. Packages with many dependents are hard to change -- protect them by making them mostly Protocols and base types. Packages with few dependents can be concrete and change freely.
• Enforce boundaries in Python. Python lacks package-private. Use import-linter contracts, _-prefixed modules, __all__ in __init__.py, and leading-underscore convention to enforce boundaries. Organization without encapsulation is just folder structure, not architecture.

Four packaging approaches from least to most robust:

Cross-Bounded-Context Communication

Bounded contexts must not share domain internals. Choose one integration style per boundary:

Rules:

• Never import directly between context_a.domain and context_b.domain
• Shared kernel must be minimal and change-controlled
• Event contracts are owned by the publisher; consumers maintain their own projections
• Use import-linter independence contracts to enforce (see review-checklists.md)

Independence & Decoupling

Decoupling Layers (horizontal)

Separate things that change for different reasons: UI concerns change independently from business rules, which change independently from persistence details.

Decoupling Use Cases (vertical)

Use cases are narrow vertical slices through all layers. Each use case changes at a different rate, for different reasons. Keep them separate: one module per use case in application/use_cases/.

Decoupling Modes

Start source-level. A good architecture allows sliding up and back down this spectrum without rewriting.

True vs Accidental Duplication

Common traps: Two use cases with similar DTOs (accidental -- each serves different actors). A DB record that looks like a view model (accidental -- keep layers decoupled). Two contexts with similar User types (accidental -- will evolve independently). Rule: When separating use cases or layers, similar code is usually accidental duplication. Verify before unifying.

Sync vs Async

The examples use async throughout. Adapt based on I/O profile:

Domain stays the same regardless: pure, sync, no I/O. The choice affects application and adapter layers only. Ports can define sync or async methods -- pick one per project and stay consistent.

Testing Strategy

The Test Boundary

Tests sit in the outermost layer -- nothing depends on them, but they depend inward. Tests that mirror production structure 1:1 (test class per production class) are structurally coupled and fragile. Refactoring production code breaks hundreds of tests.

Solution: Test through a Testing API -- a dedicated API that hides the structure of the application from tests.

Testing API

The Testing API has superpowers: seed data directly, freeze time, bypass auth, force testable states. It is a superset of the application's use cases.

# testing/api.py -- NOT deployed to production
class TestingAPI:
    def __init__(self, app: Application) -> None:
        self._app = app

    async def seed_user(self, user_id: str, name: str, role: str = "admin") -> User:
        """Bypass registration, create user directly."""
        ...

    def freeze_time(self, at: datetime) -> None:
        """Replace clock port with frozen clock."""
        self._app.clock = FrozenClock(at)

    async def execute_as(self, user_id: str, use_case, input):
        """Execute any use case as any user, bypassing auth."""
        ctx = RequestContext(user_id=user_id, roles=("admin",))
        return await use_case.execute(input, ctx=ctx)

# tests/conftest.py
@pytest.fixture
async def api() -> TestingAPI:
    app = create_test_app()  # In-memory adapters, no real infra
    return TestingAPI(app)

Test through behavior, not structure:

# GOOD: Tests business rules through the Testing API
class TestPlaceOrder:
    async def test_valid_order_succeeds(self, api: TestingAPI):
        await api.seed_user("u1", "Alice")
        order_id = await api.execute_as("u1", api.place_order, PlaceOrderCommand(...))
        assert order_id is not None

Keep the Testing API and its fakes/builders in a separate testing/ package that is never included in production builds.

Framework Isolation

Keep Frameworks at the Edge

Frameworks want deep integration. Protect your core by keeping them in outer layers only.

The Solution: Constrain Framework Imports

# WRONG: Framework in domain
class Order(DeclarativeBase): ...  # Entity married to SQLAlchemy

# RIGHT: Framework stays in adapters
# domain/entities.py
@dataclass(frozen=True, slots=True)
class Order: ...

# adapters/persistence/models.py
class OrderModel(DeclarativeBase):
    def to_domain(self) -> Order: ...
    @classmethod
    def from_domain(cls, order: Order) -> "OrderModel": ...

Frameworks you must depend on (stdlib, typing, dataclasses): that's fine -- but it should be a decision, not an accident.

Typed Facade for Untyped Third-Party Libraries

An untyped or partially-typed dependency makes a strict type checker (pyright strict, mypy --strict) report errors at its call sites (reportUnknownMemberType, "type is partially unknown", reportUnknownArgumentType). Do NOT reach for the blunt fix of disabling the rule -- globally, via a per-directory executionEnvironments block, or scattered # type: ignore / # pyright: ignore at every call site. Disabling blinds the checker to real regressions in your OWN code too, and the suppression spreads.

Instead wrap the untyped surface behind a small typed facade / adapter module at the boundary: re-export the handful of affected callables through thin wrappers with explicit, fully-known signatures, and confine the single unavoidable # pyright: ignore (or # type: ignore) to that one file. The rest of the codebase imports the typed wrappers and stays strict-clean; the checker still guards everything else. Same edge-quarantine habit as the rest of Framework Isolation, applied to the type layer.

Keep the facade minimal: wrap only the members the checker flags (import cleanly-typed members directly), and give each wrapper a concrete, fully-known return type -- never an unbound generic that re-introduces "unknown" inference at call sites.

# adapters/cli/typed_click.py -- the ONE file carrying the suppression
import rich_click as click
from collections.abc import Callable
from typing import Any

# Non-generic alias: works both as @option(...) and as a value in a
# shared-options functools.reduce list.
_Decorator = Callable[[Callable[..., Any]], Callable[..., Any]]

def option(*param_decls: str, **attrs: Any) -> _Decorator:
    return click.option(*param_decls, **attrs)  # pyright: ignore[reportUnknownMemberType]

# Call sites import the typed wrapper and stay strict-clean:
#     from .typed_click import option
# Cleanly-typed members stay direct: click.command, click.echo, click.Context, ...

Disable the rule only as a genuine last resort -- when the untyped surface is too broad to wrap sensibly (an entire dynamic web framework emitting hundreds of unrelated findings, not a few decorators). Then scope the suppression as narrowly as possible (a single executionEnvironments root, never the whole project) and disable only that one rule, never the others.

Composition Root

The composition root is the outermost wiring point. Nothing depends on it -- it depends on everything.

Responsibilities: Create factories/strategies, wire concrete adapters to abstract ports, hand control to the application.

You can have multiple composition roots for different environments:

# composition/main_prod.py
def create_app() -> Application:
    config = ProdConfig.from_env()
    return Application(uow=PostgresUoW(config.db_url), cache=RedisCache(config.redis_url))

# composition/main_test.py
def create_app() -> Application:
    return Application(uow=InMemoryUoW(), cache=InMemoryCache())

# composition/main_dev.py
def create_app() -> Application:
    return Application(uow=SqliteUoW(":memory:"), cache=InMemoryCache())

The composition root is deliberately dirty. It imports every concrete class so the rest of the system stays abstract. Keep it thin: parse config, construct, wire, start. No business logic.

Deployment vs Architecture

Architecture is defined by boundaries and dependencies, not by process boundaries. Microservices are a deployment choice, not an architectural choice.

Rule: Start monolith with clean internal boundaries. Extract services only when you have a proven need for independent deployment. A well-structured monolith is architecturally superior to poorly structured microservices.

Details (What Is NOT Architecture)

Key distinction: The data is architecturally significant. The database is not.

Observability & Security

• Structured logging at adapter boundaries; thread trace_id via contextvars
• Health (/health) and readiness (/ready) checks when transport exists
• RED metrics (Rate, Errors, Duration) per use case/adapter
• Centralize config in one module; never read os.environ outside it
• AuthN/Z in adapters; pass RequestContext (claims/roles) to use cases
• PII/PHI boundary mappers redact by default

Non-Negotiables Checklist

• Dependencies point inward only (enforced via import-linter)
• Domain pure (no I/O, logs, frameworks, mutable state)
• No concrete paths, URLs, hostnames, or env-var reads in domain/application (config passed in from the composition root)
• Use cases free of framework/driver types
• Request/response DTOs independent of entities and framework types
• UoW binds transaction-scoped repos
• Deterministic lock ordering for multi-aggregate ops
• Outbox + Idempotency implemented
• Boundary validation (Pydantic at edges)
• Boundary input sanitized and length-bounded (types, encoding; non-ASCII/emoji/CJK/binary handled safely) and tested
• Large/unbounded reads (big files, huge DB results, huge logs) stream/paginate; no full materialization unless provably bounded
• Money as integer minor units (cents)
• Observability hooks at boundaries (trace_id via contextvars)
• UoW generic over typed deps (UnitOfWork[D]); no Mapping[str, Any] for deps
• Domain events extend DomainEvent base TypedDict
• In-memory unit tests + contract test outline
• Top-level folders reveal the domain, not the framework
• Testing API with superpowers (seed, freeze time, bypass auth)
• No framework base classes in domain entities
• Package dependency graph is acyclic

Operating Modes

Selecting a Mode

The agent

…(truncated)