🚀 Sincpro Framework: Application Layer Framework within Hexagonal Architecture

⚡ Quick Start

Here's a quick example to get you started with the Sincpro Framework:

🏁 Quick Example

from sincpro_framework import DataTransferObject, Feature, UseFramework

# 1. The bus of one bounded context
framework = UseFramework("greetings")


# 2. A dependency: every Feature reads it as self.greeter
class Greeter:
    def greet(self, name: str) -> str:
        return f"Hello, {name}!"


framework.add_dependency("greeter", Greeter())


# 3. A use case: its command, its response, and the Feature that answers it
class CommandGreet(DataTransferObject):
    name: str


class ResponseGreet(DataTransferObject):
    message: str


@framework.feature(CommandGreet)
class Greet(Feature):
    def execute(self, dto: CommandGreet) -> ResponseGreet:
        return ResponseGreet(message=self.greeter.greet(dto.name))


# 4. Execute it through the bus
result = framework(CommandGreet(name="Alice"), ResponseGreet)
print(result.message)  # Hello, Alice!

That is the whole framework: a use case, its DTO, and a bus that executes it. Observability comes with it — a span per DTO, errors reported, logs correlated — without configuring anything.

When the same catalog has to be reachable from outside the process, an entrypoint publishes it over a protocol without touching the use case. That is transport, and it comes later.

Now you are ready to explore more complex use cases! 🚀

🤖 Skills for AI coding agents

This repository is a Claude Code plugin marketplace. The sincpro-framework plugin carries 12 skills: the framework's context, vocabulary and architecture, the mistakes to avoid, and one skill per component (persistence, criteria, events, caching, auth, observability, settings, entrypoints, operations, analytics). An agent working in any service that uses the framework reads them before writing code.

Install once per machine (needs read access to this repository through your git credentials):

claude plugin marketplace add Sincpro-SRL/sincpro_framework
claude plugin install sincpro-framework@sincpro

Pull the skills of the latest commit:

claude plugin marketplace update sincpro
claude plugin update sincpro-framework@sincpro

From a clone of this repository, make skills-install, make skills-update and make skills-validate run the same commands. Inside a Claude Code session the equivalents are /plugin marketplace add Sincpro-SRL/sincpro_framework and /plugin install sincpro-framework@sincpro.

To give every collaborator of a service the skills with no command, commit this to that service's .claude/settings.json; Claude Code offers to install it when the folder is trusted:

{
  "extraKnownMarketplaces": {
    "sincpro": {
      "source": { "source": "github", "repo": "Sincpro-SRL/sincpro_framework" }
    }
  },
  "enabledPlugins": { "sincpro-framework@sincpro": true }
}

The plugin has no version: each commit is a version, so an update always brings the skills that match the code on main. Agents other than Claude Code reach the same guidance through Sincpro's knowledge MCP server.

Updating existing services

The persistence and event refactors in main (e6dbe87, manifest 3.14.5) change public imports and storage recipes. See the upgrade guide for Forge and MCP Odoo before changing their dependency locks: it separates compatibility, history migration, event sourcing and durable delivery, and lists the verified limitations.

Updating the skills plugin does not update the installed Python package. Conversely, editing this clone does not update a cached plugin or the knowledge MCP server. Check the consumer's installed artifact and the skill revision together; a version label alone does not prove that the refactored APIs were published.

📑 Table of Contents

1. Overview of Hexagonal Architecture
   • Key Layers of Hexagonal Architecture
   • Why Use a Unified Bus Pattern?
2. Key Features of the Sincpro Framework
   • DTO Validation with Pydantic
   • Dependency Injection
   • Inversion of Control (IoC)
   • Context Manager for Metadata Propagation
   • Interceptors
   • Error Handling at Different Levels
   • Bus Pattern for Component Communication
   • Decoupled Logic Execution
   • Application Service Orchestration
   • IDE Support with Typing
   • Crons
   • Registry of one process
   • One process, the loops it runs
   • Migrations
   • Data analysis
   • Runtime use cases
   • Workflows (experimental)
   • Persistence and the ORM
3. Features vs. Application Service
4. Example Usage for a Payment Gateway
   • Configuring the Framework
   • Best Practices for Imports
   • Sample Configuration in __init__.py
5. Recommended Infrastructure Structure
   • dependencies.py — Adapter Registration
   • framework.py — Wiring with DependencyContextType
   • __init__.py — Bootstrap the Bounded Context
   • Testing Dependency Consistency
   • Handler Lifetime — one instance serves every execution
   • Swapping a Dependency in a Test
   • Asserting What a Use Case Published
   • Checking Imports and Layers
6. Creating a Feature
7. Creating an Application Service
8. Executing a Use Case
9. Summary
10. Error Handling
11. Persistence (ORM) — aggregates, repository, queries, hooks, events
    • What it covers
    • How to, by topic
12. Crons — a registry per bounded context, a background process, one run per tick
13. Migrations — every context and store on one timeline, one command
14. Data analysis — a query read once, merged page by page, handed to pandas, polars or DuckDB
15. Runtime use cases — use cases stored as source, loaded onto the bus without a deploy
16. Workflows (experimental) — Commands composed as data, for a node editor's preview
17. Documentation
18. Entrypoints: exposing the bus — transport, not domain
    • MCP tools
    • JSON-RPC
    • gRPC
    • Bounded contexts across services — the context map: a context hosted by another service
19. Observability — tracing (OTLP) + errors (Sentry/GlitchTip)
20. Configuration or settings
21. Variables
22. Tests & coverage
23. Python 3.14 & Free-Threading Notes

🔍 Overview of Hexagonal Architecture

Hexagonal Architecture, also known as Ports and Adapters, is an architectural approach that aims to decouple core business logic from external dependencies. It organizes the system into distinct layers: domain, application, and infrastructure, enhancing maintainability, scalability, and adaptability.

🏗️ Key Layers of Hexagonal Architecture

• Core Domain: This layer encapsulates essential entities, value objects, and domain services representing the core business rules and behaviors. It is kept strictly independent from infrastructure concerns, preserving business logic integrity.
• Application Layer: Orchestrates user requests, processes domain responses, and mediates interactions between domain and external systems to ensure effective workflow execution.
• Infrastructure Layer: Contains adapters for interacting with databases, APIs, messaging systems, and other services. It handles data transformation to ensure compatibility with the domain and application layers.

🤔 Why Use a Unified Bus Pattern?

The Sincpro Framework adopts a unified bus pattern as a single point of entry for managing use cases, dependencies, and services within a bounded context. This simplifies the architecture by encapsulating all requirements of a given context, ensuring a clear and consistent structure.

Using a unified bus allows developers to access all dependencies through a single environment, eliminating the need for repeated imports or initialization. This approach ensures each bounded context is self-sufficient, independently scalable, and minimizes coupling while enhancing modularity.

🔑 Key Features of the Sincpro Framework

The Sincpro Framework follows hexagonal architecture principles, promoting modularity, scalability, and development efficiency. Here are its core features:

✅ DTO Validation with Pydantic

• Utilizes Pydantic to validate Data Transfer Objects (DTOs).
• Ensures only well-structured data is allowed into core business logic, reducing errors and maintaining data integrity.

🧩 Dependency Injection

• Facilitates integration of user-defined dependencies, promoting modular design.
• Enhances unit testing by allowing easy mocking or replacement of dependencies.

🔄 Inversion of Control (IoC)

• Automates the instantiation and configuration of components, reducing boilerplate code.
• Encourages loose coupling, making systems more adaptable and maintainable.

Interceptors

See docs/core/interceptors.md for the contract and runnable recipes.

• @bus.interceptor(CommandX) runs a function around one use case — @bus.interceptor() around all of them — however it is executed: at the entry, from an ApplicationService, a workflow step or a scheduled tick.
• It can veto, adjust the Command (model_copy), adjust the response, or answer by itself; it never changes their class.
• Recipes: credit check, audit trail, idempotency, cache, retry on a stale write, feature flags, timing.
• @bus.feature(CommandX, replaces=CoreFeature) has another handler answer instead of the core's — explicitly: the build log, the span and introspection.describe() say who replaced it.

📡 Context Manager for Metadata Propagation

• Provides automatic metadata propagation across Features and ApplicationServices without manual parameter passing.
• Uses Python's contextvars for thread-safe context storage and isolation.
• Supports nested contexts with override capabilities for complex workflows.
• Enriches exceptions with context information for better debugging and observability.

# Simple context usage
with app.context({"correlation_id": "123", "user_id": "admin"}) as app_with_context:
    result = app_with_context(some_dto)  # Context automatically available in handlers

# Nested contexts with overrides
with app.context({"env": "prod", "user": "admin"}) as outer_app:
    with outer_app.context({"env": "staging"}) as inner_app:  # Override env, inherit user
        inner_app(dto)  # env="staging", user="admin"

# Access context in Features and ApplicationServices
class PaymentFeature(Feature):
    def execute(self, dto: PaymentDTO) -> PaymentResponse:
        correlation_id = self.context.get("correlation_id")
        user_id = self.context.get("user_id")
        # Use context in business logic...

Propagating context into a ThreadPoolExecutor

• The context overlay lives in a ContextVar, which is isolated per OS thread. A plain executor.submit(bus.execute, dto) runs execute in a new thread that never saw the overlay's set() — every Feature's self.context there silently falls back to the (usually empty) shared context.
• bus.thread_context() on a built bus (framework.bus or self.feature_bus, not UseFramework itself) captures the calling thread's current context and returns a ThreadContextBus — pass .execute (not the raw bus) to the executor instead.
• Call thread_context() once per task you submit, not once for a whole batch: a captured contextvars.Context can only be entered by one thread at a time, so sharing a single one across concurrent workers raises RuntimeError.

from concurrent.futures import ThreadPoolExecutor

class SyncManyFeature(ApplicationService):
    def execute(self, dto: SyncManyDTO) -> SyncManyResponse:
        with ThreadPoolExecutor(max_workers=3) as executor:
            futures = [
                # captured here, in this thread, once per task
                executor.submit(self.feature_bus.thread_context().execute, item_dto)
                for item_dto in dto.items
            ]
            results = [f.result() for f in futures]
        return SyncManyResponse(results=results)

Async fan-out with get_async_bus()

Current limitation: the async facade does not create the same per-call context boundary as UseFramework.__call__. Handler writes can persist between calls without an explicit scope; parallel calls must not share a mutable context mapping. See the current audit.

• thread_context() is for sync code manually managing a ThreadPoolExecutor. If the caller is already async def (an async host handler, a script) and wants to fan out several DTOs concurrently instead, use bus.get_async_bus() (or the shortcut framework.get_async_bus()).
• Solves the same context-propagation problem as thread_context(), via asyncio.to_thread itself (stdlib already propagates contextvars into the worker thread — no ThreadContextBus involved here). Opposite reuse rule: an AsyncBus is stateless, so get it once and await/asyncio.gather many calls on it — each call gets its own fresh context snapshot, unlike ThreadContextBus's single-use-per-snapshot restriction.
• Runs each call via asyncio.to_thread (stdlib), so Feature/ApplicationService stay 100% sync — no "async Feature" variant to maintain.

import asyncio

async def handle_request(framework, dto_a, dto_b, dto_c):
    async_bus = framework.get_async_bus()
    result_a, result_b, result_c = await asyncio.gather(
        async_bus(dto_a), async_bus(dto_b), async_bus(dto_c),
    )
    return result_a, result_b, result_c

Cancellation and fan-out reliability

• Cancelling the awaiting coroutine (e.g. a asyncio.wait_for(...) timeout) does not stop the Feature/ApplicationService already running in its worker thread — Python cannot forcibly kill a thread. Design for that (idempotency, no assumption that a timeout actually aborted the work) rather than relying on cancellation to stop it.
• For fan-out with proper partial-failure handling, prefer asyncio.TaskGroup (3.11+) over asyncio.gather: it cancels sibling tasks on the first failure and raises an ExceptionGroup, instead of gather's default of leaving siblings running and swallowing all-but-the-first exception unless return_exceptions=True is passed.

⚠️ Error Handling at Different Levels

• Provides centralized error handling at three independent levels: global, app service, and feature.
• First registered handler executes first. On re-raise, the framework delegates to the next handler in the chain.
• Handlers can be registered at any point — before or after the first execution — and take effect immediately.
• Ensures consistent error management, improving overall reliability.

🚌 Bus Pattern for Component Communication

• Implements a bus mechanism to facilitate communication between Feature and ApplicationService components.
• Decouples component interactions, resulting in more flexible and scalable business logic.

🧩 Decoupled Logic Execution

• Supports independent execution of use cases through the Feature component, promoting separation of concerns.
• For example, a user registration workflow can be broken down into steps like input validation, profile creation, and email notification.

🎻 Application Service Orchestration

• Uses a feature bus to orchestrate multiple features into complex business workflows (e.g., customer onboarding).
• Integrates smaller use cases into cohesive flows to manage entire business processes effectively.

📝 Reads and forms of one aggregate, built in

• Get, GetMany, LiteralSearch, Search, Preview and DomainEvents: one Feature (EntityReads) answers them from what the entity says about itself — see docs/persistence/entity-reads.md.
• Edit: a form's whole life in one command — keep the caller's draft, save it, discard it, find it again after the browser closed (Get(draft=True), Search(drafts=True)) — answered by EntityEdits; drafts live where the context is kept, by user or session — see docs/persistence/edits.md.

💻 IDE Support with Typing

• Uses type hints to enhance code quality and support features like autocompletion and type checking.
• Parameterize the bus as UseFramework[DependencyContextType]. Features get self.token_adapter; callers outside a Feature get the same instance as framework.deps.token_adapter.

Events over brokers

See docs/events/brokers.md.

• The domain keeps publish(event) and registering an event on a bus; Kafka, RabbitMQ, Redis or NATS carry it through FastStream, behind the [faststream] extra.
• FastStreamQueue(broker).start() for synchronous code, aput for async; subscribe(broker, Subscriber(...)) subscribes every event the buses registered.

Caching

See docs/caching/.

• Infrastructure the use case never sees: the composition says which Queries keep their answers — QueryCaching(store).on(bus, Query, CachePolicy(ttl=…, vary_by=("tenant_id",))).
• An answer is tagged by what the repository noted it read and let go of when a commit writes it; KeyValueStore is the provider contract — memory in the core, Redis/Valkey and Memcached as extras.

Registry of one process

See docs/registry/README.md.

• registry.add(billing) in the context, once the instance exists. An entrypoint imports the context packages and then reads registry.all() — a gateway, a Subscriber, serve().
• The registry is this process's memory, in the order the contexts were added. It does not build them and it does not sort them. Another replica builds its own. A context another service hosts stays on the context map.

One process, the loops it runs

See docs/process/README.md.

• Process stays up and runs each loop (run / stop) on its own thread until SIGINT, SIGTERM, or stop. CronGateway is a loop. Poll calls a function every interval.
• The function is the project's: a relay pass, or the Commands that claim a row. There is no job type. A broker stays faststream run. A server stays its gateway.

Crons

See Crons for the full section.

• A registry per bounded context (cron_payments = Crons("cron-payments")), one Cron class per cron, the buses it calls injected by their own names — a component beside the buses, not on one.
• By default in a background process: an in-memory orchestrator looks every 5 seconds and gives each run a thread of its own; one run per tick across replicas, explicit policies for overlapping and missed ticks.

Migrations

See Migrations for the full section.

• Every context and every store on one timeline: status says where the whole system stands, upgrade and downgrade --to move it as one, from a Makefile.
• The core needs no database: MigrationEngine is one abstract class for any store; Alembic is the engine shipped for SQL, behind the [migrations] extra.

Data analysis

See Data analysis for the full section.

• What a Criteria answered, held as a DataFrame under the repository's fingerprint of the read: the next page continues it, asking again reads nothing, a narrower filter is answered from a complete read.
• Handed on as Parquet or Arrow to a client, or through the Arrow PyCapsule interface to pandas, polars and DuckDB — pyarrow behind the [data-analysis] extra.

Runtime use cases

See Runtime use cases for the full section.

• Commands, Responses and a Feature or ApplicationService stored as Python source, loaded onto the bus while the service runs; a stored handler may replaces= one in code.
• A new generation of the bus is built, checked and swapped in whole; a refused one leaves the bus as it was.

Workflows (experimental)

See Workflows (experimental) for the full section.

• The Commands a bus answers composed as JSON — execute, code, for_each, fail, when — validated against the live bus, run with a trace, drawn as a graph for a node editor.

Persistence and the ORM

See Persistence (ORM) for the full section.

• An aggregate is a plain dataclass; a Feature reads and writes it through self.repository.
• One way to ask — Criteria — and one shape back — a page with its cursor, count and definition.
• Hooks, domain events, change tracking, event sourcing and an outbox on the same repository.

entrypoint_mcp

See docs/entrypoints/ for the full section.

• One line publishes the bus as MCP tools: build_mcp_server(instance).run().
• Features and ApplicationServices become typed tools. Docstrings are the LLM context.
• Domain code stays host-agnostic. This extra is MCP only.

entrypoint_rpc

See docs/entrypoints/ for the full section.

• One process, several instances: RpcGateway({"qr": qr, "cybersource": cybersource}).run().
• Methods are qr.features.CommandCreateQREconomico / siat.app_services.CommandGenerateCUFD.
• context on the JSON-RPC request is framework.context + optional with_trace. OpenRPC 1.4 discovery.

entrypoint_grpc

See docs/entrypoints/grpc.md for the full section.

A bounded context hosted by another service. When two services run the same code, the context map places a context in the other one, and the caller's code does not change:

context_map:                                   # the conf file; SINCPRO_CONTEXT_MAP wins per context
  - context: billing
    at: grpc://billing-service:50051?timeout=5

billing(CommandIssueInvoice(...), ResponseIssueInvoice) is then answered by the service at that address, which hosts it in one line — billing.serve(address), billing.serve(address, Attach.THREAD) beside a REST API, Attach.PROCESS in a subprocess of its own, or open_host_routes([billing]) on the REST API's own port for http:// addresses. Here billing is a reference: never built, it forwards every call. DTOs, bytes, Decimal and the whole request context travel as they are, in chunks; the answer is the DTO, an exception raised there is raised here as itself, and only a DTO that does not fit fails — versions are never compared. Nothing configured, every context runs here; in code, billing.hosted_by("grpc://…") does the same without configuration (the last one applied wins: configuration, then hosted_by, then serving). The component is sincpro_framework.bus_communication.remote_execution. See docs/entrypoints/bounded-contexts-across-services.md.

• Same catalog, gRPC wire: GrpcGateway({"qr": qr}).run("0.0.0.0:50051").
• Methods are /qr.Features/CommandCreateQREconomico — unary, google.protobuf.Struct in and out.
• No protoc and no generated stubs: server reflection for discovery, .write_proto_files() when a Go/TS client wants the contract.

Observability (tracing + errors)

• Tracing (optional): OpenTelemetry spans on every DTO, export via OTLP (sincpro-framework[opentelemetry] + OTEL_EXPORTER_OTLP_ENDPOINT).
• Errors (optional): Sentry/GlitchTip capture on bus exceptions (sincpro-framework[sentry] + SENTRY_PYTHON_DSN in conf). Isolated client — does not call sentry_sdk.init(), does not reuse Odoo's client.
• Metrics (optional): every use case timed by bounded context and outcome, with nothing to declare; @metrics.counts / sums / measures on a use case by field reference (of(Command).field); Prometheus /metrics (sincpro-framework[prometheus]) or OTLP ([opentelemetry]), picked by OTEL_METRICS_EXPORTER (alias SINCPRO_METRICS_BACKEND) — guide.
• Independent: you can enable traces, errors, both, or neither.

⚙️ Features vs. Application Service

• Feature: Represents a discrete, self-contained use case focused on specific functionality, easy to develop and maintain.
• ApplicationService: Orchestrates multiple features for broader business objectives, providing reusable components and workflows.

💳 Example Usage for a Payment Gateway

The following example shows how to configure the Sincpro Framework for a payment gateway integration, such as CyberSource. It is recommended to name the framework instance to clearly represent the bounded context it serves.

🔧 Configuring the Framework

To set up the Sincpro Framework, configuration should be performed at the application layer within the use_cases directory of each bounded context.

sincpro_payments_sdk/
├── pyproject.toml
├── README.md
├── apps/
│   ├── cybersource/
│   │   ├── adapters/
│   │   │   ├── cybersource_rest_api_adapter.py
│   │   │   └── __init__.py
│   │   ├── domain/
│   │   │   ├── card.py
│   │   │   ├── customer.py
│   │   │   └── __init__.py
│   │   ├── infrastructure/
│   │   │   ├── logger.py
│   │   │   ├── aws_services.py
│   │   │   ├── orm.py
│   │   │   └── __init__.py
│   │   └── use_cases/
│   │       ├── tokenization/
│   │       │   ├── new_tokenization_feature.py
│   │       │   └── __init__.py
│   │       ├── payments/
│   │       │   ├── token_and_payment_service.py
│   │       │   └── __init__.py
│   │       └── __init__.py
│   ├── qr/
│   ├── sms_payment/
│   ├── bank_api/
│   ├── online_payment_gateway/
│   └── paypal_integration/
└── tests

📋 Best Practices for Imports

Each use case should import both the DTO for input parameters and the DTO for responses to maintain clarity and consistency.

📝 Sample Configuration in __init__.py

from typing import Type

from sincpro_framework import Feature as _Feature
from sincpro_framework import UseFramework as _UseFramework
from sincpro_framework import ApplicationService as _ApplicationService

from sincpro_payments_sdk.apps.cybersource.adapters.cybersource_rest_api_adapter import (
    ESupportedCardType,
    TokenizationAdapter,
)
from sincpro_payments_sdk.infrastructure.orm import with_transaction as db_session
from sincpro_payments_sdk.infrastructure.aws_services import AwsService as aws_service

# Create an instance of the framework
cybersource = _UseFramework()

# Register dependencies
cybersource.add_dependency("token_adapter", TokenizationAdapter())
cybersource.add_dependency("ECardType", ESupportedCardType)
cybersource.add_dependency("db_session", db_session)
cybersource.add_dependency("aws_service", aws_service)


# Define a custom Feature class to access the dependencies
class Feature(_Feature):
    token_adapter: TokenizationAdapter
    ECardType: Type[ESupportedCardType]
    db_session: ...
    aws_service: ...
    logger: ...


# Define a custom Application Service class to access dependencies
class ApplicationService(_ApplicationService):
    token_adapter: TokenizationAdapter
    ECardType: Type[ESupportedCardType]
    db_session: ...
    aws_service: ...
    logger: ...
    feature_bus: ...


# Add use cases (Application Services and Features)
from . import tokenization

__all__ = ["cybersource", "tokenization", "Feature"]

🏗️ Recommended Infrastructure Structure

When bootstrapping a bounded context with UseFramework, the recommended practice is to split framework wiring into three dedicated files under apps/<domain>/infrastructure/:

apps/
└── my_domain/
    ├── infrastructure/
    │   ├── dependencies.py   # registers adapters; declares DependencyContextType
    │   ├── framework.py      # defines local Feature/ApplicationService + config_framework()
    │   └── error_handler.py  # (optional) registers error handlers
    ├── services/
    │   ├── feature_a.py
    │   └── feature_b.py
    └── __init__.py           # creates the framework instance and imports services

dependencies.py — Adapter Registration

Declare all external adapters in one place and expose a DependencyContextType typing helper. This class is not instantiated — it is used only as a mixin to give Feature and ApplicationService subclasses IDE autocomplete for injected attributes.

# apps/my_domain/infrastructure/dependencies.py
from sincpro_framework import UseFramework

from my_sdk.adapters import PaymentAdapter, TokenizationAdapter


class DependencyContextType:
    """Typing helper — gives Features/AppServices IDE autocomplete for injected deps."""

    token_adapter: TokenizationAdapter
    payment_adapter: PaymentAdapter


def register_dependencies(framework: UseFramework[DependencyContextType]) -> UseFramework[DependencyContextType]:
    """Register all adapters with the framework instance."""
    framework.add_dependency("token_adapter", TokenizationAdapter())
    framework.add_dependency("payment_adapter", PaymentAdapter())
    return framework

framework.py — Wiring with DependencyContextType

Combine the framework base classes with DependencyContextType using multiple inheritance so that every Feature, ApplicationService and repository hook in this bounded context automatically inherits the typed attributes. Each also takes the request's context type as a parameter — Feature[Command, Response, MyContext], Hook[MyContext] — for a typed self.context.

# apps/my_domain/infrastructure/framework.py
from sincpro_framework import ApplicationService as _ApplicationService
from sincpro_framework import DataTransferObject  # re-exported for convenience
from sincpro_framework import Feature as _Feature
from sincpro_framework import UseFramework
from sincpro_framework.ddd import Hook as _Hook

from .dependencies import DependencyContextType, register_dependencies


class Feature(_Feature, DependencyContextType):
    """Base Feature for this bounded context — typed deps included."""

    pass


class ApplicationService(_ApplicationService, DependencyContextType):
    """Base ApplicationService for this bounded context — typed deps included."""

    pass


class Hook(_Hook, DependencyContextType):
    """Base repository hook for this bounded context — typed deps, `self.context`, `self.bus`."""

    pass


def config_framework(name: str) -> UseFramework[DependencyContextType]:
    """Create and configure the framework instance."""
    instance = UseFramework[DependencyContextType](name)
    register_dependencies(instance)
    return instance

Hooks are lazy, like the bus. A repository reads its hooks the first time it saves or reads, not when it is built — so it is registered in dependencies.py like any other dependency, and a hook imports its base from the context, as a Feature does:

services/hooks/__init__.py      billing_hooks = Hooks()
services/hooks/invoices.py      from my_domain import Hook
                                @billing_hooks.on(Invoice) class NumbersInvoices(Hook): ...
infrastructure/dependencies.py  billing_hooks.inject(framework)
                                framework.add_dependency("repository", Repository(database, billing_hooks))

The same names Features receive as self.token_adapter are available on the root as my_framework.deps.token_adapter. Use self.<name> inside a Feature / ApplicationService; use .deps from SDK callers, tests, and entrypoints.

__init__.py — Bootstrap the Bounded Context

Create the framework instance first, then import the service modules so that the @framework.feature and @framework.app_service decorators register against the already-created instance.

# apps/my_domain/__init__.py
from .infrastructure.framework import (
    ApplicationService,
    DataTransferObject,
    Feature,
    config_framework,
)

my_framework = config_framework("my-domain")

# Import services AFTER creating the instance so decorators register against it
from .services import feature_a, feature_b  # noqa: E402, F401

Testing Dependency Consistency

Every name a Feature or ApplicationService declares — on itself or on the DependencyContextType it inherits — has to be registered with add_dependency, or the handler fails the first time it runs. One assertion checks the whole bus, every handler at once:

# tests/my_domain/test_framework_setup.py
from sincpro_framework.runtime.testing import unregistered_dependencies

from my_sdk.apps.my_domain import my_framework


def test_every_declared_dependency_is_registered():
    assert unregistered_dependencies(my_framework) == {}

A failure names each handler with the dependencies it declares and nobody registered. The framework's own attributes (feature_bus, context, the logger) are never reported.

Handler Lifetime — one instance serves every execution

A Feature or ApplicationService is built once per UseFramework and reused by every execution, on every thread. Anything written to self inside execute is therefore shared: two concurrent calls overwrite each other's value, and the next call sees the last one.

# ❌ request data on self — concurrent executions read each other's value
def execute(self, dto):
    self.order = self.repository.get(dto.order_id)
    return self._total()

# ✅ request data in locals — self holds only the injected dependencies
def execute(self, dto):
    order = self.repository.get(dto.order_id)
    return self._total(order)

self.context is safe: it is isolated per execution.

Swapping a Dependency in a Test

override_dependencies keeps the production wiring and replaces only the adapters the test names, for the length of the block. Every Feature and ApplicationService of that bus sees the double, and so does framework.deps; the real adapter is back afterwards, even if the test fails. A name that was never registered is refused, so a typo cannot leave the real adapter talking to the outside world.

from sincpro_framework.runtime.testing import override_dependencies

from my_sdk.apps.payments import payments


def test_declined_card_is_not_retried():
    gateway = FakeGateway(declines=True)

    with override_dependencies(payments, gateway=gateway):
        result = payments(CommandCharge(amount=10), ResponseCharge)

    assert result.status == "declined"
    assert gateway.calls == 1

It patches the instances the bus holds, so it is meant for tests, not for live traffic.

Asserting What a Use Case Published

RecordingQueue is a queue like any other: put it behind the Publisher the bus already takes, and it keeps every event in order — handing each one on to the queue it wraps, when given one, so the subscribers still hear it.

from sincpro_framework.event_driven import Publisher, Subscriber, SyncQueue
from sincpro_framework.runtime.testing import RecordingQueue


def test_issuing_publishes_invoice_issued():
    published = RecordingQueue(SyncQueue(Subscriber(accounting)))
    billing = build_billing(publisher=Publisher(published))

    billing(CommandIssue(invoice_id="F-1"))

    assert [one.invoice_id for one in published.of(InvoiceIssued)] == ["F-1"]

Checking Imports and Layers

Two checks read the package's source — without importing it, so an SDK that opens a network connection on import is safe to check — and return what they find:

from sincpro_framework.runtime.testing import import_cycles, layer_violations


def test_no_import_cycles():
    assert import_cycles("my_sdk") == []


def test_layers():
    assert layer_violations("my_sdk") == []

import_cycles reports modules that import each other while loading, e.g. my_sdk.domain.codes → my_sdk → my_sdk.domain.codes. Imports inside a function or under if TYPE_CHECKING: are not counted — they are how a cycle is deliberately broken. The bootstrap this framework asks for (the context __init__ creates the bus, then imports services, which take the bus back) is not a cycle while the bus is bound before that import.

layer_violations applies the layering conventions. A module belongs to the first domain / adapters / services / infrastructure / entrypoints in its path, and to the context named by the path before it:

They are conventions, not a cage: pass the ones a project does not follow in ignore.

assert layer_violations("my_sdk", ignore=["adapters-are-independent"]) == []

Fixing a cycle. Break it where the dependency points the wrong way, not with a local import:

• Two modules share a type → move the type down, into the domain/ both already import.
• A module takes a name from its package __init__ → bind that name before the __init__ imports anything that leads back to the module (the bus goes first), or import it from the module that defines it.
• Two contexts need each other → the more foundational one stops reaching up: move the shared piece into it, or pass the value in the Command.
• An adapter module needs a different adapter module → a Feature composes both; the record they exchange goes to domain/. Files inside one folder under adapters/ may import each other, and that folder exposes one facade.

🛠️ Creating a Feature

To create a new Feature, follow these steps:

1. Create a Module for the Feature: Add a new Python file in the appropriate folder under use_cases.
2. Import the Framework and Required Classes: Import the configured framework instance and DataTransferObject.
3. Define the Parameter and Response DTOs: Use DataTransferObject to create classes for input parameters and responses.
4. Create the Feature Class: Define the Feature class by inheriting from the custom Feature class.

🖋️ Example of Creating a Feature

from sincpro_payments_sdk.apps.cybersource import cybersource, DataTransferObject, Feature


# Define parameter DTO
class TokenizationParams(DataTransferObject):
    card_number: str
    expiration_date: str
    cardholder_name: str


# Define response DTO
class TokenizationResponse(DataTransferObject):
    token: str
    status: str


# Create the Feature class
@cybersource.feature(TokenizationParams)
class NewTokenizationFeature(Feature):
    def execute(self, dto: TokenizationParams) -> TokenizationResponse:
        # Example usage of dependencies
        cybersource.logger.info("Starting tokenization process")
        token = self.token_adapter.create_token(
            card_number=dto.card_number,
            expiration_date=dto.expiration_date,
            cardholder_name=dto.cardholder_name
        )
        return TokenizationResponse(token=token, status="success")

🔄 Creating an Application Service

ApplicationService is used to coordinate multiple features while maintaining reusability and consistency. It orchestrates features into cohesive workflows.

💡 Example of Creating an Application Service

from sincpro_payments_sdk.apps.cybersource import cybersource, DataTransferObject, ApplicationService
from sincpro_payments_sdk.apps.cybersource.use_cases.tokenization import TokenizationParams


# Define parameter DTO
class PaymentServiceParams(DataTransferObject):
    card_number: str
    expiration_date: str
    cardholder_name: str
    amount: float


# Define response DTO
class PaymentServiceResponse(DataTransferObject):
    status: str
    transaction_id: str


# Create the Application Service class
@cybersource.app_service(PaymentServiceParams)
class PaymentOrchestrationService(ApplicationService):
    def execute(self, dto: PaymentServiceParams) -> PaymentServiceResponse:
        # Create the command DTO for tokenization
        tokenization_command = TokenizationParams(
            card_number=dto.card_number,
            expiration_date=dto.expiration_date,
            cardholder_name=dto.cardholder_name
        )
        tokenization_result = self.feature_bus.execute(tokenization_command)

        # Example usage of dependencies
        cybersource.logger.info("Proceeding with payment after tokenization")
        # Proceed with payment using the token (pseudo code for payment processing)
        transaction_id = "12345"  # Simulated transaction ID
        return PaymentServiceResponse(status="success", transaction_id=transaction_id)

⚙️ Executing a Use Case

Once a Feature or ApplicationService is defined, it can be executed by passing the appropriate DTO instance.

📌 Example of Executing a Use Case

from sincpro_payments_sdk.apps.cybersource import cybersource
from sincpro_payments_sdk.apps.cybersource.use_cases.tokenization import TokenizationParams, TokenizationResponse
from sincpro_payments_sdk.apps.cybersource.use_cases.payments import PaymentServiceParams, PaymentServiceResponse

# Example of executing a Feature
feature_dto = TokenizationParams(
    card_number="4111111111111111",
    expiration_date="12/25",
    cardholder_name="John Doe"
)

# Execute the feature
feature_result = cybersource(feature_dto, TokenizationResponse)
print(f"Tokenization Result: {feature_result.token}, Status: {feature_result.status}")

# Example of executing an Application Service
service_dto = PaymentServiceParams(
    card_number="4111111111111111",
    expiration_date="12/25",
    cardholder_name="John Doe",
    amount=100.00
)

# Execute the application service
service_result = cybersource(service_dto, PaymentServiceResponse)
print(f"Payment Status: {service_result.status}, Transaction ID: {service_result.transaction_id}")

📚 Summary

The Sincpro Framework provides a robust solution for managing the application layer within a hexagonal architecture. By focusing on decoupling business logic from external dependencies, the framework promotes modularity, scalability, and maintainability.

• Features: Handle specific, self-contained business actions.
• ApplicationServices: Orchestrate multiple features for cohesive workflows.
• entrypoint_mcp: Publish that catalog as MCP tools (sincpro-framework[mcp]).
• entrypoint_rpc: Publish one or more instances as JSON-RPC 2.0 methods (sincpro-framework[rpc]).
• entrypoint_grpc: Publish that same catalog as unary gRPC services (sincpro-framework[grpc]).

This structured approach ensures high-quality, maintainable software that can adapt to evolving business needs. 🚀

⚠️ Error Handling

The framework provides three independent error handler scopes: global (framework bus), feature, and app service. Register a handler with the corresponding method — handlers can be added before or after the first execution and always take effect immediately.

Basic usage

An error handler receives the exception. Return a value to suppress it:

from sincpro_framework import UseFramework

framework = UseFramework("my_app")

def handle_error(error: Exception):
    return {"error": str(error)}  # suppresses the exception

framework.add_global_error_handler(handle_error)

Scoped handlers

Each scope intercepts only the errors produced at that level:

# Only feature errors
framework.add_feature_error_handler(feature_handler)

# Only app service errors
framework.add_app_service_error_handler(app_service_handler)

# Everything that reaches the root bus
framework.add_global_error_handler(global_handler)

Registration lifecycle

Handlers can be registered at any point — before the bus is built or after — and take effect immediately:

framework = UseFramework("my_app")

framework.add_global_error_handler(base_handler)  # before first execution

framework(some_dto)  # first call triggers build

framework.add_global_error_handler(extra_handler)  # after build — also works

🔗 Advanced: Handler chaining

Every call to add_*_error_handler adds the handler to a chain. The first registered handler executes first. If it re-raises, the framework automatically delegates to the next handler in the chain.

Example: three-layer chain

# Registration order: auth → observability → base
# Execution order: auth → observability → base

def auth_handler(error: Exception):
    """First — intercepts auth errors; delegates everything else."""
    if isinstance(error, AuthenticationError):
        return {"ok": False, "detail": "unauthenticated", "code": 401}
    raise error  # delegates to observability_handler

def observability_handler(error: Exception):
    """Second — logs error, then delegates to base_handler."""
    log.error("unhandled error", exc_info=error)
    raise error  # delegates to base_handler

def base_handler(error: Exception):
    """Last — always returns a structured error, never raises."""
    return {"ok": False, "detail": str(error)}

framework.add_global_error_handler(auth_handler)          # 1st = runs first
framework.add_global_error_handler(observability_handler) # 2nd
framework.add_global_error_handler(base_handler)          # 3rd = final fallback

Persistence (ORM)

sincpro_framework.ddd is the vocabulary — aggregates, Criteria, repositories, events — and needs nothing installed. sincpro_framework.data_layer.orm is the SQLAlchemy adapter: pip install sincpro-framework[sqlalchemy].

What it covers

Example

from dataclasses import dataclass

from sqlalchemy import Column, Integer, Text
from sqlalchemy.orm import registry

from sincpro_framework.ddd import Criteria, Entity, EntityCollection
from sincpro_framework.data_layer.orm import Database, Repository, map_aggregates, template_table


@dataclass
class Product(Entity):                     # id, created_at, updated_at, version come from Entity
    name: str = ""
    price: int = 0


class Products(EntityCollection[Product]): ...


catalog = registry()
product_table = template_table.entity_table(
    "product",
    catalog.metadata,
    Column("name", Text, nullable=False),
    Column("price", Integer, nullable=False),   # a page is only ordered by NOT NULL columns
)
map_aggregates(catalog, {Product: product_table})

database = Database("sqlite:///catalog.sqlite3")
catalog.metadata.create_all(database.engine)          # a project runs its migrations instead
repository = Repository(database)                     # add_dependency("repository", repository)

repository.save([Product(name="Coffee", price=30), Product(name="Tea", price=20)])

page = repository.search(Products, Criteria.model_validate({
    "where": {"field": "price", "operator": ">", "value": 25},
    "order": [{"field": "price", "descending": True}],
    "pagination": {"limit": 10},
}))
print([product.name for product in page.items])        # ['Coffee']

How to, by topic

Every row links the step-by-step section of the guide — where every example runs as part of the test suite — and the page that explains it in depth.

Crons

A cron is one more caller of the buses, like an RPC method: a class registered on the registry of its bounded context, the buses it orchestrates injected by name.

from sincpro_framework.entrypoints.adapters.cron import Cron as _Cron
from sincpro_framework.entrypoints.adapters.cron import CronGateway, CronProcess, Crons, Tick


class CronDependencyContextType:
    cybersource: UseFramework


class Cron(_Cron, CronDependencyContextType):          # the bounded context's base cron
    pass


cron_payments = Crons[CronDependencyContextType]("cron-payments")
cron_payments.add_dependency("cybersource", cybersource)


@cron_payments.cron("0 2 * * *", timezone="America/La_Paz")
class ReconcileTransactions(Cron):
    def run(self, tick: Tick) -> None:
        self.cybersource(CommandReconcile(day=tick.scheduled_for.date()))


def build_crons() -> CronGateway:                       # module-level: runs in the child
    return CronGateway([cron_payments])


crons = CronProcess(build_crons).start()                # at startup; crons.stop() at shutdown

The whole of it, runnable: docs/cron/.

Migrations

The project writes each migration step; the framework orders every chain of every context and store into one timeline and moves the whole system along it.

from sincpro_framework.data_layer.migrations import ContextMigrations, Migrations, command_line
from sincpro_framework.data_layer.orm import AlembicEngine            # the [migrations] extra

billing_migrations = ContextMigrations("billing", Path("domains/billing/entrypoints/migrations"))
billing_migrations.store("main", AlembicEngine(billing_tables, database))

migrations = Migrations([common_migrations, billing_migrations])      # entrypoints/migrations.py
raise SystemExit(command_line(migrations))                           # status, upgrade, downgrade …

The whole of it, runnable: docs/migrations/.

Data analysis

The same question never sent to the database twice: a read is held, continued page by page, and narrowed without asking again. A utility, not an engine — the computing is pandas', polars' or DuckDB's.

from sincpro_framework.data_layer.data_analysis import QueryCache

cache = QueryCache(max_rows=2_000_000)
page = cache.fetch(repository, InvoiceLine, posted)                  # one read
more = cache.fetch(repository, InvoiceLine, posted, pages=3)         # only pages 2 and 3
sales = cache.fetch_all(repository, InvoiceLine, posted)             # complete
sales.narrow({"field": "journal", "operator": "=", "value": "SAL"})  # no read at all
polars.DataFrame(sales)                                              # or sales.to_parquet()

The whole of it, runnable: docs/data_analysis/.

Runtime use cases

A use case stored as source — its Commands, its Responses and the one handler that answers them — loaded onto the bus without a deploy.

from sincpro_framework.runtime.runtime_use_cases import BusRegistry, RuntimeUseCase

registry = BusRegistry(billing, store)                              # the code's bus + what is stored
registry.check(RuntimeUseCase(name="quote", source=source, version=2))          # refused before it is saved
store.save(RuntimeUseCase(name="quote", source=source, version=2))
registry.reload()                                                   # a new generation, swapped in whole
registry.execute("sincpro_runtime.billing.quote.CommandQuote", {"amount": 100})

The whole of it, runnable: docs/runtime_use_cases/.

Workflows (experimental)

The Commands a bus answers composed as JSON — what a node editor previews and an agent tries. The vocabulary may change.

from sincpro_framework.runtime.workflows import FileWorkflows, Workflows

workflows = Workflows(billing, FileWorkflows(Path("workflows")))    # one <name>.json each
run = workflows.run("bill_order", {"order_id": 1})                    # a trace of every step
workflows.draw("bill_order")                                          # the graph, as Mermaid

The whole of it, runnable: docs/workflows/.

📖 Documentation

This repository's documentation is generated with openwiki and lives in openwiki/ as browsable Markdown.

make docs-init   # first generation (once per repository)
make docs        # regenerate from code changes
make docs-view   # local explorer: node graph + Markdown reader

Requires Node >= 22 — nothing else. The Makefile runs openwiki through npx --yes openwiki@$(OPENWIKI_VERSION), so there is no global install and the version is pinned per run.

Provider, model and endpoint are already set in the Makefile (OPENWIKI_PROVIDER, OPENWIKI_MODEL_ID, OPENAI_COMPATIBLE_BASE_URL), so the only thing you have to supply is the API key:

export OPENAI_COMPATIBLE_API_KEY=<key>   # or store it in ~/.openwiki/.env (chmod 600)

The key is the one value that is never committed — not in the Makefile, not anywhere in the repository. make docs fails with a clear message when it is missing. Any of the other parameters can be overridden from the shell. The scope of the wiki is controlled in openwiki/INSTRUCTIONS.md, and what the agent is not allowed to read, in .openwikiignore.

Hand-written documentation stays in docs/, organised by layer (core, persistence, events, entrypoints, observability), and remains authoritative: the wiki references it, it does not replace it.

Migration note (4.0.0) — up to 3.x the framework shipped its own documentation generator (sincpro_framework.generate_documentation, with build_documentation() and the ai_context/ JSON files). It has been removed. Projects that used it replace their scripts/generate_doc.py with the make docs above.

Observability

The bus always instruments. Extras and env vars only decide where data goes.

Missing extra or missing DSN in conf → no-op, the bus still raises. Framework events are independent from the host: Odoo may also capture the same exception with its own release. That is intended.

What works without any extra install

Every framework(dto) call automatically tags all internal log lines with a trace_id and span_id. These are UUID-based identifiers — enough to correlate all logs produced by a single execution even without an external tracing backend.

You can also read them inside any Feature or ApplicationService:

class MyFeature(Feature):
    def execute(self, dto: MyDTO) -> MyResponse:
        trace_id = self.context.get("trace_id")  # always present
        ...

Installing with OpenTelemetry

pip install sincpro-framework[opentelemetry]

This installs:

• opentelemetry-api + opentelemetry-sdk
• opentelemetry-exporter-otlp-proto-grpc (primary)
• opentelemetry-exporter-otlp-proto-http (fallback)

When a use case fails

A failure is described by the handler that raised it and logged once, by the outermost bus — however many ApplicationServices or other bounded contexts' buses it crossed:

{
  "event": "SendInvoice failed: Fault: CUFD vencido",
  "dto": "CommandSendInvoice(nit=123, cuf='ABC')",
  "failed_in": "siat-soap-sdk",
  "handler": "SendInvoice",
  "layer": "feature",
  "chain": "CommandBillOrder → CommandSendInvoice",
  "error_type": "Fault",
  "error_at": "sincpro_siat_soap.infrastructure.soap_client:88 in SoapClient._send",
  "raised_at": "zeep.proxy:52 in OperationProxy.__call__",
  "correlation_id": "req-42",
  "exception": "Traceback ..."
}

• app_name is the bus that logged it — the outermost — and failed_in the bus whose handler raised: siat called from inside another context's Feature still says failed_in: "siat-soap-sdk".
• error_at is the last line of the handler's own package the error went through; raised_at is the line that raised, when it is somewhere else (a SOAP or HTTP library). The logger's own filename / func_name name whoever executed the DTO, never bus.py.
• The exception leaving the bus carries a note with the same facts, so whoever catches it — Odoo, a test, an error handler that maps it to the SDK's own exception — sees where it came from at the end of the traceback: [sincpro] CommandBillOrder → CommandSendInvoice: SendInvoice (siat-soap-sdk) failed at ….
• Every key set with context() is on every log line. Keep some out: UseFramework("siat", hide_in_logs=["TOKEN"]).
• A failure describes the DTO by its repr — on the error line, in GlitchTip, in the note. A field that must not leave the process is kept out the way pydantic keeps it out of a repr: password: SecretStr, or otp: str = Field(repr=False).
• An error handler that re-raises changes nothing above. One that answers leaves a warning — … (answered by an error handler) — because the caller never sees the error.
• An expected error, one passed to ignore_sentry_exceptions, is logged at info with no traceback: it is traffic, not a fault.
• A transport (JSON-RPC, gRPC, the background queue) logs only what no bus logged: process.was_reported(error).

Sentry / GlitchTip (errors)

Same silent contract as OTel. The bus always tries to report exceptions; if sentry-sdk is missing or the DSN in conf is unset, it is a no-op.

pip install sincpro-framework[sentry]
export SENTRY_PYTHON_DSN=https://KEY@glitchtip.sincpro.dev/1

Conf (sincpro_framework/conf/sincpro_framework_conf.yml) resolves sentry_dsn from SENTRY_PYTHON_DSN. If that env is set, app.observability.status.sentry is on:init, not off. The framework does not call sentry_sdk.init() and does not reuse the host client.

Each framework event uses an isolated Sentry Client whose release is the deployed artifact and its version — literally APP_RELEASE (sincpro_mcp_odoo:0.8.0), or {distribution}:{version} for an SDK (sincpro-payments-sdk:5.0.3). The bus is not part of the release: two buses of one deployment ship the same release and are told apart by the sincpro.instance tag. Framework-internal errors use sincpro-framework:<framework version>.

APP_RELEASE is the standard on every deployed service, so it answers first. Only when there is no artifact at all does the bus name stand in for it, so events stay separable per bounded context.

GlitchTip environment is TENANT (same value as the tenant tag) so events can be filtered by tenant in the UI.

Odoo may capture the same exception with Odoo's release. That second event is intended — two products, two releases, same traceback.

The bus reports before the error handler runs, once, from the handler that raised: an ApplicationService the error crosses sends nothing more. The event carries a sincpro.handler tag and a sincpro context with the DTO, the chain, error_at and the execution's context. A handler that swallows an unexpected exception still produces a GlitchTip event. Expected domain errors can be excluded per instance:

app = UseFramework("payment-cybersource")  # release auto-detected from the caller package
app.ignore_sentry_exceptions(ValidationError, InsufficientFunds)

Pass package="sincpro-payments-sdk" to UseFramework when the caller is not the library itself (tests, a thin adapter).

Observability is optional and must never break the bus. After build_root_bus() (or the first app(dto) call) every instance exposes a probe:

status = app.observability.status          # ObservabilityStatus
status.sentry.state                        # off | on | failed
status.otel.reason

• off — extra not installed or conf DSN missing (sdk_missing, dsn_missing)
• on — isolated client ready (init)
• failed — DSN present but client construction broke; the bus still runs. Logged as warning.

The instance logger emits: observability sentry=on:init otel=off:no_endpoint.

Do not send traces to GlitchTip (traces_sample_rate=0); Tempo stays on OTLP.

What OpenTelemetry adds

When OTel is installed, auto-adoption of outer spans happens transparently — any active span already in the OTel context (set by FastAPI OpenTelemetry middleware, a Celery task decorator, etc.) becomes the parent of all sincpro spans without any extra setup.

Configuring the OTLP exporter

Set the endpoint in your environment or the framework config file:

export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317

Or in sincpro_framework/conf/sincpro_framework_conf.yml:

otlp_endpoint: $ENV:OTEL_EXPORTER_OTLP_ENDPOINT

Nothing registers it by hand: a bus brings tracing up when it starts, from the endpoint the configuration names above. The provider is a process-level singleton — the first bus to start installs it, and a second bounded context in the same process finds it already there.

What each door is for, and how a transport span joins a bus's trace, is in the observability docs.

The with_trace() context manager

Use with_trace() when you need explicit control over the trace boundary — for example, to group multiple framework(dto) calls under a single trace, or to accept a trace from an upstream caller.

Fresh trace (useful in CLI scripts, workers, or test harnesses):

with framework.with_trace() as fw:
    result = fw(MyDTO(...), MyResponse)
    # all logs inside this block share the same trace_id/span_id

Propagate from upstream HTTP headers (W3C traceparent):

# headers = {"traceparent": "00-<trace_id>-<parent_id>-01", ...}
with framework.with_trace(carrier=request.headers) as fw:
    result = fw(ProcessOrderDTO(...), OrderResult)

Explicit IDs (re-use IDs from a previous system that doesn't speak W3C):

with framework.with_trace(trace_id="abc123", span_id="def456") as fw:
    result = fw(MyDTO(...), MyResponse)

Span attributes

Every span produced by the framework carries:

What only the use case knows — a NIT, a merchant, the code a provider answered — it puts on its own span with traces: declared with @traces.attributes(of(Command).nit, namespace="siat"), or by hand with traces.annotate({"siat.cuf": cuf}). High-cardinality values a metric refuses as labels belong there; see span attributes.

The observability API: automatic first, two doors when you need them

Nothing has to be called. Creating a UseFramework and building it is the whole setup:

• It auto-configures. The endpoint, the DSN, the release, the tenant and the sampling rate are read from the environment by the framework itself (OTEL_EXPORTER_OTLP_ENDPOINT, SENTRY_PYTHON_DSN, APP_RELEASE, TENANT, OTEL_TRACES_SAMPLER_ARG, OTEL_SERVICE_NAME, SINCPRO_FRAMEWORK_LOG_LEVEL). A service does not assign anything into the framework's settings: if the pod has the variable, the bus has it too.
• It auto-instruments. Every DTO gets a span named after it, with sincpro.layer and sincpro.instance; every unhandled exception reaches GlitchTip under the right release; every internal log line carries the active trace_id. Identity is resolved once — the calling library, or APP_RELEASE for a service.
• It degrades on its own. Extras missing, endpoint absent, collector down, a span processor that raises: all of it is a status you can read, never an exception the bus has to survive.

So a library embedded in Odoo calls none of this. UseFramework(...), and done.

Two doors exist for what the framework cannot know on its own:

# 1. Per bus — you already have it
framework.observability.identity   # sincpro-odoo-mcp:0.8.0:sales_mcp
framework.observability.status     # {"sentry": {...}, "otel": {...}}

# 2. The transport — only for a service that also serves HTTP
from sincpro_framework.observability import process

process.identity                     # sincpro-odoo-mcp:0.8.0 — no bus segment
process.status                       # on:installed | on:host | off:...
process.tracer("asgi")               # for OpenTelemetryMiddleware / HTTPXClientInstrumentor
process.bind_logger(access_log)      # that logger now stamps the request's trace_id
process.trace_ids()                  # the same ids, for a structlog processor
process.record_error(error, "asgi")  # a 500 that dies before any Feature runs

Everything else under sincpro_framework.observability is an implementation detail of those two. There is nothing to monkeypatch and no private module to import.

A bus never takes the global TracerProvider

Transport instrumentation asks OTel for the global tracer. If a bus answered there, every request would be exported as if it belonged to that bounded context — with three buses in one process, POST /mcp came out labelled as the first one that booted.

So the rule is: when nobody owns the global, the framework installs a process provider there — same endpoint, same tenant, identity without the bus segment. When a host (Odoo) already owns it, nothing is installed. Building any bus is what triggers this, which is why a service builds its buses before it serves the first request.

The result is one trace, several identities, and no shared Resource:

service.name=sincpro-odoo-mcp:0.8.0              POST /mcp                  (ASGI)
  └── service.name=...:0.8.0:common_mcp            CommandListAvailableTools  (bus)
  └── service.name=...:0.8.0:sales_mcp             CommandCreateQuote         (bus)
        └── service.name=sincpro-odoo-mcp:0.8.0      GET /odoo/registry       (httpx)
access log / uvicorn                              same trace_id, process logger

They share the trace_id because OTel propagates through contextvars — not because they share a provider. Filter the deployment by sincpro-odoo-mcp:0.8.0, a bounded context by its full service.name or by the sincpro.instance attribute.

What a service's boot looks like

from sincpro_framework.observability import process

def main() -> None:
    configure_process_logging(settings.log_level)   # your own stdlib/structlog routing

    # Eager: building a bus installs the process provider, which the ASGI middleware
    # needs before it opens its first span.
    for bus in ALL_BUSES:
        bus.build_root_bus()

    process.bind_logger(access_log)
    logger.info("otel proceso=%s:%s", process.status.state, process.status.reason)

    serve(middleware=[Middleware(OpenTelemetryMiddleware), Middleware(JsonAccessLog)])

What stays yours: which URLs to exclude, which library loggers to route, what not to log, the origin guard, the routes. The framework gives the sink, the identity, the tracer per layer and the ids for the logs; the service wires its own transport.

Embedding sincpro inside another instrumented service (Odoo, FastAPI, etc.)

Every bus gets its own TracerProvider, whose service.name is artifact:version:bus — e.g. sincpro_mcp_odoo:0.8.0:common_mcp. The root span created by with_trace() and the DTO spans under it all come from that provider, so one trace reports one identity. If the host application (Odoo, FastAPI, Celery) already registered the global OTel provider, sincpro leaves it untouched and keeps its own provider internal.

The trace relationship is still preserved: OTel propagates the active parent span via contextvars (process-wide), so sincpro spans are automatically children of whatever span the host has active at call time. In Tempo/Jaeger the full tree is visible and filterable:

service.name=odoo                            →  GET /web/dataset/call_kw    (Odoo HTTP span)
service.name=sincpro_mcp_odoo:0.8.0:sales    →    └── CreateOrderDTO         (application_service)
service.name=sincpro_mcp_odoo:0.8.0:sales    →         └── ValidateStockDTO  (feature)

Sampling is respected across the boundary: sincpro uses ParentBased, so a decision already taken upstream — by the host, or by an incoming traceparent — always wins and a sampled request is never truncated halfway through.

How much of the traffic this bus starts on its own is recorded comes from OTel's standard variable:

export OTEL_TRACES_SAMPLER_ARG=0.1   # record 10% of the traces born in this bus

1.0 (the default) records everything, 0.0 records nothing. An unusable value falls back to 1.0 with an info log — settings is built at import time, so a typo in one deployment variable must not make the framework unimportable.

Configuration or settings

A project's configuration is one document — a YAML file plus the environment — and a settings class is any SincproConfig. build_config_obj(Shape, file, path) builds one from the document; the full guide is docs/core/settings.md.

from sincpro_framework.sincpro_conf import SincproConfig, build_config_obj


class PostgresConf(SincproConfig):
    host: str = "localhost"
    port: int = 5432
    user: str = "my_user"


class MyConfig(SincproConfig):
    log_level: str = "DEBUG"
    token: str = "default_my_token"
    postgresql: PostgresConf = PostgresConf()


config = build_config_obj(MyConfig, "/path/to/your/config.yml", "my_project")

my_project:
  log_level: "INFO"
  token: "$ENV:MY_SECRET_TOKEN"
  postgresql:
    host: localhost
    port: 12345

path names the project's section of the file (optional: without it the whole file is the section) and may be dotted to reach a section inside it. A nested class is read from the section of its name; one with a default keeps it when its section is absent. The object is assignable unless its class sets model_config = ConfigDict(frozen=True).

Environment variables

$ENV:NAME reads the variable. When it is unset, or holds what the field cannot accept, the field takes its default and an info line is logged — the settings are built at import time, and a typo in one deployment variable never makes the process fail to start:

class ApiConfig(SincproConfig):
    api_key: str = "dev_default_key"

# config.yml:  api_key: "$ENV:API_KEY"
# API_KEY unset → info: "Environment variable [API_KEY] is not set for field [api_key].
#                        Using default value: dev_default_key"

A root class that declares env_prefix: ClassVar[str] = "PAYMENTS" also reads variables by path (PAYMENTS__QR__TIMEOUT sets qr.timeout); without it, nothing is read by path.

A project with bounded contexts

The shared settings, a class per context inheriting them, and the global holding every context:

class SharedSettings(SincproConfig):
    environment: Environment = Environment.TEST


class QRSettings(SharedSettings):
    timeout: float = 10.0


class PaymentsSettings(SharedSettings):
    qr: QRSettings


settings = build_config_obj(PaymentsSettings, FILE, "sincpro_payments_sdk")   # the global
qr_bus.add_dependency("settings", settings.qr)                                  # a context's own
qr = build_config_obj(QRSettings, FILE, "sincpro_payments_sdk.qr")             # a shape alone

A section inherits a shared field it does not set (qr.environment) from the nearest section above, and overrides it by setting it. Everything required and missing is one validation error naming each path. Secret[str] masks a value everywhere it is printed. describe_settings(settings) lists each value with where it came from — file, section inherited, variable or default. sincpro_framework.runtime.testing.settings_scope_violations reports a context reading another context's section, for a team that wants the rule. A shared class inheriting FrameworkSettings hands the framework its log, OTLP, Sentry and release settings — no framework_settings.x = … lines.

📦 Variables

The framework use a default setting file where live in the module folder inside of sincpro_framework/conf/sincpro_framework_conf.yml where you can define some behavior currently we support the following settings:

• sincpro_framework_log_level: Log level for the framework logger. Default: DEBUG.
• otlp_endpoint: OTLP exporter endpoint for distributed tracing. Resolved from OTEL_EXPORTER_OTLP_ENDPOINT env var. Default: null (tracing disabled). Requires sincpro-framework[opentelemetry].
• sincpro_framework_log_level: INFO or DEBUG. Resolved from SINCPRO_FRAMEWORK_LOG_LEVEL. Default: DEBUG. A service sets its own level through the environment, or through its own settings by inheriting FrameworkSettings (docs/core/settings.md), which configures the log again when the level is among what it sets.
• otlp_traces_sample_rate: share of new traces to record, 0.0-1.0. Resolved from OTEL_TRACES_SAMPLER_ARG. Default: 1.0. An upstream sampling decision always wins over this ratio.
• app_release: deployed artifact and version, artifact:version. Resolved from APP_RELEASE — the standard on every Sincpro service. Feeds both the GlitchTip release and the OTel service.name.
• otel_service_name: names the artifact when APP_RELEASE carries only a version. Resolved from OTEL_SERVICE_NAME.
• tenant: GlitchTip environment and the tenant tag. Resolved from TENANT.
• otel_metrics_exporter / otel_traces_exporter / otel_sdk_disabled: OpenTelemetry's own switches, from OTEL_METRICS_EXPORTER (otlp, prometheus, none), OTEL_TRACES_EXPORTER (none builds no provider of the framework's) and OTEL_SDK_DISABLED.
• metrics_backend: auto, prometheus, otel or off. Resolved from SINCPRO_METRICS_BACKEND — kept as an alias; set to anything but auto, it wins over OTEL_METRICS_EXPORTER. Default: auto — OpenTelemetry when a meter provider or an OTLP endpoint is configured, otherwise nothing. prometheus requires sincpro-framework[prometheus] and serves /metrics (docs/observability/metrics.md).
• sentry_dsn: GlitchTip/Sentry DSN. Resolved from SENTRY_PYTHON_DSN. Default: null (error reporting disabled). Requires sentry-sdk (or sincpro-framework[sentry]). The framework uses an isolated client with release=APP_RELEASE and never calls sentry_sdk.init(). Odoo may capture the same error separately. Use UseFramework.ignore_sentry_exceptions(...) for expected errors.

Override the config file using another

export SINCPRO_FRAMEWORK_CONFIG_FILE = /path/to/your/config.yml

🧪 Tests & coverage

The Makefile is the single entry point — CI calls the same targets you run locally.

make test                 # unit suites + coverage report in the terminal + coverage.xml
make test-realworld       # the ledger cases over a populated database (docs/persistence/testing.md)
make test-stress          # the same at 25 000 entries, timed; SINCPRO_REALWORLD_ENTRIES and DATABASE_URL apply
make test-coverage        # the above + HTML report in htmlcov/
make test-coverage-open   # the above + opens htmlcov/index.html in the browser
make test_one t=tests/aio/test_async_bus.py   # a single file/test, verbose, no coverage
make clean-coverage       # remove .coverage, coverage.xml and htmlcov/

make test fails when total coverage drops below COVERAGE_MIN (65% by default), so a regression breaks the build instead of passing silently. Raise the floor as coverage grows, or override it for a single run:

make test COVERAGE_MIN=80

Coverage is measured over sincpro_framework with branch coverage enabled; the settings live in [tool.coverage.*] in pyproject.toml. All generated artifacts (.coverage, coverage.xml, htmlcov/) are git-ignored.

🧵 Python 3.14 & Free-Threading Notes

Regular Python 3.14 (GIL build): fully supported today, no breaking changes. Verified end-to-end on 3.14.7 — every dependency (core and all extras: opentelemetry, sentry, mcp, rpc) installs from prebuilt wheels with no source builds, pyright reports 0 issues, and the full test suite passes (215/215). CI (.github/workflows/02-check_code.yaml) now runs "3.12", "3.13", "3.14".

Free-threaded Python (python3.14t, PEP 703/779): not implemented or targeted yet — this section documents current findings only, so the challenges are visible before anyone attempts it. Everything below was verified hands-on (3.14.7 vs. 3.14.7t), not inferred:

• Dependency gap, not a code gap. dependency-injector (core, required — see ioc.py) and grpcio (transitive dep of the optional opentelemetry-exporter-otlp-proto-grpc) have no free-threaded wheel yet. Importing either on python3.14t makes CPython print RuntimeWarning: The global interpreter lock (GIL) has been enabled to load module '...', which has not declared that it can run safely without the GIL and silently re-enable the GIL for the rest of the process. pydantic-core (the framework's other compiled dependency) is fine — it already ships a cp314t wheel.
• contextvars propagation to a bare ThreadPoolExecutor.submit differs by build. Verified with a minimal repro outside this framework: on 3.12.11 and 3.14.7 (GIL builds) a bare executor.submit(fn) loses the caller's contextvars.Context, same as always — this is the bug ThreadContextBus/thread_context() exists to fix. On 3.14.7t (free-threaded), it's already propagated with no thread_context() involved. This doesn't make thread_context() wrong or unnecessary — most users run a GIL build, and code shouldn't silently depend on a free-threaded-only behavior — but it does mean tests/context/test_thread_context_bus.py::test_bare_submit_loses_context_in_new_thread and ::test_fan_out_without_thread_context_loses_context_for_every_worker encode a GIL-build-specific assumption and would legitimately fail if that suite is ever run on a free-threaded interpreter.
• asyncio's free-threading support only matured in 3.14. Relevant to get_async_bus() / AsyncBus, which is built on asyncio.to_thread: prefer 3.14+ over 3.13t for that path.
• Registries built once at startup (feature_registry, app_service_registry, dynamic_dep_registry, the interceptor chains) are safe as long as nothing mutates them concurrently with in-flight executions — true today, not enforced. See the docstring on UseFramework.add_dependency.

Every spot above is also marked in the source with a # PYTHON 3.14 FREE-THREADING: comment — grep -rn "PYTHON 3.14 FREE-THREADING" sincpro_framework/ tests/ finds all of them.