aoa-action-machine

The core of AOA — a framework where a business operation is described as an executable contract. Every operation is an Action class with a typed input, output, and a straight chain of steps; the machine reads this description and executes it literally.

This README is a quick start: from installation to an operation published over HTTP and as an AI-agent tool. It deliberately does not cover everything — the full picture is in the tutorial, with links to individual topics placed along the way.

Installation

pip install aoa-action-machine

Optional extensions are installed as needed:

pip install "aoa-action-machine[fastapi]"   # HTTP API
pip install "aoa-action-machine[mcp]"        # tools for AI agents
pip install "aoa-action-machine[postgres]"   # asyncpg connections
pip install aoa-ocel                  # OCEL 2.0 event log

Your first action

An operation is an atomic business operation. Let us assemble the simplest one: it takes a name and returns a greeting. Three declarations are mandatory: the domain (@meta), access (@check_roles), and a single exit point (@summary_aspect).

from aoa.action_machine.auth import GuestRole
from aoa.action_machine.context import Context
from aoa.action_machine.domain.base_domain import BaseDomain
from aoa.action_machine.intents.aspects import summary_aspect
from aoa.action_machine.intents.check_roles import check_roles
from aoa.action_machine.intents.meta import meta
from aoa.action_machine.model import BaseAction, BaseParams, BaseResult
from aoa.action_machine.runtime.action_product_machine import ActionProductMachine
from pydantic import Field


class GreetingDomain(BaseDomain):
    name = "greeting"
    description = "Greeting domain"


class GreetParams(BaseParams):
    name: str = Field(description="Recipient name")


class GreetResult(BaseResult):
    message: str = Field(description="The assembled greeting")


@meta(description="Greet by name", domain=GreetingDomain)
@check_roles(GuestRole)            # GuestRole — the operation is open to everyone (declared explicitly)
class GreetAction(BaseAction[GreetParams, GreetResult]):

    @summary_aspect("Assemble the greeting")
    async def greet_summary(self, params, state, box, connections):
        return GreetResult(message=f"Hello, {params.name}!")

BaseParams and BaseResult are Pydantic models; field descriptions go into the external schema. Inside the operation there is no state — everything arrives through params.

Running it

The single entry point is ActionProductMachine. It reads the declarations and guides the operation along the pipeline.

import asyncio

async def main() -> None:
    machine = ActionProductMachine()
    result = await machine.run(Context(), GreetAction(), GreetParams(name="Alice"))
    print(result.message)   # Hello, Alice!

asyncio.run(main())

Context() is the call environment (user, roles, metadata); here it is empty.

Several steps and the state contract

Usually an operation consists of several steps. An intermediate step — @regular_aspect — returns a dict that becomes the new state. A checker @result_* declares the contract: what must appear in state after the step.

from aoa.action_machine.intents.aspects import regular_aspect
from aoa.action_machine.intents.checkers import result_string

@meta(description="Greet by name", domain=GreetingDomain)
@check_roles(GuestRole)
class GreetAction(BaseAction[GreetParams, GreetResult]):

    @regular_aspect("Normalise the name")
    @result_string("name", required=True, min_length=1)
    async def normalise_aspect(self, params, state, box, connections):
        return {"name": params.name.strip().title()}

    @summary_aspect("Assemble the greeting")
    async def greet_summary(self, params, state, box, connections):
        return GreetResult(message=f"Hello, {state['name']}!")

Aspects run strictly top to bottom. If a step does not fulfil the checker's contract, the machine stops the pipeline at its boundary — the next step does not run. In detail — in the tutorial: Action and pipeline and State: an x-ray of the operation.

Dependencies and logs

Everything external is declared by the operation in its header via @depends and obtained through box.resolve(...); the machine will not hand out an undeclared dependency. Business events are written through box, not print:

from aoa.action_machine.intents.depends import depends
from aoa.action_machine.logging import Channel

@meta(description="Greet by name", domain=GreetingDomain)
@check_roles(GuestRole)
@depends(GreeterService)
class GreetAction(BaseAction[GreetParams, GreetResult]):

    @summary_aspect("Assemble the greeting")
    async def greet_summary(self, params, state, box, connections):
        greeter = await box.resolve(GreeterService)
        await box.info(Channel.business, "greeting name={%var.name|cyan}", name=params.name)
        return GreetResult(message=greeter.greet(params.name))

Where to deliver events — console, queue, Telegram — is decided by the machine's logger, not by the operation code.

Service: FastAPI and MCP

The operation knows nothing about transport, so the same GreetAction can be exposed both over HTTP and as an AI-agent tool — by adding an adapter. The business code does not change.

HTTP (FastAPI):

from aoa.action_machine.adapters.fastapi import FastApiAdapter
from aoa.action_machine.auth import NoAuthCoordinator

machine = ActionProductMachine()

app = (
    FastApiAdapter(machine=machine, auth_coordinator=NoAuthCoordinator(), title="Greetings API")
    .post("/greet", GreetAction, tags=["greetings"])
    .build()
)
# uvicorn app:app

FastApiAdapter publishes the operation as a REST endpoint with a ready-made OpenAPI schema derived from Params/Result.

MCP (AI agents):

from aoa.action_machine.adapters.mcp import McpAdapter
from aoa.action_machine.auth import NoAuthCoordinator

server = (
    McpAdapter(machine=machine, auth_coordinator=NoAuthCoordinator(), server_name="Greetings MCP")
    .tool("greetings.greet", GreetAction)
    .build()
)
# python mcp_server.py

The detailed chapters: Step 13 — FastAPI adapter · Step 14 — MCP adapter.

What's next

That was the quick start. The full tutorial is on the Contents page: it leads from the first operation to the service layer and the domain model, with examples and review questions.

Useful by topic:

• Action and pipeline · State: an x-ray of the operation — the core of the model.
• Sagas and compensations, error handling, context, cache, plugins — the «Business logic» part (in progress).
• Action, aspect, or resource · Migrating a legacy codebase to AOA — practical decisions.
• Glossary · Comparison with other frameworks · FAQ.

License

MIT