Devlog: LCAV / Hellcats

LCAV stands for LLM Code Analysis and Validation. I built it during the era when applying a model’s output meant copying a large block of code, or an entire file, over the top of your existing code and hoping nothing quietly disappeared in the process.

What it does

It parses Python and TypeScript, walks the code with an AST parser and some supporting tooling, and builds a map of relations and dependencies. Then you paste in a proposed change and it tells you what that change actually touches.

It auto-detects which files and functions are being modified based on what you pasted, then produces two views: a dependency graph and a text analysis, both at configurable granularity. Impact renders as a heat map, where red means this code was affected and the other colors mark different kinds of effect.

What it catches

The failure modes of that workflow were all the same shape. The model silently drops a function. It drops your comments. It leaves out a variable or a constant. It forgets an import. It replaces something it had no business replacing.

Every one of those is invisible in a diff you’re skimming at speed, and obvious in a dependency map.

The tradeoff

It adds a step between generating code and applying it. You paste into the tool, read the analysis, then go back. That’s real friction and I felt it every single time.

For a production codebase, or any codebase you intend to keep for years, I still think it’s worth paying. The damage from this failure mode isn’t a broken build you notice immediately. It’s slow, silent degradation you find out about much later.

Honest current status

I hardly use it. Models are pretty good these days.