Skip to content

Starting with rlang 1.0, abort() includes the erroring function in the message by default:

my_function <- function() {
  abort("Can't do that.")
}

my_function()
#> Error in `my_function()`:
#> ! Can't do that.

This works well when abort() is called directly within the failing function. However, when the abort() call is exported to another function (which we call an "error helper"), we need to be explicit about which function abort() is throwing an error for.

Passing the user context

There are two main kinds of error helpers:

  • Simple abort() wrappers. These often aim at adding classes and attributes to an error condition in a structured way:

    stop_my_class <- function(message) {
      abort(message, class = "my_class")
    }

  • Input checking functions. An input checker is typically passed an input and an argument name. It throws an error if the input doesn't conform to expectations:

    check_string <- function(x, arg = "x") {
      if (!is_string(x)) {
        cli::cli_abort("{.arg {arg}} must be a string.")
      }
    }

In both cases, the default error call is not very helpful to the end user because it reflects an internal function rather than a user function:

my_function <- function(x) {
  check_string(x)
  stop_my_class("Unimplemented")
}

my_function(NA)
#> Error in `check_string()`:
#> ! `x` must be a string.

my_function("foo")
#> Error in `stop_my_class()`:
#> ! Unimplemented

To fix this, let abort() knows about the function that it is throwing the error for by passing the corresponding function environment as call argument:

stop_my_class <- function(message, call = caller_env()) {
  abort(message, class = "my_class", call = call)
}

check_string <- function(x, arg = "x", call = caller_env()) {
  if (!is_string(x)) {
    cli::cli_abort("{.arg {arg}} must be a string.", call = call)
  }
}

my_function(NA)
#> Error in `my_function()`:
#> ! `x` must be a string.

my_function("foo")
#> Error in `my_function()`:
#> ! Unimplemented

Input checkers and caller_arg()

The caller_arg() helper is useful in input checkers which check an input on the behalf of another function. Instead of hard-coding arg = "x", and forcing the callers to supply it if "x" is not the name of the argument being checked, use caller_arg().

check_string <- function(x,
                         arg = caller_arg(x),
                         call = caller_env()) {
  if (!is_string(x)) {
    cli::cli_abort("{.arg {arg}} must be a string.", call = call)
  }
}

It is a combination of substitute() and rlang::as_label() which provides a more generally applicable default:

my_function <- function(my_arg) {
  check_string(my_arg)
}

my_function(NA)
#> Error in `my_function()`:
#> ! `my_arg` must be a string.

Side benefit: backtrace trimming

Another benefit of passing caller_env() as call is that it allows abort() to automatically hide the error helpers

my_function <- function() {
  their_function()
}
their_function <- function() {
  error_helper1()
}

error_helper1 <- function(call = caller_env()) {
  error_helper2(call = call)
}
error_helper2 <- function(call = caller_env()) {
  if (use_call) {
    abort("Can't do this", call = call)
  } else {
    abort("Can't do this")
  }
}

use_call <- FALSE
their_function()
#> Error in `error_helper2()`:
#> ! Can't do this

rlang::last_error()
#> <error/rlang_error>
#> Error in `error_helper2()`:
#> ! Can't do this
#> ---
#> Backtrace:
#>     x
#>  1. \-rlang (local) their_function()
#>  2.   \-rlang (local) error_helper1()
#>  3.     \-rlang (local) error_helper2(call = call)
#> Run rlang::last_trace(drop = FALSE) to see 1 hidden frame.

With the correct call, the backtrace is much simpler and let the user focus on the part of the stack that is relevant to them:

use_call <- TRUE
their_function()
#> Error in `their_function()`:
#> ! Can't do this

rlang::last_error()
#> <error/rlang_error>
#> Error in `their_function()`:
#> ! Can't do this
#> ---
#> Backtrace:
#>     x
#>  1. \-rlang (local) their_function()
#> Run rlang::last_trace(drop = FALSE) to see 3 hidden frames.

testthat workflow

Error snapshots are the main way of checking that the correct error call is included in an error message. However you'll need to opt into a new testthat display for warning and error snapshots. With the new display, these are printed by rlang, including the call field. This makes it easy to monitor the full appearance of warning and error messages as they are displayed to users.

This display is not applied to all packages yet. With testthat 3.1.2, depend explicitly on rlang >= 1.0.0 to opt in. Starting from testthat 3.1.3, depending on rlang, no matter the version, is sufficient to opt in. In the future, the new display will be enabled for all packages.

Once enabled, create error snapshots with:

expect_snapshot(error = TRUE, {
  my_function()
})

expect_snapshot_error(my_function())

You'll have to make sure that the snapshot coverage for error messages is sufficient for your package.