DiffSense

An AI code reviewer that pulls in repository context before it comments on a diff, with auto-generated Playwright regression tests.

GitHub

Overview

Most AI code review tools look at a diff on its own. DiffSense first retrieves the surrounding repository context (related files, call sites, past changes) and only then comments on the change. It flags likely defects, explains why a change is risky, and suggests a fix.

Results

• 97% precision, 73% recall on 30 held-out JavaScript bug cases reconstructed from real BugsJS fixes.
• 40% lower median review time across 30 paired review tasks, while finding a comparable number of confirmed defects.
• 75% of seeded regressions caught across 40 test scenarios by automatically generated Playwright tests.

How it works

1. Context retrieval. Repository code is embedded into PostgreSQL with pgvector, so the reviewer can pull the most relevant code for each changed hunk.
2. Review pipeline. A Node.js pipeline sends the diff and its retrieved context to the model, which returns defects, risk explanations and suggested fixes.
3. Regression tests. For risky changes, DiffSense generates Playwright tests and runs them inside a containerized CI workflow.

Stack

TypeScript · Node.js · PostgreSQL · pgvector · Playwright · Docker