AgentForge

AgentForge is an open-source agentic AI system for reliable software debugging and code repair.

Instead of allowing an AI model to directly modify a repository, AgentForge follows a controlled workflow:

1. Understand the reported issue.
2. Create an isolated working copy.
3. Inspect the relevant source files.
4. Apply restricted, auditable edits.
5. Run the maintained test suite.
6. Reject incorrect repairs.
7. Save the final diff, validation output, and evidence report.

The system is designed to handle practical software failures such as:

• Incorrect invoice calculations
• Duplicate orders after network retries
• CSV imports that lose the final record
• API crashes caused by missing optional fields
• Network clients that retry forever

AgentForge includes a web dashboard where users can select a repair scenario, monitor the workflow, inspect the generated diff, review test output, and download the evidence report.

For evaluation, AgentForge was tested against five pinned open-source Python repositories. All five original test suites passed, and all five injected regressions were detected. One guided Ollama repair was verified end-to-end with the complete maintained test suite. A trusted local fixture was also successfully repaired using qwen2.5-coder:7b.

The project is presented honestly as a working open-source prototype focused on controlled edits, validation, and evidence-based repair.