I wanted to know when an AI coding change got bigger than the prompt

I've been using coding agents more often lately, and one thing kept bothering me.
Sometimes I'd ask for a very small frontend change — change a button style, adjust a hero section, tweak a navbar — and the result would work, but the Git diff would be much larger than I expected.
Git already tells me what changed. What I was missing was the context of what I had actually asked for.
That became a small open-source project called VibeTrace.

The basic idea is to record the prompt before an agent edits the project, capture the resulting Git changes, and compare the expected scope of the request with what actually happened.
For example:

Prompt: "Change the primary button color"

Expected:
UI/styles, probably a few files

Actual:
12 files
6 modules
routing + auth + dependencies touched

That gets surfaced as a large "Blast Radius", with the reasons shown rather than hidden behind an AI-generated confidence score.
I also wanted the checkpoint itself to be safe. VibeTrace uses temporary Git indexes and private refs, so creating a before/after checkpoint doesn't move HEAD, stash the worktree, or modify the real index.
The current v0.2 alpha also has optional Playwright captures, basic pixel/layout/DOM comparison, sessions, timeline replay, standalone reports, and JSON output.
There are still plenty of rough edges. The intent rules need calibration on real repositories, visual captures can be affected by fonts and dynamic content, and I haven't exposed one-click restore yet because I don't want it overwriting later work without drift detection and a dry run.
The code is here if you'd like to try it:

https://github.com/feifeing/vibetrace

I'm mostly interested in whether the core idea is useful: when you're working with a coding agent, would you want to see the difference between what you asked for and how far the resulting change actually spread?