Triumvirate: A Collaborative Intelligent Code Review System with Three Models

Triumvirate is a multi-model collaborative intelligent code review system launched by ClosedLoop Technologies. Its core lies in integrating three large language models to conduct code reviews from different professional perspectives, addressing issues such as high time costs, limited coverage, and subjective judgment differences in traditional manual reviews. The system has full repository analysis capabilities, supports CI/CD integration and private deployment, and provides development teams with comprehensive and in-depth code quality assessments and improvement suggestions.

Traditional manual code review faces challenges like high time costs, limited coverage, and subjective judgment differences. The core concept of the Triumvirate project is derived from the concept of 'triumvirate'—it simulates role divisions such as architects, security experts, and performance engineers through collaborative decision-making of three independent and equal models, thereby enhancing review coverage and accuracy while reducing biases and blind spots of a single model.

Triumvirate adopts a multi-model architecture, calling three independent large language models to analyze the same codebase: Model 1 focuses on functional correctness and business logic integrity; Model 2 emphasizes code quality and maintainability; Model 3 concentrates on security and performance. After the three models output independently, the system intelligently integrates consensus and points of disagreement. Disagreements are marked for manual focus, improving review reliability.

Unlike single-file review tools, Triumvirate has full repository analysis capabilities. It can understand the project's overall structure, module dependencies, cross-file call chains, and other contexts, and identify architecture-level issues (such as circular dependencies and unclear responsibilities). It supports multiple programming language frameworks, builds a code semantic graph to understand function calls and data flows, and provides constructive suggestions beyond surface syntax.

It generates structured reports, classifying issues by severity, type, and location, with accompanying explanations, impact analysis, repair suggestions, and code comparison examples. It supports Markdown/HTML/JSON output and can be integrated into tools like Jira and GitHub. A knowledge base is built for repeated issues, improvement trends are tracked, and historical data can be viewed to understand the evolution of code quality.

It supports CI/CD integration, providing CLI and Docker images to embed into existing pipelines. Common scenarios: automatic review and comment posting when a PR is created/updated; regular full scans of the main branch to monitor health; pre-commit checks to detect issues early. Review rules and thresholds can be configured, strictness can be customized, and checks can be enabled/disabled for specific files/directories.

It supports private deployment; code analysis is performed in a local/controlled environment and not uploaded to third-party services. When cloud models are needed, code desensitization is provided to replace sensitive information such as API keys and database connection strings. Review logs and reports follow the principle of least privilege—only authorized users can access them, and administrators can configure data retention policies to clean up expired records.

Triumvirate represents a new direction in AI-assisted code review. It provides comprehensive and objective perspectives through multi-model collaboration, preserving human decision-making rights while improving efficiency and coverage. In the future, it is expected to integrate multi-dimensional analysis such as runtime performance, user feedback, and security intelligence. Its modular architecture reserves space for expansion and will continue to evolve.