Running intermediate try scripts with the original shell

[angelhof]

try currently runs the intermediate temporary scripts that it creates using /bin/sh. This causes a problem if we want the internal script to run using a different shell and therefore inherit its state, e.g., bash functions.

A running example of the wanted behavior follows:

$ func() { echo hi; }
$ export -f func
$ bash ./deps/try/try func
hi

I think that a good portable solution would be to run the intermediate scripts with whatever shell try is running too. What is a portable way to determine which shell we are running on? Here are some alternatives:

• The $SHELL variable will contain the name of the login shell (it doesn't work if the subprocess is different than the login shell)
• Using ps, e.g., ps -o args= -p "$$" or ps -o command= -p "$$". This seems to me like the most portable and correct solution.

Here are some SO posts that discuss this issue:

• https://stackoverflow.com/questions/3327013/how-to-determine-the-current-interactive-shell-that-im-in-command-line
• https://unix.stackexchange.com/questions/71121/determine-shell-in-script-during-runtime

Note that we don't really care about the underlying shell, we just want the executable (or even a link to it), so if try was invoked using /bin/sh we can just use that, we don't care that it was bash under the hood.

[angelhof]

@mgree and @dspinellis any thoughts? I can implement the ps solution, I just need a second opinion on whether it is adequately portable :)

[dspinellis]

I don't think the ps way is accurate enough, because it gives the name of the command that invoked the shell, but not the actual shell path. For example, if I execute a script with sh script I will get sh without a path, and won't know whether this is a link to, say, bash or sh. Here are the ways I recommend for commonly used systems.

• For Linux I recommend using stat --format='%N' /proc/$$/exe, which will output something like '/proc/3339774/exe' -> '/bin/bash'
• For FreeBSD I recommend using procstat -bh $$, which will output something like 78297 sh 1400073 /bin/sh (we need to use all fields after the third to take into account spaces in paths)
• On macOS consider using lsof -p $$ -Fnf | sed -n '/^ftxt/{n;/dlyd/d;s/^n//;p;}' (The sed command prints the output of txt entries, removing the dynamic loader entry, which I've seen appear in older macOS versions).

[mgree]

I agree that ps doesn't quite cut it. These detection approaches seem good; we should also have a flag that lets you pick the shell that runs.

Also: these APIs only exist on Linux and FreeBSD, so we don't need to worry about macOS portability.

[kwakubiney]

@mgree which APIs do you speak of? Are you referring to the APIs exposed by try or APIs to detect the shell path?

[mgree]

I'm talking about unshare.