AI Agent Architecture-Based OpenAPI Automated Test Generator: Self-Healing and Intelligent Validation

This project is an OpenAPI automated test generation tool based on AI agent architecture. It can read OpenAPI/Swagger specifications to automatically generate complete test suites, featuring self-healing JSON payloads, strict type checking, and Gemini-based generative error correction capabilities. It aims to address pain points in API testing such as manual repetition, missing edge cases, and high maintenance costs.

Manual API test writing has many issues: repetitive work (similar code needs to be written for each endpoint), missing edge cases (e.g., missing required fields, type errors), high maintenance costs (tests need to be updated synchronously when APIs change), and type safety issues (JSON payload type errors are often found at runtime). This project provides an automated solution through an AI agentic architecture.

Adopts a dual-stack architecture: the backend uses a Python agentic workflow (Streamlit), including modules like parser.py (parses API specifications), validator.py (type checking and AST validation), self_healing.py (two-stage repair: deterministic + generative), and test_agent.py (reflective retry loop); the frontend is a React+Tailwind dashboard that provides real-time monitoring and visualization.

1. Agent loop mechanism: Generate code → AST validation → Error feedback → Self-correction → Regenerate (up to 3 times); 2. Self-healing loop: First perform deterministic repair (type coercion), then call the Gemini API for generative repair if it fails; 3. Deep integration with Google Gemini API (gemini-2.5-flash) to drive code generation and validation.

Repair Example: If the email in the input payload is a number and age is a string, they are converted to the corresponding types after repair; Test Code Example: Generates code that follows pytest best practices, including normal path validation, self-healing tests, etc.; Effective prompts are used during development to assist AI collaboration.

Assumptions: The input schema is a simple dictionary (does not support nested lists/objects), default local port 3000 mapping; Limitations: The TestGenerationAgent relies on LLM determinism, and AST validation cannot guarantee logical correctness (active execution is required).

Applicable Scenarios: Microservice architecture (quick test generation), API version migration (update tests), legacy system documentation (generate tests), team standardization (unify test styles); Summary: This project combines rule engines and generative AI, providing a reference for AI-assisted development tools and freeing developers from repetitive work.