The .gitignore Explainer shows which rule decides each path. Its file box says to paste the output of git ls-files --others or git status --porcelain --ignored. Today's red-team run tested that advice against git 2.43. Every rule was matched correctly, but the paths themselves were read wrong.
Here are three lines git really prints:
M src/a.js
R x.js -> "y z.js"
?? "caf\303\251.txt"
The checker read these as the paths M src/a.js, x.js -> "y z.js and caf/303/251.txt. The last one is a folder that doesn't exist.
- The status column can start with a space. A file that changed but isn't staged prints as
M path. The parser trimmed the line before it looked for the two-letter status code, soMbecame part of the path, and an anchored rule like/src/missed it. - Git quotes unusual names. With
core.quotePathon, which is the default, git writes any name with non-ASCII bytes, a tab, a quote or a backslash in C style, with octal escapes. The parser removed the quotes but kept the escapes, then turned each backslash into a/(meant for Windows paths). - A rename has two names. Ignore rules apply to the new one.
#at the start of a line isn't always a comment.git ls-files --othersprints#draft.md, or an Emacs autosave like#notes.txt#, without quotes. The box dropped both. Now only#followed by a space, or a bare#, counts as a comment.
The command-line version had a related problem. It reads exact names with git ls-files -z, then joined them with newlines and sent them through the same paste parser. So a file really named back\slash.log was checked as a path inside a back folder. Now it passes the list to the engine unchanged.
The sixth find came from a seeded bug that the tests didn't catch. A ! rule in .git/info/exclude, blocked by a folder rule in the root .gitignore, was flagged but never given a fix. The fix is the same line copied into the root .gitignore, since git reads that file first and both match paths from the repository root. Git agrees:
$ git check-ignore --no-index -v build/keep.txt
.gitignore:4:!build/keep.txt build/keep.txt
The guard
Tests written by hand only check the cases someone thought of, so the new test asks git. It builds a repository with a modified file, a deleted file, a renamed file and eight oddly named ones (café.txt, a tab, a quote, a backslash, a #, a non-ASCII name in an ignored folder). It pastes git's own status --porcelain --ignored and ls-files --others output into the engine. Then it checks that each path comes back under the name git status -z gives it, never quoted, with git's own ignored or untracked verdict. On the old parser that test fails, and on the new one it passes.
Seeded bugs: 20 across two campaigns, and the existing tests caught 16 (80%). The four that escaped were the [[:print:]] and [[:graph:]] character classes, which had never been tested, an edge case in splitting a pasted bundle, and the missing exclude-file fix above. Each now has a test that fails with the bug in place, so the catch rate on this file is now 100%.
Results are a reading of git's documented rules, checked against one git version. They are not a guarantee about every git build or setting.