Fix Windows clipboard paste mojibake for CJK search text

[kirahan]

Problem

ComputerActionExecutor._type_text() pastes text on Windows via clip.exe and Ctrl+V. The clipboard was filled with text.encode("utf-16le") (UTF-16 without BOM).

On localized Windows (e.g. code page 936 / GBK), clip.exe can treat that byte stream as ANSI/system code page instead of Unicode. Pasting into the taskbar search box or other text fields then shows mojibake (e.g. 企业微信 → 吅OON胜醇).

Agent output is unaffected: logs and action_desc still show the correct Chinese string (type: 企业微信). The bug is in local input, not model responses.

Solution

Encode clipboard input as UTF-16 with BOM using Python's utf-16 codec (adds FF FE BOM + UTF-16LE payload), which clip.exe recognizes as Unicode.

Changes

# visual/computer/computer_action_executor.py — _type_text(), Windows branch

# Before
subprocess.run(["clip"], input=text.encode("utf-16le"), check=True)

# After
# clip.exe expects UTF-16 with BOM; utf-16le alone is misread as ANSI on
# localized Windows and pastes mojibake into Search / text fields.
subprocess.run(["clip"], input=text.encode("utf-16"), check=True)

Typing still uses clipboard + Ctrl+V (unchanged); only the Windows clipboard encoding is fixed. macOS (pbcopy / UTF-8) and Linux (xclip) paths are unchanged.

Files Changed

• visual/computer/computer_action_executor.py

Testing

• On Windows 11 (GBK console, chcp 936), run a task that types a CJK app name into taskbar search (e.g. 企业微信).
• Confirm search box shows the correct characters, not mojibake.
• Confirm macOS/Linux paste behavior is unchanged (no regression).