Review Copilot V2

Review Copilot V2 is an evidence-backed, risk-aware pull request review system designed to make code reviews more reliable without removing humans from the decision-making process. Reviewing a pull request properly requires understanding the change, checking its downstream impact, validating test coverage, and ensuring that coding and security standards are followed. Under time pressure, teams may either rush reviews and miss bugs or spend too much time reviewing every change. Review Copilot V2 addresses this by first understanding the intent of a change and creating a shared review brief. It then assesses the risk of the change and assigns specialized reviewers for correctness, impact, tests, standards, and security. These reviewers investigate the same change from different perspectives while sharing their findings through a coordinator. A key principle of the system is “evidence before conclusions.” AI-generated suspicions are not automatically treated as valid findings. Findings are validated against the actual code and, when available, supported by real evidence such as tests, linters, or type checks. Unverified issues are explicitly marked as unresolved instead of being presented as confirmed problems. The final output is a decision-ready report containing a change summary, risk assessment, confirmed blocking findings, non-blocking suggestions, verification records, and open questions requiring human judgment. The system does not automatically approve or merge pull requests, bypass branch protections, or silently modify business logic or tests. A human reviewer remains in control of the final decision. The initial version is designed as a local CLI tool using TypeScript, Node.js, Git, SQLite, Vitest, and a provider-neutral AI model adapter, with future support planned for GitHub integration, sandboxed verification, and IBM Bob once those integrations are verified.