Windows: choose the child argument encoder per child — MSYS/Cygwin children lose one backslash of each pair inside quoted arguments

[KuSh]

What happens

On Windows rtk builds every child's command line with one encoder, the MSVCRT/libuv convention that native programs (git.exe, rg.exe, node, dotnet) parse with. MSYS2/Cygwin programs (Git for Windows' grep, find, ls, wc, sed, …) parse the command line with Cygwin's own rules (build_argv / quoted / globify in msys2-runtime winsup/cygwin/dcrt0.cc). The two disagree in one case: inside a quoted argument, Cygwin collapses every \\ to \, while MSVCRT keeps the backslashes that do not precede a quote.

So an argument that has to be quoted (it holds a space or a character MSYS would reinterpret) and contains a run of two or more backslashes not followed by a quote reaches an MSYS child with one backslash of each pair missing. No single encoding is exact for both kinds of child:

• MSYS needs the run doubled.
• MSVCRT needs it as written.
• Moving the run outside the quotes depends on Cygwin's drive-letter heuristic, and fails for one parser or the other.

Evidence

Checked with a Linux harness built from msys2-runtime's real build_argv/quoted/globify and glob.cc (commit c770e1b9), next to an MSVCRT-rule parser. The harness was fed the exact command lines rtk's encoder produces:

argument	sent	MSYS child receives	native child receives
\\(	"\\("	\(	\\(
C:\\Temp (x86)	"C:\\Temp (x86)"	C:\Temp (x86)	C:\\Temp (x86)
a\\b(	"a\\b("	a\b(	a\\b(
\\srv\share\(x)	"\\srv\share\(x)"	\srv\share\(x)	\\srv\share\(x)
C:\a\b "x"	"C:\a\b \"x\""	C:\a\b "x"	C:\a\b "x"

Single backslashes (ordinary Windows paths) are exact for both. Only runs of two or more are affected: a UNC path or a regex \\ in an argument that is quoted anyway. For example, rtk grep '\\(' f from PowerShell searches for \( in MSYS grep.

Switching everything to the Cygwin encoding would break native children instead: rtk git commit -m "fix: a\\b" would reach git.exe as a\\\b. That is worse, since git is the most common child.

Proposed

Choose the encoder per child. An MSYS/Cygwin program can be recognised from its executable: its PE import table names msys-2.0.dll or cygwin1.dll.

• MSYS children get the Cygwin-exact encoding.
• Native children keep today's encoding, which is exact for them.
• The choice stays a pure function tested on every platform. Only the PE import scan is Windows-only I/O.
• Cost: reading the import table of the resolved executable, about 1 ms per spawn uncached, less with a per-path cache.

Related

• Windows: arguments containing a double quote are corrupted when rtk spawns MSYS/Cygwin children (rtk grep, rtk proxy) #3727, Windows: MSYS/Cygwin children also lose apostrophes and glob metacharacters, not just double quotes (sequel to #3727) #4102: the MSYS reinterpretation class this encoder addresses.
• refactor(core): make MSYS-safe child argument encoding unforgettable #4103: the ChildCommand encoder refactor. It keeps today's encoding and documents this limit.

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

🔍 Automated Triage

🐛 Category	bug
🟠 Priority	high
🎯 Confidence	85%

Summary

Windows: single argument encoder mis-serializes backslash runs for MSYS/Cygwin children; needs per-child encoder selection via PE import table detection

📁 Relevant files

• src/core/child_command.rs
• src/windows/argv_encode.rs

Suggested Actions

1. Coordinate with refactor(core): make MSYS-safe child argument encoding unforgettable #4103 (ChildCommand encoder refactor) since it documents this exact limitation
2. Review linked Windows: arguments containing a double quote are corrupted when rtk spawns MSYS/Cygwin children (rtk grep, rtk proxy) #3727/Windows: MSYS/Cygwin children also lose apostrophes and glob metacharacters, not just double quotes (sequel to #3727) #4102 for prior MSYS reinterpretation reports this addresses
3. Prototype PE import table scan (msys-2.0.dll/cygwin1.dll) with per-path caching as proposed
4. Reuse author's Linux harness (msys2-runtime build_argv/quoted/globify) as a regression test fixture

---

Triaged automatically by wshm · This is an automated analysis, not a human review.