Package cli use of "Suggests" rather than "Imports" seems odd

[bnaras]

Clearly rlang functions are being used:

cli/R/rlang.R

Line 46 in 3dd94c0

rlang::abort(

I encountered this when my package which just began importing cli failed CRAN checks.

[msberends]

Yea, had this too, but I would only add rlang to Imports, not any of the others packages to keep the dependencies as few as possible. A solution for the other pkgs is to call rlang::check_is_installed() in cli functions you’re using other deps in.

[gaborcsardi]

What exactly is the issue?

[msberends]

If you rely on cli, eg:

#’ @importFrom cli cli_abort
my_abort <- function() {
  cli_abort(“Something”)
}

Then add cli to Imports as devs should, then:

1. The package will only install cli, not rlang
2. yet rlang is required for cli_abort
3. The function my_abort will fail since rlang is not installed

So it only works if rlang was installed because users also installed tidyverse for example. I’m afraid this is a design flaw in the dependency tree.

[gaborcsardi]

Right. rlang is an optional dependency of cli, and it is only needed for some features, namely cli_abort() and co. Other packaging systems, e.g. Rust crates, have ways to express this, but unfortunately the R package system does not.

I cannot currently add rlang as a hard dependency because pak uses cli, but pak does not work with rlang. I can try to document better that you'll also need to import rlang if you use cli_abort().

Do you see any other way to improve the situation? Maybe cli_abort() could check that the called package depends on rlang, but that could be overkill and/or fragile.

[msberends]

I see.

Then documenting + checking within cli_inform (and friends) would be best I think. Checks deps of others packages is definitely overkill.

You could add a simple:

if (!"rlang" %in% rownames(utils::installed.packages())) {
  stop("This requires the rlang package. Install it using install.packages()")
}

Or something like that?