[Bug]: Gemini edit corrupts non-UTF-8 content during unrelated edits

Is there an existing issue for the same bug?

• I have searched existing issues and this is not a duplicate.

Bug Description

The Gemini edit tool can change unrelated bytes when editing a file that is not valid UTF-8. It reads the file with replacement decoding, applies the requested edit, then writes everything back as UTF-8. For example, changing False to True in a Latin-1 file also changes café to caf�, but the tool reports success.

Expected Behavior

If the file cannot be decoded as UTF-8, return an error and leave its bytes unchanged.

Actual Behavior

Running uv run python reproduce.py with the script below reports success, but replaces undecodable bytes with U+FFFD when the file is written back.

Steps to Reproduce

From a checkout of release v1.49.0, run uv sync --frozen --dev. Save this as reproduce.py and run uv run python reproduce.py. It uses the real executor and a temporary file; no model or API key is needed.

from pathlib import Path
from tempfile import TemporaryDirectory

from openhands.tools.gemini.edit.definition import EditAction
from openhands.tools.gemini.edit.impl import EditExecutor

with TemporaryDirectory() as workspace:
    path = Path(workspace) / "settings.txt"
    original = b"caption = caf\xe9\nenabled = False\n"
    path.write_bytes(original)

    result = EditExecutor(workspace)(
        EditAction(
            file_path="settings.txt", old_string="False", new_string="True"
        )
    )

    print("is_error:", result.is_error)
    print("replacements_made:", result.replacements_made)
    print("before:", original)
    print("after:", path.read_bytes())
    print("bytes_preserved:", path.read_bytes() == original)

Acceptance Criteria

• An edit on a file containing invalid UTF-8 returns an error without changing any bytes.
• Valid UTF-8 edits still work, including files containing an actual U+FFFD character.
• Existing file creation and replacement-count behavior still passes its tests.

Installation Method

Source checkout, UV_FROZEN=true make build (uv sync --dev using the repository lockfile).

If you selected "Other", please specify

Not applicable.

SDK Version

openhands-sdk==1.49.0, openhands-tools==1.49.0; reproduced on main at 97dbce5c37865dd7a3020d2b9662f4a76d11fa71 (release v1.49.0).

Version Confirmation

• I have confirmed this bug exists on the LATEST version of OpenHands SDK.

Python Version

3.13.15 and 3.12.13.

Model Name (if applicable)

Not applicable. Direct tool execution without a model.

Operating System

macOS 26.6.2, arm64.

Logs and Error Messages

is_error: False
replacements_made: 1
before: b'caption = caf\xe9\nenabled = False\n'
after: b'caption = caf\xef\xbf\xbd\nenabled = True\n'
bytes_preserved: False

Screenshots and Additional Context

This report is limited to the Gemini edit tool. Rejecting invalid UTF-8 before writing would prevent the data loss; adding support for other encodings is separate work. I have a small fix and regression tests ready.

Start with openhands/tools/gemini/edit/impl.py and the EditExecutor entry point, then run the provided reproduce.py script after installing the locked dependencies. Add regression coverage for invalid UTF-8, valid UTF-8 containing U+FFFD, file creation, and replacement counts; done means invalid bytes remain unchanged while existing valid edits still pass.