interaxions 0.0.3

A library for agent and environment protocol interactions

Interaxions

A modern, extensible framework for orchestrating AI agents and environments on Kubernetes/Argo Workflows, inspired by HuggingFace Transformers.

✨ Features

• 🎯 Job-Based Configuration - Unified Job schema for complete workflow definition
• 🚀 Dynamic Loading - Load components from built-in, local, or remote Git repositories
• 🔄 Unified API - All Auto* classes use consistent from_repo() interface
• 📦 Three-Layer Architecture - Scaffolds, Environments, and Workflows
• 🏷️ Version Control - Support for Git tags, branches, and commits
• 🔒 Multi-Process Safe - File locks for concurrent access
• 💾 Smart Caching - Three-level cache system for optimal performance
• 🌐 Flexible Sources - GitHub, GitLab, Gitea, or any Git service via IX_ENDPOINT
• ✅ Comprehensive Testing - 99 tests with 72% coverage

🚀 Quick Start

Installation

# Basic installation
pip install interaxions

# With optional dependencies
pip install interaxions[hf]    # HuggingFace datasets
pip install interaxions[oss]   # OSS storage support

# For development
pip install -e ".[dev]"

Basic Usage (Job-Based API)

from interaxions import AutoWorkflow
from interaxions.schemas import Job, Scaffold, Environment, Workflow, Runtime, LiteLLMModel

# Define a complete job configuration
job = Job(
    name="fix-django-bug",
    description="Fix Django bug using SWE-agent",
    tags=["swe-bench", "django"],
    labels={"priority": "high", "team": "research"},
    
    # Model configuration
    model=LiteLLMModel(
        type="litellm",
        provider="openai",
        model="gpt-4",
        base_url="https://api.openai.com/v1",
        api_key="your-api-key",
    ),
    
    # Scaffold (agent) configuration
    scaffold=Scaffold(
        repo_name_or_path="swe-agent",
        params={"max_iterations": 10},
    ),
    
    # Environment configuration
    environment=Environment(
        repo_name_or_path="swe-bench",
        environment_id="astropy__astropy-12907",
        source="hf",
        params={
            "dataset": "princeton-nlp/SWE-bench",
            "split": "test",
        },
    ),
    
    # Workflow configuration
    workflow=Workflow(
        repo_name_or_path="rollout-and-verify",
        params={},
    ),
    
    # Runtime configuration
    runtime=Runtime(
        namespace="experiments",
        service_account="argo-workflow",
        ttl_seconds_after_finished=3600,
    ),
)

# Create and submit workflow
workflow_template = AutoWorkflow.from_repo(job.workflow.repo_name_or_path)
workflow = workflow_template.create_workflow(job)
workflow.create()  # Submit to Argo

Quick API (One-Step Loading)

from interaxions import AutoScaffold, AutoEnvironment, AutoWorkflow

# Load scaffold
scaffold = AutoScaffold.from_repo("swe-agent")

# Load environment (unified API)
env = AutoEnvironment.from_repo(
    repo_name_or_path="swe-bench",
    environment_id="astropy__astropy-12907",
    source="hf",
    dataset="princeton-nlp/SWE-bench",
    split="test",
)

# Load workflow
workflow_template = AutoWorkflow.from_repo("rollout-and-verify")

📚 Core Concepts

1. Job - Unified Configuration

Job is the central schema that encapsulates all information needed to run a workflow:

from interaxions.schemas import Job

job = Job(
    # Metadata
    name="my-job",
    description="Job description",
    tags=["tag1", "tag2"],
    labels={"key": "value"},
    
    # Components (all use from_repo pattern)
    model=...,        # LLM configuration
    scaffold=...,     # Agent/scaffold configuration
    environment=...,  # Environment/data configuration  
    workflow=...,     # Workflow orchestration
    runtime=...,      # Kubernetes/Argo settings
)

2. Three-Layer Architecture

Scaffolds (formerly Agents)

• High-level orchestration logic
• Can manage single or multiple agents internally
• Example: swe-agent

Environments

• Test environments and evaluation datasets
• Support HuggingFace, OSS, and custom sources
• Example: swe-bench

Workflows

• Define execution order and dependencies
• Generate Argo Workflows
• Example: rollout-and-verify

3. Dynamic Loading

All components use the unified from_repo() pattern:

# Built-in (by name, no path separators)
component = Auto*.from_repo("component-name")
component = Auto*.from_repo("swe-agent")  # Uses default config

# Local path (contains /, ., or ~)
component = Auto*.from_repo("./my-component")
component = Auto*.from_repo("/absolute/path/to/component")

# Remote repository (Github/Gitlab)
component = Auto*.from_repo("username/repo-name")

# With specific version
component = Auto*.from_repo("username/repo", revision="v1.0.0")

Loading Logic:

1. If name contains no path separators (/, ., ~) → Try as built-in first
2. If built-in not found or path provided → Load from filesystem/remote
3. Built-in modules use default config; external modules require config.yaml
4. Remote repositories (format: username/repo) use IX_ENDPOINT (defaults to GitHub)

🎨 Loading Sources

Built-in Components

Built-in components are Python packages in interaxions/scaffolds/, interaxions/environments/, etc.

Two ways to load:

# Method 1: Direct import (for advanced customization)
from interaxions.scaffolds.swe_agent import SWEAgent, SWEAgentConfig

config = SWEAgentConfig(max_iterations=20)
scaffold = SWEAgent(config=config)

# Method 2: Unified interface (uses default config)
from interaxions import AutoScaffold, AutoWorkflow

scaffold = AutoScaffold.from_repo("swe-agent")  # name only, no paths
workflow = AutoWorkflow.from_repo("rollout-and-verify")

Characteristics:

• ✅ No config.yaml needed (config in Python code)
• ✅ Use simple name: "swe-agent", "swe-bench", "rollout-and-verify"
• ✅ Automatically use default configuration
• ❌ Cannot use paths: "./swe-agent" will load from filesystem

External Components

External components are loaded from filesystem paths or remote repositories.

from interaxions import AutoScaffold

# From local directory
scaffold = AutoScaffold.from_repo("./my-custom-scaffold")
scaffold = AutoScaffold.from_repo("/absolute/path/to/scaffold")

# From remote Git repository (GitHub by default)
scaffold = AutoScaffold.from_repo("username/my-scaffold")
scaffold = AutoScaffold.from_repo("username/my-scaffold", revision="v1.0.0")

# From GitLab or other Git services (set IX_ENDPOINT)
# export IX_ENDPOINT=https://gitlab.com
scaffold = AutoScaffold.from_repo("username/my-scaffold")

Characteristics:

• ✅ Must have config.yaml file
• ✅ Must have specific filename: scaffold.py, env.py, or workflow.py
• ✅ Can specify version via revision parameter
• ✅ Supports Git tags, branches, commits
• ✅ Uses IX_ENDPOINT env var to specify Git service (defaults to GitHub)
• 📦 Cached in ~/.interaxions/hub/

Built-in vs External Comparison

Feature	Built-in	External
Loading	from_repo("swe-agent")	from_repo("./my-scaffold") or from_repo("user/repo")
Location	interaxions/scaffolds/	Filesystem / Remote (GitHub/GitLab/etc.)
Config	Python code	config.yaml required
Versioning	Package version	Git revision (tags, branches, commits)
Git Service	N/A	Configurable via IX_ENDPOINT (default: GitHub)
Customization	Direct import + custom config	Via config.yaml
Use Case	Production-ready, battle-tested	Custom experiments, research

Environment Loading (Unified API)

from interaxions import AutoEnvironment

# From HuggingFace
env = AutoEnvironment.from_repo(
    repo_name_or_path="swe-bench",
    environment_id="astropy__astropy-12907",
    source="hf",
    dataset="princeton-nlp/SWE-bench",
    split="test",
)

# From OSS (optional dependency)
env = AutoEnvironment.from_repo(
    repo_name_or_path="swe-bench",
    environment_id="astropy__astropy-12907",
    source="oss",
    dataset="swe-bench-data",
    split="test",
    oss_region="cn-hangzhou",
    oss_endpoint="oss-cn-hangzhou.aliyuncs.com",
    oss_access_key_id="your-key-id",
    oss_access_key_secret="your-secret",
)

Batch Loading (Factory Pattern)

from interaxions import AutoEnvironmentFactory

# Load factory once
factory = AutoEnvironmentFactory.from_repo("swe-bench")

# Create multiple environments efficiently
env1 = factory.get_from_hf("astropy__astropy-12907", "dataset", "test")
env2 = factory.get_from_hf("django__django-11039", "dataset", "test")
env3 = factory.get_from_hf("sympy__sympy-18199", "dataset", "test")

🔧 Environment Variables (Optional)

All environment variables have sensible defaults and are optional:

Variable	Description	Default
IX_HOME	Base directory for Interaxions data	~/.interaxions
IX_HUB_CACHE	Cache directory for hub modules	~/.interaxions/hub
IX_OFFLINE	Enable offline mode (no network)	false
IX_ENDPOINT	Git service endpoint for remote repos	https://github.com

Using Custom Git Services

By default, remote repositories use GitHub. To use GitLab, Gitea, or enterprise Git:

# GitLab
export IX_ENDPOINT=https://gitlab.com
scaffold = AutoScaffold.from_repo("username/my-scaffold")
# → Clones from https://gitlab.com/username/my-scaffold.git

# Enterprise GitLab
export IX_ENDPOINT=https://git.company.com
scaffold = AutoScaffold.from_repo("team/project")
# → Clones from https://git.company.com/team/project.git

# Gitea
export IX_ENDPOINT=https://gitea.io

Note: The system automatically appends .git to repository paths.

📦 Creating Custom Components

Custom components are external modules loaded via filesystem paths or remote repositories.

See Repository Standards for detailed requirements.

Minimum Requirements

Scaffold Repository:

my-scaffold/
├── config.yaml           # Required: type: my-scaffold, templates: {...}
├── scaffold.py           # Required: Class inheriting from BaseScaffold
└── templates/            # Optional: Jinja2 templates
    └── main.j2

Example config.yaml:

type: my-scaffold
image: my-scaffold:latest
templates:
  main: templates/main.j2
  sidecar: templates/sidecar.j2

Environment Repository:

my-environment/
├── config.yaml           # Required: type: my-environment, templates: {...}
├── env.py                # Required: Factory inheriting from BaseEnvironmentFactory
└── templates/            # Optional: Jinja2 templates
    └── verify.j2

Example config.yaml:

type: my-environment
templates:
  evaluation: templates/evaluation.j2
  setup: templates/setup.j2

Workflow Repository:

my-workflow/
├── config.yaml           # Required: type: my-workflow, templates: {...}
├── workflow.py           # Required: Class inheriting from BaseWorkflow
└── templates/            # Optional: Jinja2 templates
    └── main.j2

Example config.yaml:

type: my-workflow
templates:
  main: templates/main.j2
  verify: templates/verify.j2

Key Points:

• ✅ config.yaml is required for external components
• ✅ Use specific filenames: scaffold.py, env.py, workflow.py
• ✅ Classes must inherit from base classes: BaseScaffold, BaseEnvironmentFactory, BaseWorkflow
• ✅ Implement required methods: create_task() for scaffolds/environments, create_workflow() for workflows
• ✅ templates/ directory is optional but recommended for all component types
• ✅ Templates in config.yaml are loaded as strings (e.g., templates/main.j2)
• ✅ Built-in components don't need config.yaml (config in Python code)

🧪 Testing

# Run all tests (99 passed, 4 skipped)
pytest

# Run specific test categories
pytest -m unit          # Unit tests
pytest -m integration   # Integration tests
pytest -m e2e           # End-to-end tests

# With coverage (currently 72%)
pytest --cov=interaxions --cov-report=html

# View coverage report
open htmlcov/index.html

Test Statistics:

• ✅ 99 tests passing
• ⏭️ 4 tests skipped (3 OSS optional dependency, 1 needs mock environment)
• 📊 72% code coverage
• ⚡ ~4s total runtime

See tests/README.md for detailed testing documentation.

📁 Project Structure

interaxions/
├── scaffolds/          # Agent scaffold implementations
│   ├── base_scaffold.py
│   └── swe_agent/
├── environments/       # Environment implementations
│   ├── base_environment.py
│   └── swe_bench/
├── workflows/          # Workflow implementations
│   ├── base_workflow.py
│   └── rollout_and_verify/
├── schemas/            # Pydantic schemas (Job, Scaffold, etc.)
│   ├── job.py
│   └── models.py
└── hub/                # Dynamic loading system
    ├── auto.py         # Auto* classes
    ├── hub_manager.py  # Repository management
    └── constants.py    # Configuration

tests/                  # Comprehensive test suite (99 passed, 72% coverage)
├── unit/               # Unit tests (schemas, models)
├── integration/        # Integration tests (auto loading, factories)
├── e2e/                # End-to-end tests (full pipeline)
├── fixtures/           # Test data and mock repositories
│   ├── mock_repos/     # Mock scaffold/workflow repos for testing
│   └── sample_data.py  # Sample data generators
└── conftest.py         # Shared fixtures and test configuration

examples/               # Usage examples
└── quickstart.py       # Complete tutorial

🔄 Development Workflow

# Clone repository
git clone https://github.com/Hambaobao/interaxions.git
cd interaxions

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests (should see 99 passed, 4 skipped)
pytest

# Run specific test categories
pytest -m unit          # Fast unit tests
pytest -m integration   # Integration tests
pytest -m e2e           # End-to-end tests

# Check coverage
pytest --cov=interaxions --cov-report=term

# Run examples
python examples/quickstart.py

# Build package
python -m build

# Check package
twine check dist/*

📖 Documentation

• Repository Standards - Complete guide for creating custom components
• Testing Guide - Comprehensive testing documentation
• Examples - Example implementations and tutorials

🤝 Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass: pytest -m unit
5. Submit a pull request

📄 License

MIT License - see LICENSE for details

🙏 Acknowledgments

• Inspired by HuggingFace Transformers
• Built on Hera for Argo Workflows
• Powered by Pydantic for data validation

🔗 Links

• Homepage: https://github.com/Hambaobao/interaxions
• Issues: https://github.com/Hambaobao/interaxions/issues
• PyPI: Coming soon

---

Made with ❤️ for the AI agent research community