CodeReview-Professional-Workflow: A Multi-Round Interactive Training Environment for Professional Code Reviews

CodeReview-Professional-Workflow is a multi-round interactive training environment for AI code review agents, simulating the professional code review process in real-world software development. Agents need to perform tasks like inspection, testing, and compliance verification, collaborate with simulated authors to fix injected defects, support DPO training based on complete trajectories, and provide a standardized training and evaluation platform for building practical AI code review assistants.

Traditional code review tools mostly stay at the static analysis level. This project breaks through this limitation, with core designs including:

1. Multi-round interaction: Simulate the repeated communication process in real collaboration;
2. Comprehensive capability requirements: Agents need to integrate skills such as code inspection, test execution, static analysis, documentation querying, and interpersonal communication;
3. Practical orientation: Inject real-type defects (from missing null checks to complex concurrency issues) to ensure consistency with production environments.

The project uses Docker containerized deployment and provides standardized HTTP API interfaces. Core endpoints include:

• POST /reset: Reset environment state
• POST /step: Execute agent decision
• GET /state: Get environment state
• Others: health, metadata, schema, mcp, etc. This design supports seamless integration of multiple training paradigms such as reinforcement learning and imitation learning.

The environment has built-in defect types with 5 difficulty levels:

• Beginner: Missing null check
• Intermediate: Inefficient loop
• Advanced: Division by zero error
• Expert: Race condition (missing lock)
• Master: Potential deadlock The progressive design allows agents to gradually master the ability to handle complex scenarios from simple problems.

The project supports Direct Preference Optimization (DPO) training, with features including:

• Long-range dependency modeling: Learn strategies across multi-round interactions
• Human preference alignment: Optimize behavior by comparing complete trajectories
• Improved sample efficiency: Extract more information from interaction history Technical implementation highlights: Containerized deployment (reproducibility), modular interface (multi-framework integration), scalable architecture, and Hugging Face platform hosting.

The project's value covers multiple aspects:

• AI researchers: Standardized benchmark environment for code review capabilities
• Developer tool vendors: High-quality training data generator
• Enterprises: Evaluate and optimize internal review processes
• Education field: Programming teaching aid (understand code quality and review skills)

Compared to benchmarks like HumanEval that focus on code generation, this project focuses on the underserved field of code review. Its multi-round interaction design and DPO training support have unique advantages. The project represents the evolution direction of AI-assisted development tools from static analysis to intelligent interactive collaborative review, laying the foundation for practical AI code review assistants.