The "select helper" functions appear to be a bit too aggressive about excluding everything before it. This may be a documentation thing, but I expect positive helper functions to be additive (which appears to be the case), and "none-found" helper functions to not be subtractive.

Note: in all examples, I'm using summarize_at (and mean) primarily for terse output. The behavior is the same with other helpers (e.g., matches, starts_with) and within other verbs (e.g., select, mutate_at).

Currently-Correct Behavior

For the record, "all matches found" works fine, as does a not-found match when it is the first one:

### expecting am, gear, vs
summarize_at(mtcars, vars(contains("am"), contains("gear"), contains("vs")), funs(mean))
#        am   gear     vs
# 1 0.40625 3.6875 0.4375
### expecting gear, vs
summarize_at(mtcars, vars(contains("FOO"), contains("gear"), contains("vs")), funs(mean))
#     gear     vs
# 1 3.6875 0.4375

Expected Behavior

I have been interpreting select(mtcars, contains("am"), contains("foo")) to return all columns that contain "am" or "foo", regardless of the order I provide the calls to contains (and other helpers).

Errant Behavior

Everything before a failing conditional is dropped, regardless of match:

### expecting am, gear
summarize_at(mtcars, vars(contains("am"), contains("gear"), contains("FOO")), funs(mean))
# data frame with 0 columns and 0 rows
### expecting am, vs
summarize_at(mtcars, vars(contains("am"), contains("FOO"), contains("vs")), funs(mean))
#       vs
# 1 0.4375

This appears to be because the helper functions, when nothing is found, return the negative vector of column indices:

contains("a", vars = colnames(mtcars))
# [1]  5  9 10 11
contains("FOO", vars = colnames(mtcars))
#  [1]  -1  -2  -3  -4  -5  -6  -7  -8  -9 -10 -11