Claude Code's scripted edit overshoots and deletes 17 unrelated tests

This is one person's account in a bug report on GitHub, not independently confirmed.

What happened

During a long session reviewing pull requests on an open-source project, the developer asked Claude Code, running Opus 5.5 on macOS, to edit and merge tests in a large Python test file. Instead of using an exact-match edit, the model wrote a small Python script to do a bulk text replacement.

How it went wrong

The script replaced everything from the start of one function up to the next top-level function, which it found by searching for three newlines followed by 'def '. That marker isn't always where a function ends. Decorators, blank lines or nested definitions can get in the way, and here the replaced span swallowed the functions that followed. 17 unrelated test functions were deleted, with no warning.

How it was caught

Only by luck, the developer says. The model happened to notice afterwards and rebuilt the file from the last commit with git show HEAD. Had it not, the tests would simply have disappeared from the suite, and a shrinking test suite doesn't fail anything.

Our take

Text-matching is a blunt tool for editing code. Have agents edit by exact match or by parsing the code properly, and add a cheap check after any bulk edit: count the test functions before and after, and stop if the number changed when it shouldn't have.