code-review-graph slashes Claude Code review tokens

This is less “AI review helper” and more “context compiler” for big repos: cache the codebase as structure, then let Claude query blast radius instead of rereading the world. The idea is strong, but the project’s own benchmarks also show a tradeoff in review quality, so this looks like a context-reduction layer rather than a replacement for careful human review.

• –Tree-sitter plus incremental SHA-256 diffing makes the graph cheap enough to stay fresh on file saves and git commits.
• –The local SQLite-backed design is practical for teams that want context without pushing code into a cloud index.
• –The 49x result on Next.js suggests the biggest gains show up in monorepos and other sprawling codebases.
• –The README’s benchmark table shows lower review quality scores in the graph-assisted mode, which is a useful reality check.
• –MCP support and Claude Code plugin install flow lower adoption friction for people already living in the terminal.