Code review as a pipeline of typed judgments

What it does

Jev Review ingests a Git diff or a full codebase and screens it for correctness, security, reliability, compatibility, and coverage, presenting results in a localhost-only dashboard. The orchestration is written as structured TypeScript workflows rather than open-ended prompts.

The interesting bit

The review chains discrete model calls—profiling files, selecting evidence, classifying mechanisms, scoring severity, and routing conditionally—so each step carries explicit decision boundaries and counterexamples. The codebase itself is strictly layered, with a custom check that fails the build on any upward or cross-layer import.

Key highlights

• Separate entry points for diff reviews and complete codebase scans
• Selects concrete diff hunks or source regions before scoring impact
• Uses changed and related tests as context when judging test gaps
• Dashboard binds to 127.0.0.1 and collapses large reports into sections
• Enforced dependency direction: cli and dashboard import downward through review and adapters to domain

Caveats

• Explicitly an experiment: no integration with compiler diagnostics, static analyzers, repository indexing, or generated explanations
• Findings are structured review prompts, not verified defects
• Requires Node.js 24+ and a TypeSafe API key

Verdict

Try it if you want to see what a typed, multi-step LLM workflow looks like when treated like actual software architecture. Look elsewhere if you need a drop-in linter or a SaaS CI gate.