rosenix 0.1.0

A modular, explicit, non-magical framework for building AI-powered software systems.

Rosenix

A modular, explicit, non-magical framework for building AI-powered software systems: assistants, agents, multi-agent systems, AI APIs, workflows, and more — on one consistent architecture.

Philosophy

• Nothing is magical, nothing is hidden. No implicit prompt templates, no hidden chains. Every behavior is a plain, typed, inspectable object.
• Composition over inheritance. Providers are typing.Protocols, not base classes — any object with the right shape works, no framework coupling required.
• Everything is a plugin. LLMs, embeddings, vector stores, tools, and memory backends all register through the same plugin system, including third-party packages via entry_points.

Architecture (4 layers)

Layer 4  kits/       Agent, Workflow, Tool decorator      (opinionated)
Layer 3  runtime/     event bus, task executor, tracing    (execution)
Layer 2  protocols/   LLMProvider, EmbeddingProvider, ...   (contracts)
Layer 1  kernel/      DI container, plugin registry, config (foundation)

Each layer only depends on the layers below it. Layer 1 has zero AI-specific code — it's a generic application kernel, on purpose, so the architecture is proven before AI complexity is added.

Quickstart

pip install -e ".[dev,openai]"

import asyncio
from rosenix.kernel.container import Container
from rosenix.providers.openai_provider import OpenAIProvider
from rosenix.kits.agent import Agent

async def main():
    container = Container()
    container.register(OpenAIProvider, lambda: OpenAIProvider(model="gpt-4o-mini"))

    agent = Agent(llm=container.resolve(OpenAIProvider), name="assistant")
    reply = await agent.run("Say hello in one sentence.")
    print(reply)

asyncio.run(main())

See examples/basic_agent.py for a full runnable example.

Development

pip install -e ".[dev,openai]"
pytest --cov=rosenix --cov-report=term-missing
mypy src/rosenix
ruff check src/ tests/ examples/

See CONTRIBUTING.md for local setup details and current test-coverage gaps that are good first contributions.

Project status

Pre-alpha (0.x). The architecture is stable; test coverage on kits/ and providers/ is still being built out — see ROADMAP.md. Not yet recommended for production use.