fix: harden capture_patch and snapshot_untracked filename handling

tests/v1/test_git_utils.py

@@ -0,0 +1,65 @@
+"""Focused coverage for the filename-handling gaps in #2196: ignore entries
+must reach git as literal pathspecs, and non-UTF-8 filenames must fail loudly
+instead of producing a wrong patch."""
+
+from types import SimpleNamespace
+
+import pytest
+
+from verifiers.v1.runtimes.base import ProgramResult
+from verifiers.v1.utils.git import capture_patch, snapshot_untracked
+
+
+class ScriptedRuntime:
+    """Duck-typed runtime: records run() argv and plays back scripted results."""
+
+    def __init__(self, results: list[ProgramResult], read_data: bytes = b""):
+        self.calls: list[list[str]] = []
+        self._results = list(results)
+        self._read_data = read_data
+
+    async def run(self, cmd: list[str], env: dict) -> ProgramResult:
+        self.calls.append(list(cmd))
+        return self._results.pop(0)
+
+    async def read(self, path: str) -> bytes:
+        return self._read_data
+
+
+@pytest.mark.asyncio
+async def test_capture_patch_passes_ignore_entries_as_literal_pathspecs():
+    # A file named `*.py` must not glob-unstage every Python file the agent
+    # touched, and a leading `:` must not be parsed as pathspec magic.
+    ok = ProgramResult(exit_code=0, stdout="", stderr="")
+    runtime = ScriptedRuntime([ok, ok], read_data=b"")
+    trace = SimpleNamespace(info={})
+
+    await capture_patch(trace, runtime, ignore=["*.py", "weird[name].txt", ":odd"])
+
+    diff_argv = runtime.calls[0]
+    assert diff_argv[4:] == [
+        ":(literal)*.py",
+        ":(literal)weird[name].txt",
+        ":(literal):odd",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_snapshot_untracked_splits_nul_delimited_names():
+    runtime = ScriptedRuntime(
+        [ProgramResult(exit_code=0, stdout="a.txt\0dir/b bin\0", stderr="")]
+    )
+    assert await snapshot_untracked(runtime) == ["a.txt", "dir/b bin"]
+
+
+@pytest.mark.asyncio
+async def test_snapshot_untracked_rejects_lossily_decoded_names():
+    # Runtimes decode stdout with errors="replace": a non-UTF-8 filename
+    # arrives as U+FFFD and can never match the real file again. Returning it
+    # would leave the file out of the ignore set and credit the agent with an
+    # image file — fail before a wrong patch can be produced.
+    runtime = ScriptedRuntime(
+        [ProgramResult(exit_code=0, stdout="caf�.bin\0", stderr="")]
+    )
+    with pytest.raises(ValueError, match="not\\s+valid UTF-8"):
+        await snapshot_untracked(runtime)

verifiers/v1/utils/git.py

@@ -100,7 +100,21 @@ async def snapshot_untracked(runtime: Runtime, env: dict | None = None) -> list[
     )
     if result.exit_code != 0:
         return []
-    return [path for path in (result.stdout or "").split("\0") if path]
+    stdout = result.stdout or ""
+    # Runtimes decode stdout with errors="replace", so a filename that isn't
+    # valid UTF-8 reaches us with U+FFFD substituted — it can never match the
+    # real file again, and capture_patch would credit the agent with an image
+    # file it didn't create. A wrong patch is worse than a loud setup failure.
+    # (A filename legitimately containing U+FFFD trips this too; that is the
+    # price of the runtimes' lossy text contract.)
+    if "�" in stdout:
+        raise ValueError(
+            "snapshot_untracked: `git ls-files` returned a filename that is not "
+            "valid UTF-8; it cannot round-trip through the runtime's text "
+            "decoding, so the pre-agent untracked set would be wrong. Rename "
+            "the offending file in the image or exclude it via .gitignore."
+        )
+    return [path for path in stdout.split("\0") if path]
 
 
 async def capture_patch(
@@ -142,9 +156,14 @@ async def capture_patch(
     nonce = uuid.uuid4().hex
     full, capped = f"{_FULL}_{nonce}", f"{_CAPPED}_{nonce}"
     cmd = _DIFF.format(full=full, capped=capped, cap=PATCH_CAP_BYTES + 1)
+    # `git reset -- <path>` treats arguments as pathspecs, so a file named
+    # `*.py` would unstage every Python file the agent touched. `:(literal)`
+    # makes git match the name byte-for-byte (and also neutralizes names that
+    # start with `:`, which git would otherwise parse as pathspec magic).
+    ignore_pathspecs = [f":(literal){path}" for path in ignore or []]
     try:
         result = await runtime.run(
-            ["sh", "-c", cmd, "vf-capture-patch", *(ignore or [])],
+            ["sh", "-c", cmd, "vf-capture-patch", *ignore_pathspecs],
             {**(env or {}), "VF_DIFF_BASE": base_commit or "HEAD"},
         )
         if result.exit_code != 0: