Process Mining & Agentic Application (PMAA)

A research project that combines process mining with agentic AI to discover, model, and automate software development workflows from GitHub repository event logs.

📄 Paper Reference

This repository implements the approach described in:

"Using Process Mining to Generate AI Agents from Software Engineering Process Records" Submitted to: BPM 2026 (double-blind review)

🎯 Project Overview

This project demonstrates a complete pipeline from raw GitHub data to an operational multi-agent system:

1. Extract GitHub repository events into OCEL (Object-Centric Event Log) format
2. Discover process models, roles, and behavioral patterns using process mining
3. Generate role-specific BPMN models and DECLARE constraints
4. Build a LangGraph-based multi-agent application grounded in discovered patterns

🏗️ Architecture

GitHub Events → OCEL → Role Mining → Process Discovery → Agentic App

Key Components

• Event Logs: OCEL 2.0 format capturing issues, commits, users, and tasks
• Process Mining: Role-based analysis using pm4py and custom algorithms
• Discovered Artifacts: BPMN models, DECLARE constraints, process descriptions
• Agentic Application: LangGraph-based multi-agent system for GitHub automation

📁 Repository Structure

pmaa/
├── 01-09_*.ipynb              # Sequential analysis notebooks
├── enrich_ocel_with_tasks.py  # Task extraction from commit messages
├── commitizen*.json            # OCEL event logs (raw, enriched, annotated)
├── role_logs/                  # Per-role OCEL splits + visualizations
├── bpmn_models/                # Discovered BPMN diagrams per role
├── declare_models/             # Mined DECLARE constraint models
└── langraph-agentic-app/       # Multi-agent application
    ├── src/                    # Python source code
    ├── docs/                   # Architecture documentation
    └── README.md               # App-specific documentation

🚀 Getting Started

Prerequisites

• Python 3.10+
• Jupyter Notebook
• Git

Installation

1. Clone the repository

   git clone https://github.com/liorlimonad/pmaa.git
   cd pmaa

2. Set up Python environment

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   pip install -r requirements.txt  # If available

3. Install Jupyter dependencies

   pip install jupyter pm4py pandas numpy matplotlib

📊 Analysis Pipeline

Run the notebooks in sequence:

1. 01_enrich-ocel-event-log.ipynb - Parse and enrich OCEL from GitHub data
2. 02_analyze-with-pm4py.ipynb - Basic process mining analysis
3. 03_resource-role-annotation.ipynb - Annotate resources with roles
4. 04_split-logs-by-role.ipynb - Split event log by discovered roles
5. 05_role-process-mining.ipynb - Per-role process discovery
6. 06_declarative-process-mining.ipynb - DECLARE constraint mining
7. 07_process_markdowns.ipynb - Generate process descriptions
8. 08_bpmn_discovery.ipynb - Discover BPMN models
9. 09_role_declare_model_discovery.ipynb - Role-specific DECLARE models

🤖 Agentic Application

The langraph-agentic-app/ directory contains a production-ready multi-agent system:

Features

• Role-based agents: Issue Reporter, Bot, Feature Developer, Quality Engineer, etc.
• GitHub integration: Real REST API calls with dry-run mode
• Policy guards: Risk scoring and DECLARE constraint checking
• Human-in-the-loop: Approval gates for high-risk operations
• Structured state: TypedDict-based state management

Quick Start

cd langraph-agentic-app
python -m venv .venv
source .venv/bin/activate
pip install -e .

# Configure environment
export GITHUB_TOKEN=your_token
export GITHUB_REPO=owner/repo
export GITHUB_ISSUE_ID=123
export GITHUB_DRY_RUN=true

# Run the app
python -m langraph_agentic_app.app

See langraph-agentic-app/README.md for detailed documentation.

🔍 Discovered Roles

The process mining analysis identified 8 distinct roles in software development:

1. Issue Reporter - Creates and clarifies issues
2. Bot - Automated workflow maintenance
3. Feature Developer - Implements new features
4. Maintainer - Reviews and merges contributions
5. Quality Engineer - Testing and validation
6. Technical Writer - Documentation
7. DevOps Engineer - CI/CD and infrastructure
8. Contributor - General contributions

Each role has:

• Dedicated OCEL event log
• OCDFG visualization
• Process description (markdown)
• BPMN model
• DECLARE constraints (where applicable)

📈 Key Insights

• Object-centric approach: Captures complex many-to-many relationships between issues, commits, users, and tasks
• Role specialization: Different roles exhibit distinct behavioral patterns
• Conventional commits: Task semantics extracted from commit message patterns
• Declarative constraints: Temporal and logical rules governing valid workflows

🛠️ Technologies

• Process Mining: pm4py, OCEL 2.0
• Visualization: matplotlib, graphviz
• Agentic Framework: LangGraph, LangChain
• Data Processing: pandas, numpy
• API Integration: requests (GitHub REST API)

📚 Documentation

• LangGraph App Architecture
• GitHub Setup Guide
• Process Descriptions
• BPMN Models

🙏 Acknowledgments

• Built on the Commitizen project event log
• Uses pm4py for process mining
• Powered by LangGraph for agent orchestration

---

Note: This project demonstrates a research prototype. The agentic application runs in dry-run mode by default to prevent unintended GitHub operations.