Jim Hester

`tools::testInstalledPackage(Ropts = "-d valgrind")` is probably the easiest way to do this.

I guess you could also use `debugonce` for this

Yeah that seems reasonable, we can use a similar approach to [with_tempfile()](https://github.com/r-lib/withr/issues/32#issuecomment-153476774), which hasn't yet made it to master...

Yeah, I guess the tests never try opening multiple devices so don't run into this, definitely a bug.

While I agree it is very confusing, this behavior is consistent with what would happen if you open and close the devices manually, so I am not sure we should...

Does the fact that the error message is translated mean this experiment failed?

This seems like a duplicate of https://github.com/r-lib/remotes/issues/431 IMHO this seems like a bug in base R to me, you can reproduce the issue without remotes at all.

Thank you for the pull request! We recently merged a fairly substantial change to how `install_version()` works internally. If you are still interested in fixing this issue could you look...

In general, the directory in `R_LIBS_USER` needs to exist _before_ the R process starts up, creating it afterwards is basically undefined behavior.

Can you give an example workflow and build log which is failing to install the second order dependencies?