carstenh asked in IRC how to make a shebang for mksh(1) scripts that works on both regular Unix and Android.

This is not as easy as it looks, though. Most Unicēs will have mksh installed, either manually or by means of the native package system, as /bin/mksh. Some put it into package manager-specific directories; I saw /sw/bin/mksh, /usr/local/bin/mksh and /usr/pkg/bin/mksh so far. Some systems have it as /usr/bin/mksh but these are usually those who got poettering’d and have /bin a symlink anyway. Most of these systems also have env(1) as /usr/bin/env.

Android, on the contrary, ships with precisely one shell. This has been mksh for a while, thankfully. There is, however, neither a /bin nor a /usr directory. mksh usually lives as /system/bin/mksh, with /system/bin/sh a symlink(7) to the former location. Some broken Android versions ship the binary in the latter location instead and do not ship anything that matches mksh on the $PATH, but I hope they merge my AOSP patch to revert this bad change (especially as some third-party Android toolkits overwrite /system/bin/sh with busybox sh or GNU bash and you’d lose mksh in the progress). However, on all official Android systems, mksh is the system shell. This will be important later.

The obvious and correct fix is, of course, to chmod -x the scripts and call them explicitly as mksh scriptname. This is not always possible or desirable; sometimes, people will wish it to be in the $PATH and executable, so we need a different solution.

There’s a neat trick with shebangs — the absence of one is handled specifically by most systems in various ways. I remember reading about it, but don’t remember where; I can’t find it on Sven Mascheck’s excellent pages… but: the C shell variants run a script with the Bourne Shell if its first line is a sole colon (‘:’), the Bourne family shells run it with themselves or ${EXECSHELL:-/bin/sh} in those cases, and the kernel with the system shell, AFAIK. So we have a way to get most things that could call the script to interpret it as Bourne/POSIX shell script on most systems. Then we just have to add a Bourne shell scriptlet that switches to mksh iff the current shell isn’t it (lksh, or something totally different). On Android, there is only ever one shell (or the toolkit installer better preserve mksh as mksh), so this doesn’t do anything (I hope — but did not test — that the kernel invokes the system shell correctly despite it not lying under /bin/sh) nor does it need to.

This leaves us with the following “shebang”:

	:
	case ${KSH_VERSION-} in
	*MIRBSD\ KSH*) ;;
	*)	# re-run with The MirBSD Korn Shell, this is an mksh-specific script
		test "${ZSH_VERSION+set}" = set && alias -g '${1+"$@"}'='"$@"'
		exec mksh "$0" ${1+"$@"}
		echo >&2 E: mksh re-exec failed, should not happen
		exit 127 ;;
	esac

The case argument not only does not need to, but actually should not be quoted; the expansion is a set -u guard; the entire scriptlet is set -e safe as well; comments and expansions are safe. exec shall not return, but if it does (GNU bash violates POSIX that way, for example), we use POSIX’ appropriate errorlevel. zsh is funny with the Bourne shell’s way of using "$@" properly. But this should really be portable. The snippet is both too short and too obvious (“only way to do it”) to be protected by copyright law.

Thanks to carstenh and Ypnose for discussing things like this with us in IRC, sending in bugfixes (and changes we decline, with reason), etc. — it feels like we have a real community, not just consuments ☺