Your AI CLI Writes Code. Mine Tells You What It'll Break.

AI CLI tools are everywhere right now. Claude Code, Gemini CLI, GitHub Copilot in the terminal — they'll write your code, refactor your modules, even run your tests. But ask any of them: "If I rename this function, what breaks?" They'll scan the files they can see, make their best guess, and probably miss the SQL view that reads the column you're about to change. Or the Java batch job that calls your Python function through a stored procedure. Or the dbt model downstream of the table your migration is about to alter. That's not a knock on AI. It's just not what LLMs are built for. Dependency analysis needs deterministic static analysis, not probabilistic text generation. The gap in every AI CLI Here's what I noticed building with these tools: they're incredible at writing code but terrible at understanding what already depends on it. Ask Claude Code to "add retry logic to the HTTP client" — brilliant. Ask it "what will break if I change the response shape of getUser" — it'll read a few