R: Messages coming from a cli subprocess are red

[DavisVaughan]

They should not be red if our global calling handler is working right

rlang::inform("hi")
callr::r(function() { rlang::inform("hi") }, show = TRUE)

Like pkgdown::build_site()

[DavisVaughan]

• pkgdown::build_site() calls callr::r() to run build_site() in an external process
• build_site() internally calls rlang::inform() to emit messages
• rlang::inform() uses rlang:::default_message_file() to determine where to cat output to and has special behavior for RStudio where it goes to stdout() rather than stderr()

> rlang:::default_message_file
function () 
{
    opt <- peek_option("rlang:::message_file")
    if (!is_null(opt)) {
        return(opt)
    }
    if ((is_interactive() || is_rstudio()) && sink.number("output") == 
        0 && sink.number("message") == 2) {
        stdout()
    }
    else {
        stderr()
    }
}

> rlang:::is_rstudio
function () 
{
    Sys.getenv("RSTUDIO_SESSION_PID") %in% c(Sys.getpid(), getppid())
}

And indeed if you set Sys.setenv(RSTUDIO_SESSION_PID = Sys.getpid()) in the parent process, then calls to pkgdown::build_site() show less red coloring (notably Reading from is no longer red) due to callr inheriting that env var and it matching getppid() (the parent process id, i.e. only looking 1 process up).

callr captures that stdout/stderr (depending on if the envvar is set) and streams it back up to the parent process, but I don't think it is emitted as an R "message" because it doesn't go through our global calling handler (I tried a breakpoint in there). You can see the difference in output channels in Positron:

x <- callr::r_process$new(callr::r_process_options(func = function() rlang::inform("hi")))
x$read_error_lines()
x$read_output_lines()

Sys.setenv(RSTUDIO_SESSION_PID = Sys.getpid())

x <- callr::r_process$new(callr::r_process_options(func = function() rlang::inform("hi")))
x$read_error_lines()
x$read_output_lines()

[DavisVaughan]

There is some additional weirdness in both RStudio and Positron - here is RStudio

One thing of note in my image above is that Building pkgdown site is red in both RStudio and Positron - that goes through cli::cli_rule() and i imagine that does not use rlang::inform(), unlike Reading from, so it probably ultimately emits to stderr in the subprocess https://github.com/r-lib/pkgdown/blob/5f9606b54aedbb36ade5a1f9db4c2780ceaf4e56/R/build.R#L437-L438

The reason why the Installing package rule at the top is white in both rstudio and positron is because it runs in the main process, not the subprocess, so positron's global calling handler runs there and captures it, and idk what happens in rstudio, but i guess something intelligent (or maybe it has to do with the fact that cli messages don't inherit from the "message" class, see cli:::cli__message_create()) https://github.com/r-lib/pkgdown/blob/5f9606b54aedbb36ade5a1f9db4c2780ceaf4e56/R/build.R#L347

[DavisVaughan]

So two tractable issues are:

• rlang::inform() shows as red when run in a subprocess with show = TRUE in Positron but not RStudio because of RSTUDIO_SESSION_PID
• cli::cli_rule() and probably other cli helpers show as red when run in a subprocess with show = TRUE in both Positron and RStudio, but I haven't fully tracked down why