Repellent AI

Repellent AI is an AI agent that learns any project on its own. It analyzes source code, contributors, and business logic to build a knowledge base. In the future, it will use this knowledge to triage and root cause issues.

Entry Point

The main entry point is src/main.py:

python src/main.py --repository /path/to/source/repo --api-key YOUR_API_KEY

Options

Option	Required	Default	Description
--repository	Yes	—	Path to the source repository
--api-key	Yes	—	API key for the LLM provider
--workspace	No	agent_workspace/<project_name>	Directory for analysis outputs
--batch-size	No	10	Batch size for structural analysis
--model_name	No	gemini-3-flash-preview	LLM model to use
--model_provider	No	google_genai	LLM provider (e.g. google_genai, anthropic)

Architecture

The onboarding pipeline runs three workflows in sequence:

1. Structural Analysis → 2. Business Analysis → 3. Contributor Analysis

Workflows

• Onboarding Workflow (onboarding_workflow.py)
  Orchestrates the full onboarding pipeline: structural analysis, business analysis, and contributor analysis. Outputs are written to the agent workspace.

• Structural Analysis Workflow (structural_analysis_workflow.py)
  Recursively discovers source files, creates parallel analysis agents, and produces a per-file analysis including responsibilities, contributors, and functions. Output: file_analysis.md.

• Business Analysis Workflow (business_analysis_workflow.py)
  Reads the structural analysis and generates a business and Critical User Journey (CUJ) overview: product summary, audience, use cases, features, and CUJ stages with linked source files. Output: business_analysis.md.

• Contributor Analysis Workflow (contributor_analysis_workflow.py)
  Gathers contributor data per file and produces an analysis of all contributors: names, accounts, commit counts, and contribution summaries. Output: contributor_analysis.md.

Agents

• Structural Analysis Agent (structural_analysis_agent.py)
  LLM agent that inspects a batch of source files and produces structured summaries (file-level responsibilities, contributors, and functions). Multiple agents run in parallel to analyze the project.

Supporting Modules

• tools.py — Utilities for file and directory operations, git contributor lookup, and file I/O, used across workflows.

Output

All outputs go under the workspace directory (default: agent_workspace/<project_name>/):

• file_analysis.md — Per-file structural analysis
• business_analysis.md — Business logic and CUJ analysis
• contributor_analysis.md — Contributor analysis