[Bug] Native folder picker truncates paths containing characters whose UTF-16 low byte is 0x00 (e.g. CJK 最/需)

[JankenLv]

Version: 0.1.0-rc.6 (also present on main)
Platform: Windows (Win32 IFileOpenDialog path)
Package: @deepseek-ai/dsh-host-directory-picker-native

Summary

Opening a directory via the native folder picker fails with:

workspace create failed: workspace-invalid-path: cannot create a workspace at "D:\...\05-03-RM\07-": ENOENT: no such file or directory, realpath 'D:\...\05-03-RM\07-'

The real folder name is 07-最終產出, but the path sent to workspace.create is truncated to 07-.

Root cause

readUtf16 in packages/host/directory-picker-native/src/win32-dialog-bindings.ts:40 scans for the NUL terminator by checking only the low byte of each UTF-16 code unit:

function readUtf16(koffi: Koffi, address: unknown): string {
  const bytes = Buffer.from(koffi.view(address, 32768))
  let end = 0
  while (end + 1 < bytes.length && bytes[end] !== 0) end += 2
  return bytes.toString('utf16le', 0, end)
}

A NUL terminator in UTF-16LE is 00 00 (both bytes zero), so the loop must check both bytes. Any character whose code point's low byte is 0x00 is mistaken for the terminator:

• 最 = U+6700 → LE bytes 00 67 → scan stops → string cut at the preceding character
• 需 = U+9700 → LE bytes 00 97 → same
• (any code point matching U?00, including Latin-1 chars like Ā U+0100, ā U+0101…)

ASCII is unaffected (each unit is XX 00, low byte non-zero), which is why English-only paths never hit this. It triggers on the first such character anywhere in the path, at any depth — not on path length.

Repro (same scan logic):

input : D:\...\05-03-RM\07-最終產出
output: D:\...\05-03-RM\07-     <- 最 (U+6700) cuts the string

Suggested fix

while (end + 1 < bytes.length && (bytes[end] !== 0 || bytes[end + 1] !== 0)) end += 2

or equivalently check the code unit as a whole:

while (end + 1 < bytes.length && bytes.readUInt16LE(end) !== 0) end += 2

The existing bound end + 1 < bytes.length already guarantees bytes[end + 1] is in range.

Notes

• The in-app browse picker (dsh-client-ui-directory-picker-browse) is unaffected — it uses the server's listing rows, not readUtf16.
• Same file appears as the built lib/worker.cjs in the published package; both copies share the bug.

[Electricitysheep]

这个坑我们手册第 12 章记录过（“Windows 路径 0x00 截断”），根因确认是 readUtf16 只按低字节查 0x00 终止——中文/日文等双字节字符的低字节可能恰为 0x00，导致路径在 07- 处截断。

临时 workaround（官方修复前）：

1. 文件夹名避免含低字节为 0x00 的字符（CJK 字如 “最/需/產”），用纯 ASCII 命名
2. 或者绕过原生选择器：手动输入路径 / 用 workspace create 命令直接建

官方讨论区同根因已有 15+ 帖独立复现（#107 #151 #210 #244 #295 #396 #428 #488 #563 #580 #617 #643 #644 #701 #727 #761 #800），建议合并成一个 issue 追踪。完整排查思路见我们手册：
https://github.com/Electricitysheep/dsh-handbook/blob/main/docs/12-limitations.md