Revet – Code review CLI that builds a dependency graph

Hey HN, I've been building Revet, a code review CLI that takes a different approach than the current wave of "send diff to LLM" tools. The core idea: parse your codebase into an AST-based dependency graph first (Tree-sitter, 11 languages), run deterministic checks against that graph, and only use LLM reasoning when explicitly asked — and even then, the LLM receives structured graph context, not your raw source code. What it catches that linters don't:

Cross-file breaking changes: you change a return type, revet tells you every caller that will break across the codebase ML pipeline issues: training/serving skew, data leakage, non-reproducible splits Security with context: not just pattern-matching SQL injection, but taint analysis tracing user input through the graph to dangerous sinks Plus: React hooks violations, async anti-patterns, dependency hygiene, error handling

What makes it different from CodeRabbit/Copilot Review/etc:

Deterministic first: 80% of checks need no LLM — they're fast, free, reproducible Offline-first: all deterministic checks work without network Your code stays local: LLM layer receives structured findings, not source Incremental: per-file graph cache, sub-second subsequent runs CI-native: SARIF, GitHub annotations, inline PR comments

Stack: Rust (tree-sitter + petgraph), single binary, no runtime dependencies. Install: cargo install revet revet review GitHub: https://github.com/umitkavala/revet Docs: https://umitkavala.github.io/revet Site: https://revet.dev Built solo over the past few months. Would love feedback, especially from anyone working with large polyglot codebases or ML pipelines.