BUG: find_diffs retains b/ prefixes after the first file in a multi-file udiff

[yifanxiong272]

Issue

Summary

find_diffs removes standard Git a/ and b/ prefixes from the first file in a fenced unified diff, but retains the b/ prefix for the second and subsequent files.

For a two-file diff targeting file1.txt and file2.txt, the parser returns file1.txt and b/file2.txt.

Steps to reproduce

1. Check out Aider main at commit 5dc9490bb35f9729ef2c95d00a19ccd30c26339c.

2. Add the following test inside TestUnifiedDiffCoder in tests/basic/test_udiff.py:

def test_find_multi_diffs_with_git_prefixes(self):
    content = """```diff
--- a/file1.txt
+++ b/file1.txt
@@ -1 +1 @@
-old one
+new one

--- a/file2.txt
+++ b/file2.txt
@@ -1 +1 @@
-old two
+new two
```
"""

    edits = find_diffs(content)

    self.assertEqual(
        [edit[0] for edit in edits],
        ["file1.txt", "file2.txt"],
    )

3. Run:

python -m pytest tests/basic/test_udiff.py -q

4. Observe that the newly added assertion fails while the four existing tests pass.

Expected behavior

find_diffs should normalize the standard Git prefixes for every file-header pair and return:

["file1.txt", "file2.txt"]

The parser already removes a/ and b/ from the initial header pair. A fenced multi-file diff contains the same standard header structure for each subsequent file, so those paths should be normalized consistently.

Actual behavior

find_diffs returns:

["file1.txt", "b/file2.txt"]

The focused test reports:

AssertionError: Lists differ: ['file1.txt', 'b/file2.txt'] != ['file1.txt', 'file2.txt']

First differing element 1:
'b/file2.txt'
'file2.txt'

1 failed, 4 passed

The second edit is therefore associated with b/file2.txt rather than the intended repository-relative path file2.txt.

Root cause

process_fenced_block normalizes the initial header pair:

a_fname = block[0][4:].strip()
b_fname = block[1][4:].strip()

if (a_fname.startswith("a/") or a_fname == "/dev/null") and b_fname.startswith("b/"):
    fname = b_fname[2:]

When a later header pair is encountered, however, the parser assigns the raw +++ path directly:

if line.startswith("+++ ") and hunk[-2].startswith("--- "):
    ...
    fname = line[4:].strip()

The later branch does not apply the normalization used for the initial pair.

UnifiedDiffCoder.get_edits consumes the filenames returned by find_diffs, and apply_edits subsequently resolves each filename as a repository-relative path. The retained prefix therefore propagates beyond the parser.

Possible fix direction

Apply the same paired-header normalization to every ---/+++ pair encountered by process_fenced_block, rather than only the initial pair.

Regression coverage should include a fenced multi-file diff using standard a/ and b/ prefixes and verify the normalized filename of every returned edit.

Version and model info

Aider: 0.86.3.dev, main@5dc9490bb35f9729ef2c95d00a19ccd30c26339c
Python: 3.12.13
Operating system: macOS 15.7.3
Installation: source checkout; focused pytest reproduction
Edit format: udiff
Model: not applicable; reproduced by directly invoking deterministic unified-diff parsing code

[KXHXK]

I'd like to work on this. I'll factor the paired ---/+++ path normalization so it is applied consistently to the initial and subsequent file headers, preserve /dev/null handling, and add focused multi-file regression coverage in ests/basic/test_udiff.py. I'll verify the targeted udiff tests and the repository's formatting checks before opening the PR.