snapshot transform "leaves a trace" when returning character(0)?

[MichaelChirico]

I am trying to deal with some Windows-specific output issues while converting to snapshot tests:

MichaelChirico/potools#257, as of 8558cf2479195e41a783677c4141eebf933a7a05

To deal with some output that is not produced on Windows, I wrote this helper:

normalize_output <- function(x) {
  # different platforms/installations of gettext apparently
  #   produce a different number of "." in "progress" output; normalize
  x <- gsub("\\.{2,}", ".", x)

  # en@quot translations are not produced on Windows (as of now)
  x <- grep("Generating en@quot translations", x, fixed = TRUE, invert = TRUE, value = TRUE)

  # this is produced alongside the previous message, but in a different iteration of transform()
  idx <- grep("running msgfmt on (?:R-)?en@quot\\.po", x)
  if (length(idx)) {
    browser()
    x <- x[-(idx + 0:1)]
  }
  x
}

Whenever either of the grep() match the snapshot input x, the resulting output x is character() (confirmed by browser()-ing normalize_output() while running tests interactively). That is as intended -- I want the snapshot to act as if that output was never produced, in order to imitate the corresponding output on Windows.

However, on Windows, my CI build still fails:

https://github.com/MichaelChirico/potools/runs/4174597881?check_suite_focus=true

-- Failure (test-translate-package.R:101:1): (code run outside of `test_that()`) --
Snapshot of code has changed:
old[2:8] vs new[2:7]
  "  translate_package(languages = \"cy\")"
  "Message <simpleMessage>"
  "  Writing R-rMsg.pot"
- "  "
  "  'cy' is not a known language. "
  "  Please help supply some metadata about it. You can check https://l10n.gnome.org/teams/<language>"
  "Output"

It seems - " " is a residual line left after running normalize_output(). Either that, or the " " line is not supplied to transform() and thus not available to remove. Either way, it seems I have no way of matching my local snapshot to that produced on Windows CI.

Laterally, is it feasible to support Windows/platform-specific snapshots? Windows in particular seems potentially fruitful given how much the Windows/non-Windows builds of R already differ (terminal access, file paths, etc). I'd rather not use skip_on_os('windows') (e.g., the same failing tests seem to be showing me some encoding issues that I need to fix, so keeping the snapshots running on windows would be ideal).

[hadley]

I'll take a look at the character() issue when I'm next working on testthat.

For the windows build challenge, you could try using a variant — when you set variant, the snapshot is created in _snaps/{variant}/{test}.md instead of _snaps/{test}.md so you could maybe use .Platform$OS.type? The key challenge is getting the variant snapshots back into your local repo. Variants are new so we don't yet have any automation around retrieving them from (e.g.) GHA.