Agent Workflows

YAML-driven workflow engine for AI agents.

Installation

# Install globally with uvx (recommended)
uvx siili-agent-workflows --help

# Or install with pip
pip install siili-agent-workflows

# Or install locally for development
cd packages/agent-workflows
uv sync

Quick Start

# Run a workflow by name
siili-agent-workflows my-workflow --param value

# List all available workflows
siili-agent-workflows --list

# Validate a workflow without running
siili-agent-workflows my-workflow --validate

# Show help
siili-agent-workflows --help

Workflow Discovery

Workflows are discovered from multiple locations (in order):

1. Current directory: ./workflow-name.yml, ./workflow-name.yaml, etc.
2. Local workflows dir: ./.agent_workflows/workflow-name.yml
3. Global config: ~/.config/siili-agent-workflows/workflow-name.yml
4. Bundled: Workflows shipped with the package

Supported extensions: .yml, .yaml, .agent-workflow.yml, .agent-workflow.yaml

Environment Variables

Environment files are loaded in order (later overrides earlier):

1. ~/.config/siili-agent-workflows/.env (global)
2. ./.env (current directory)
3. --env-file <path> (explicit)

Workflow Format

Create my-workflow.yml:

name: my-workflow
env:
  API_KEY: ${{ env.MY_API_KEY }}

vars:
  project_name: my-project

jobs:
  setup:
    steps:
      - uses: terminal/run
        with:
          command: npm install

  generate:
    needs: setup
    steps:
      - uses: agents/claude_code
        with:
          prompt: "Create a README for ${{ vars.project_name }}"
          cwd: "."

  test:
    needs: generate
    steps:
      - uses: terminal/run
        with:
          command: npm test

CLI Options

You can also pass variables as positional args:

siili-agent-workflows apply-template --template agent --path ./my-agent

Built-in Plugins

Terminal

# Run shell commands
- uses: terminal/run
  with:
    command: "echo hello"
    shell: true              # optional, default true
    continue_on_error: false # optional
    timeout: 30              # optional, seconds

# Run scripts
- uses: terminal/script
  with:
    path: "./scripts/setup.sh"
    args: ["--verbose"]      # optional
    interpreter: "/bin/bash" # optional
    cwd: "."                 # optional

Git

# Clone/download from repository
- uses: git/download
  with:
    repo: "https://github.com/user/repo.git"
    ref: "main"              # optional branch/tag
    files:                   # optional, specific files/dirs
      - "templates/agent"
    dest: "./output"         # destination directory

# Push changes
- uses: git/push
  with:
    message: "chore: automated update"
    branch: "main"           # optional

Agents

# Claude Code agent (requires ANTHROPIC_API_KEY)
- uses: agents/claude_code
  with:
    prompt: "Implement feature X"
    cwd: "."                 # working directory

# General LLM agent (returns text response only)
- uses: agents/general
  with:
    prompt: "Analyze this code"
    system_prompt: "You are a code reviewer"  # optional
    output_var: analysis     # variable to store response

# Structured JSON response
- uses: agents/structured
  with:
    prompt: "Extract topics from this text: ..."
    schema:
      type: object
      properties:
        topics:
          type: array
          items:
            type: string
        summary:
          type: string
      required:
        - topics
        - summary
    output_var: extracted_data

Audio (Speech)

# Speech-to-text recording
- uses: audio/stt
  with:
    language: "en"           # optional
    output_var: "user_input" # variable to store result

# Text-to-speech
- uses: audio/tts
  with:
    text: "Hello, world!"
    voice: "alloy"           # optional
    output: "./output.mp3"   # optional

Variable Interpolation

Supports two syntaxes:

# GitHub Actions style
prompt: "${{ vars.name }} - ${{ env.API_KEY }}"

# Python format style
prompt: "{name} - {API_KEY}"

Custom Plugins

Create agent_workflows/plugins/my_namespace/my_plugin.py:

from typing import Any, Dict

async def execute(ctx: Dict[str, Any]) -> Dict[str, Any]:
    logger = ctx['logger']
    workspace = ctx['workspace']
    params = ctx.get('with', {})
    
    logger("Running my plugin...")
    
    # Do work...
    result = "some output"
    
    # Return values are available to subsequent steps
    return {"output": result}

Features

• Parallel Execution: Jobs without dependencies run concurrently
• DAG Support: Declare dependencies with needs:
• Variable Interpolation: Use ${{ vars.x }} or {x} syntax
• Environment Inheritance: Global → Job → Step
• Run History: Track execution history in .agent-workflows/history/
• Fail-fast: Stop on first error (configurable per step)

Development

cd packages/agent-workflows
uv sync --dev
uv run pytest tests/
uv run ruff check .
uv run ty check agent_workflows

License

MIT - Copyright (c) 2025 Siili Solutions Oyj