Per discussion, setting PWD correctly when using os/exec is likely the right solution here rather than adding new API.

And the way to set PWD correctly is to set cmd.Dir and to either not set cmd.Env or to set it by appending things to the result of cmd.Environ().

FWIW, getcwd(3), the usual way for a C program to find the name of the current directory, does not do the $PWD shortcut trick, nor is it allowed to by POSIX.1: "The pathname copied to the array shall contain no components that are symbolic links."

xtools$ cat cwd.c
#include <stdio.h>
#include <unistd.h>

int main() {
	char buf[1000];
	puts(getcwd(buf, sizeof buf));
}

xtools$ gcc cwd.c
xtools$ exe=$(pwd)/a.out
xtools$ $exe
/Users/adonovan/w/xtools
xtools$ mkdir /tmp/real
xtools$ ln -sf /tmp/real /tmp/link 
xtools$ ( cd /tmp/link && $exe )
/private/tmp/real
xtools$ ( cd /tmp/link && PWD=/tmp/link $exe )
/private/tmp/real

I realize that the odds of the driver program being in a language other than Go are low, but I don't think this interface should assume anything about the implementation language of the child process or require it to know that it should use os.Getenv("PWD") in preference to getcwd when constructing the filenames that it prints.

The C library function that looks at PWD is get_current_dir_name.

Checking PWD is a pretty wide-spread convention even if getcwd and getwd don't do it.

If we set PWD, is this causing actual problems?

The C library function that looks at PWD is get_current_dir_name.
Checking PWD is a pretty wide-spread convention even if getcwd and getwd don't do it.

FWIW, that's a GNUism rather than a standard feature of C, and I don't think I ever used it (or was much aware of it, or the motivating subtlety) when I was working in C/C++. But perhaps you're right that there is a convention.

If we set PWD, is this causing actual problems?

Not at present. Perhaps this can be downgraded to a documentation issue. I think there are two aspects:

1. os/exec.Cmd should document its treatment of Dir and PWD, explain that os.Getwd checks PWD, and recommend this approach for all subprocesses regardless of implementation language. I've filed os/exec: document subtleties around Cmd.Dir, PWD, os.Getwd and C's getcwd #68000.
2. packages.DriverRequest should explain that it sets PWD and expects that the child process will honor it, redirecting the reader to comment number 1 for more detail.

Retracting the proposed change; leaving the issue open until item 2 in the previous note is addressed.

Change https://go.dev/cl/616057 mentions this issue: go/packages: document the role of PWD