Coding agents now write a large share of the files in a normal repository. You read the visible text, and you approve it.
Some Unicode characters render as nothing at all. A zero-width space or a private-use codepoint can sit between two letters and carry bits. Several tools claim to strip "AI watermarks" from text, so people assume the marks are there.
Are any actually hidden in your files?
One command answers that now:
watermarks-remover
An experiment in measuring how much invisible AI provenance marking is actually present in real codebases, and removing it losslessly where it is.
The tool "Finds data that someone hid inside text using characters you cannot see, and removes it."
