fix: force C locale for internal git

[nalajcie]

Summary

Force LC_ALL=C on internal git subprocesses so parser-facing git output is stable across user locales, and normalize C-quoted --numstat path tokens before matching them.

Problem

Some internal flows parse git stderr to distinguish expected conditions from real failures. One example is authorship notes fetch handling, where a missing remote refs/notes/ai ref should be treated as "no remote notes yet" instead of a command failure.

That parsing currently assumes English git output. I observed local test failures when git emitted localized Polish stderr for the missing notes ref, e.g. fatal: nie znaleziono zdalnej referencji refs/notes/ai, which did not match the existing English missing-ref checks.

Forcing LC_ALL=C fixes the localized stderr problem, but it also makes Git C-quote non-ASCII filenames in git diff --numstat output. The unified diff parsers already normalize quoted Git path tokens via unescape_git_path; the numstat parsers did not, so non-ASCII filenames could fail status pathspec post-filtering or ignore matching in status/range stats.

Fix

• Set LC_ALL=C centrally for internal git command execution in src/git/repository.rs.
• Keep GIT_EXTERNAL_DIFF and GIT_DIFF_OPTS cleanup in the same helper so internal git command environment normalization remains in one place.
• Unescape --numstat filename tokens before status pathspec post-filtering.
• Unescape --numstat filename tokens before status and range-authorship ignore matching.
• Add focused regression tests for C-quoted non-ASCII --numstat paths.

I kept the current line-oriented --numstat parsing instead of switching to -z, because the existing parser structure only needed the same Git path-token normalization already used by unified diff parsing. A NUL-separated parser would be a larger behavioral rewrite for the same issue.

Testing

The new focused regression tests failed before the unescape fix:

• commands::status::tests::numstat_post_filter_matches_c_quoted_non_ascii_paths
• commands::status::tests::numstat_ignore_matcher_matches_c_quoted_non_ascii_paths
• authorship::range_authorship::tests::numstat_ignore_matcher_matches_c_quoted_non_ascii_paths

Validated after the fix with:

• task test TEST_FILTER=numstat_ NO_CAPTURE=true
• task test TEST_FILTER=status_ignore NO_CAPTURE=true
• task test TEST_FILTER=range_authorship NO_CAPTURE=true
• task test TEST_FILTER=utf8_filenames NO_CAPTURE=true (101/102 passed; one daemon completion timeout passed on exact retry)
• task test TEST_FILTER=test_rtl_with_ltr_mixed_filename_in_worktree NO_CAPTURE=true
• task fmt
• task lint

Earlier locale-specific validation:

• task test TEST_FILTER=internal_git_command_forces_c_locale NO_CAPTURE=true
• task test TEST_FILTER=fetch_notes::test_fetch_notes_no_remote_notes NO_CAPTURE=true
• task test TEST_FILTER=fetch_notes NO_CAPTURE=true
• task test TEST_FILTER=ci_local_skip_fetch NO_CAPTURE=true
• task test TEST_FILTER=internal_machine_commands::test_fetch_and_push_authorship_notes_internal_commands_json NO_CAPTURE=true

---

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.

[nalajcie]

the windows CI seems a little bit flaky, those changes passed in my fork:

• https://github.com/nalajcie/git-ai/pull/1/checks

Note: this is a side issue I've found during implementing submodule support (nalajcie#2), but I'll test it locally a little bit before submitting official PR. On this PR - the windows (daemon) tests failed (the same issue: lost connection with the worker), but passed upon rerun.

[svarlamov]

Thank you @nalajcie -- I do think that Devin finding was a real issue so good to see that patched with regression tests. Windows CI is indeed a bit flaky, but if you rebase on latest main it should be mostly addressed

[svarlamov]

Thanks for the contribution again @nalajcie

Added @heapwolf to help take a review pass and should go out probably in the release after next

[devin-ai-integration]

Devin Review found 1 new potential issue.

View 9 additional findings in Devin Review.

[devin-ai-integration]

🟡 parse_two_path_tokens splits on spaces within unquoted filenames, breaking diff --git header parsing for files with spaces

The new parse_two_path_tokens function splits unquoted tokens on any whitespace character. Git does NOT C-quote paths that only contain spaces (space 0x20 is in the printable ASCII range and doesn't trigger quote_c_style), so a file named hello world.txt produces the header diff --git a/hello world.txt b/hello world.txt. The old code line.split(" b/").last() correctly extracted hello world.txt, but parse_two_path_tokens("a/hello world.txt b/hello world.txt") returns tokens ("a/hello", "world.txt"), yielding the wrong path world.txt. This causes current_diff_file in run_diff_tree_with_hunks (src/authorship/rebase_authorship.rs:2837) to be set incorrectly, so hunks are tracked against a nonexistent file path, leading to silently lost attribution data during rebase for any file whose name contains spaces.

Prompt for agents

The parse_two_path_tokens function in src/authorship/rebase_authorship.rs:2579-2633 incorrectly splits unquoted path tokens on whitespace. Git only C-quotes paths containing non-printable characters, backslashes, double-quotes, or non-ASCII bytes — NOT spaces. So filenames with spaces (e.g. 'hello world.txt') appear unquoted in diff --git headers.

The diff --git header format is inherently ambiguous for unquoted paths with spaces. Git itself resolves this by relying on the --- and +++ lines, not the diff --git line. One approach to fix this:

1. For quoted paths (starting with "), the current logic is correct.
2. For unquoted paths, you cannot simply split on whitespace. Instead, leverage the fact that in diff --git headers, the a-side and b-side paths are identical (for non-rename cases). You can try to find the split point by looking for the pattern where the second half starts with 'b/' and the remainder equals the first half with 'a/' prefix replaced. For example, scan for ' b/' separators and check if a/X matches b/X.
3. Alternatively, fall back to the old approach of splitting on ' b/' for the unquoted case only, while keeping the quoted-path parsing for C-quoted headers.

The function is called from parse_diff_git_new_path at line 2567, which is used in run_diff_tree_with_hunks at line 2837 to set current_diff_file for hunk tracking during rebase.

---

Was this helpful? React with 👍 or 👎 to provide feedback.