fix(StrReplaceFile): refuse to edit files that are not valid UTF-8

CHANGELOG.md

@@ -12,6 +12,7 @@ Only write entries that are worth mentioning to users.
 ## Unreleased
 
 - Kosong: Stop sending an empty `anthropic-beta` header when no beta features are declared — adaptive thinking removes the interleaved-thinking beta, which previously left an empty header value that some backends reject
+- Tools: `StrReplaceFile` now refuses to edit a file that is not valid UTF-8 instead of silently corrupting it — the whole-file round trip replaced every undecodable byte with U+FFFD, anywhere in the file, including bytes far from the edit and invisible in the approval diff
 
 ## 1.49.0 (2026-07-16)
 

src/kimi_cli/tools/file/replace.py

@@ -131,6 +131,29 @@ async def __call__(self, params: Params) -> ToolReturnValue:
             # Read the file content
             content = await p.read_text(errors="replace")
 
+            # This tool reads the whole file, edits the string, and writes the whole
+            # string back, so every undecodable byte in the file — including bytes
+            # nowhere near the edit — would come back as U+FFFD and be written out as
+            # EF BF BD. Refuse rather than silently rewrite bytes the edit never asked
+            # to touch. A U+FFFD present in the decoded text is only a symptom: it may
+            # equally be a real U+FFFD stored in the file, so confirm against the raw
+            # bytes before rejecting. The strict decode below is deliberate and is
+            # caught, not propagated, so it cannot panic on malformed UTF-8.
+            if "�" in content:
+                try:
+                    (await p.read_bytes()).decode("utf-8")
+                except UnicodeDecodeError as decode_error:
+                    return ToolError(
+                        message=(
+                            f"`{params.path}` is not valid UTF-8 "
+                            f"(byte 0x{decode_error.object[decode_error.start]:02x} at offset "
+                            f"{decode_error.start}). Editing it with StrReplaceFile would "
+                            "replace that byte, and every other undecodable byte in the file, "
+                            "with U+FFFD. No changes were made."
+                        ),
+                        brief="File is not valid UTF-8",
+                    )
+
             original_content = content
             edits = [params.edit] if isinstance(params.edit, Edit) else params.edit
 

tests/tools/test_str_replace_file.py

@@ -246,3 +246,59 @@ async def test_replace_empty_strings(
     assert not result.is_error
     assert "successfully edited" in result.message
     assert await file_path.read_text() == "Hello !"
+
+
+async def test_replace_refuses_file_with_undecodable_bytes(
+    str_replace_file_tool: StrReplaceFile, temp_work_dir: KaosPath
+):
+    """A file that is not valid UTF-8 is left byte-for-byte alone."""
+    file_path = temp_work_dir / "invalid.txt"
+    # The undecodable byte is nowhere near the edit, on a line the edit never mentions.
+    original_bytes = b"alpha\nbeta \xff gamma\ndelta\n"
+    await file_path.write_bytes(original_bytes)
+
+    result = await str_replace_file_tool(
+        Params(path=str(file_path), edit=Edit(old="alpha", new="ALPHA"))
+    )
+
+    assert result.is_error
+    assert "not valid UTF-8" in result.message
+    # Without the guard the file is rewritten with \xff replaced by \xef\xbf\xbd,
+    # growing by two bytes on an edit that only asked to touch "alpha".
+    assert await file_path.read_bytes() == original_bytes
+
+
+async def test_replace_allows_file_containing_real_replacement_character(
+    str_replace_file_tool: StrReplaceFile, temp_work_dir: KaosPath
+):
+    """U+FFFD stored in the file is legitimate content, not a failed decode."""
+    file_path = temp_work_dir / "fffd.txt"
+    original_content = "alpha\nbeta � gamma\ndelta\n"
+    await file_path.write_text(original_content)
+
+    result = await str_replace_file_tool(
+        Params(path=str(file_path), edit=Edit(old="alpha", new="ALPHA"))
+    )
+
+    assert not result.is_error
+    assert await file_path.read_text() == "ALPHA\nbeta � gamma\ndelta\n"
+
+
+async def test_replace_allows_crlf_file(
+    str_replace_file_tool: StrReplaceFile, temp_work_dir: KaosPath
+):
+    """CRLF files must not be mistaken for undecodable ones.
+
+    Reads translate CRLF to LF, so any detection that compares the decoded text
+    against the raw bytes would reject every Windows-line-ending file. (That the
+    write then normalizes the endings to LF is a separate bug, #2191.)
+    """
+    file_path = temp_work_dir / "crlf.txt"
+    await file_path.write_bytes(b"alpha\r\nbeta\r\n")
+
+    result = await str_replace_file_tool(
+        Params(path=str(file_path), edit=Edit(old="alpha", new="ALPHA"))
+    )
+
+    assert not result.is_error
+    assert b"ALPHA" in await file_path.read_bytes()