Skip to content

every Windows dispatcher accepts your .sh and exits 0 without running it, because Git for Windows registered the extension and gave it nothing to do

env paths · case

bothsilentagentciscriptthird-partyrepro: verifiedinert-file-association
Affectscmd, powershell 51, windows
Fails asEFTYPE
Mechanisminert file association
Safe fixdispatch by declared interpreter

A config, a hook, a task or an npm script points at something.sh. It never runs. There is no error. The exit code is 0. Whatever launched it records success, and any check that asserts “the step completed” passes.

If you instead spawn the file directly from a runtime you get EFTYPE — an exec format error, not ENOENT — which sends you looking for a corrupt file rather than a missing interpreter.

One script, six ways to start it. Measured on Windows 11 with Git for Windows installed:

hook.sh contains: #!/bin/bash
echo MARKER_OK
CreateProcess direct (spawn) code=null err=EFTYPE
cmd /c hook.sh code=0 out="" <- ran nothing, said success
cmd /c call hook.sh code=0 out=""
cmd /c start /wait hook.sh code=0 out=""
powershell & hook.sh code=0 out=""
bash hook.sh code=0 out="MARKER_OK"

And the two registry facts behind it:

assoc .sh -> .sh=sh_auto_file
echo %PATHEXT% -> .COM;.EXE;.BAT;.CMD;.VBS;.VBE;.JS;.JSE;.WSF;.WSH;.MSC;.CPL

Only the last row does the thing you asked for.

Two facts stack into silence.

.sh is not an executable type. It is not a PE image, it is not .bat or .cmd, and it is not in PATHEXT. CreateProcess therefore refuses it outright, which is the EFTYPE. Nothing in the Windows loader reads a shebang.

But .sh is associated — Git for Windows registers it as sh_auto_file. So the shell-level dispatchers do not refuse it. They resolve the association, find it has no executing verb, do nothing, and return success. An association that exists but does nothing is worse than no association at all, because “no association” produces an error someone can read.

The result is that the loud path and the silent path disagree, and the one your automation is more likely to take is the silent one.

Name the interpreter and pass the script as an argument. Never hand a .sh to a Windows dispatcher and hope:

bash "C:/path/to/hook.sh" # not: hook.sh
"%WH_BASH%" "%~dp0hook.sh" # from a .cmd, with an absolute interpreter

Resolve that bash to an absolute path rather than trusting the bare name — see bash-on-path-may-be-wsl for what the bare name can resolve to on a stock machine.

If you are generating dispatch for other people’s scripts, the robust shape is to record the interpreter next to the script rather than inferring it from the suffix, which is caller-picks-interpreter-not-shebang.

Why registering an association is the unsafe fix

Section titled “Why registering an association is the unsafe fix”

The tempting repair is to point .sh at an interpreter through assoc/ftype, or to have your tool prepend bash to any command containing .sh. Both look like they work and both are traps: the first is a machine-wide change your software has no business making and that behaves differently on every user’s box, and the second double-dispatches when the command already names an interpreter, so bash script.sh becomes bash bash script.sh. win-hooks hit exactly that and works around it by keeping its hook names extensionless.

Why this is not the other two extension cases

Section titled “Why this is not the other two extension cases”

pathext-bare-name-enoent is about a name with no extension: PATHEXT is consulted, finds nothing, and you get ENOENT. Here the name has an extension, so PATHEXT is never consulted at all.

caller-picks-interpreter-not-shebang is about an interpreter that was already chosen and runs the file in the wrong language, failing loudly with a SyntaxError. Here nobody chose an interpreter and nothing ran.

Same family, three different outcomes: loud refusal, loud mis-execution, and silence.