What happens
With ZERO_FORMAT_ON_WRITE=1, writing any .py file launches ruff without the argument that tells it to read standard input, and with its working directory set to the written file's own directory. Ruff's stdin marker is a trailing -; --stdin-filename only names the input. With no path argument and no -, ruff format formats the current directory, which is documented ("ruff format (format current directory)").
So a single-file write can rewrite every .py file in that directory and below, while the tool reports one file written.
Where
internal/tools/format_on_write.go, the .py entry of formatterCommands:
".py": {argv: []string{"ruff", "format", "--quiet"}, stdin: true, filenameFlag: "--stdin-filename"},and formatWithStdin, which sets formatter.Dir to the destination's directory.
Measured
Recorded from the real launch on main at 99721c7, with a recording shim standing in for ruff, writing pkg/new_file.py into a workspace where pkg/already_here.py already exists:
ARGV format --quiet --stdin-filename=...\pkg\new_file.py
CWD ...\pkg
No - anywhere in the argv, and the working directory is the one holding the neighbouring file.
There is a second effect. Under --quiet ruff prints nothing to stdout in that mode, so the written file is published unformatted and the call reports success, which is why this does not show up as a formatting failure.
Suggested fix
Add - to the .py argv, after the flags, so ruff reads the bytes it is given:
".py": {argv: []string{"ruff", "format", "--quiet", "-"}, stdin: true, filenameFlag: "--stdin-filename"},.tf and .lua already pass - for the same reason, and .zig, .gleam and .kt pass --stdin.
While there: dart format has no - either. That one errors out rather than reaching for the directory, so it fails the format and publishes the written bytes, but it means .dart files are never actually formatted.
A regression test wants to assert the argv each stdin adapter is launched with, rather than only that formatting happened, since a formatter acting on its working directory still looks like success from the outside.
What happens
With
ZERO_FORMAT_ON_WRITE=1, writing any.pyfile launches ruff without the argument that tells it to read standard input, and with its working directory set to the written file's own directory. Ruff's stdin marker is a trailing-;--stdin-filenameonly names the input. With no path argument and no-,ruff formatformats the current directory, which is documented ("ruff format(format current directory)").So a single-file write can rewrite every
.pyfile in that directory and below, while the tool reports one file written.Where
internal/tools/format_on_write.go, the.pyentry offormatterCommands:and
formatWithStdin, which setsformatter.Dirto the destination's directory.Measured
Recorded from the real launch on
mainat99721c7, with a recording shim standing in for ruff, writingpkg/new_file.pyinto a workspace wherepkg/already_here.pyalready exists:No
-anywhere in the argv, and the working directory is the one holding the neighbouring file.There is a second effect. Under
--quietruff prints nothing to stdout in that mode, so the written file is published unformatted and the call reports success, which is why this does not show up as a formatting failure.Suggested fix
Add
-to the.pyargv, after the flags, so ruff reads the bytes it is given:.tfand.luaalready pass-for the same reason, and.zig,.gleamand.ktpass--stdin.While there:
dart formathas no-either. That one errors out rather than reaching for the directory, so it fails the format and publishes the written bytes, but it means.dartfiles are never actually formatted.A regression test wants to assert the argv each stdin adapter is launched with, rather than only that formatting happened, since a formatter acting on its working directory still looks like success from the outside.