Files
metabuilder/workflow/examples/python/default_app_workflow
johndoe6345789 51cc840420 feat: Add workflow examples from AutoMetabuilder
Add 19 example workflow packages demonstrating various patterns:

Templates:
- blank: Empty starter template
- single_pass: Single AI request with tool execution
- iterative_loop: AI loop until completion
- contextual_iterative_loop: Loop with repository context
- plan_execute_summarize: Planning and execution pattern

Data Processing:
- data_processing_demo: Filter, map, reduce operations
- conditional_logic_demo: Branching logic
- repo_scan_context: Repository scanning

Plugin Test Suites:
- dict/list/logic/math/string_plugins_test: Unit tests for plugins

Infrastructure:
- backend_bootstrap: Initialize backend services
- default_app_workflow: Full application workflow
- web_server_bootstrap/json_routes: Flask server setup

Specialized:
- game_tick_loop: Game loop pattern
- testing_triangle: CI/CD test pipeline

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-21 16:19:08 +00:00
..

Default Application Workflow

The Default Application Workflow is AutoMetabuilder's production-ready system that combines backend initialization with an iterative AI agent loop. It demonstrates the framework's self-referential approach where internal logic is expressed as declarative workflows.

Overview

This workflow package replaces imperative Python code with a declarative JSON-based approach, providing a complete end-to-end workflow that bootstraps the backend and executes the AI loop.

Key Components

Phase 1: Backend Bootstrap (9 nodes)

Initializes all required services:

  • Message loading from storage
  • Metadata configuration
  • Prompt template loading
  • GitHub client initialization
  • OpenAI client initialization
  • Tool definitions loading
  • Plugin loading and initialization
  • Context seeding
  • Message seeding

Phase 2: AI Agent Loop (8 nodes)

Executes the core agent through iterative cycles:

  1. Loading context
  2. Seeding messages
  3. Making LLM requests
  4. Executing tool calls
  5. Appending results

The loop continues for up to 10 iterations or until no tool calls are returned.

Main Advantages

The workflow-based architecture provides:

  • Separation of Concerns: Clear boundaries between initialization and execution
  • Flexibility: Easy to modify individual nodes without affecting others
  • Observability: Each node execution can be logged and monitored
  • Extensibility: New nodes can be added without changing existing ones
  • Visual: The declarative format enables visual workflow editors
  • Testable: Individual nodes can be unit tested in isolation
  • Modular: Components can be reused across different workflows

File Structure

default_app_workflow/
├── package.json      # Package metadata and configuration
├── workflow.json     # Workflow definition with nodes and connections
└── README.md         # This documentation file

Customization

To create a custom variant:

  1. Copy this package to a new directory
  2. Edit the workflow.json file to modify nodes or connections
  3. Update the package.json with new name and description
  4. Update any configuration references
  • backend_bootstrap - Initialization only
  • single_pass - Single AI request without loop
  • iterative_loop - Loop-only without bootstrap
  • plan_execute_summarize - Advanced planning workflow

Architecture Notes

The system distinguishes between:

  • Immutable Context: Configuration and dependencies that don't change during execution
  • Mutable Store: Execution state that changes as the workflow progresses

This separation enables both workflow data flow and programmatic access patterns.

License

MIT