Don't resolve columns with tidyselect when tbl cannot be materialized

[yjunechoe]

Summary

{tidyselect} requires a data frame present in order to resolve the column selections against. Therefore, this PR ensures that when create_agent(tbl = ) receives a table that cannot (yet) materialize, the validation steps dont't/can't use tidyselect.

This preserves v0.11 behavior, where use of tidyselect triggers materialization of the tbl and thus always errors if tbl cannot (yet) materialize:

# remotes::install_github("rstudio/[email protected]")
create_agent(~blah) %>% rows_distinct(columns = starts_with("x"))
#> Error: object 'blah' not found

In the PR, this now errors with a slightly more informative message:

create_agent(~blah) %>% rows_distinct(columns = starts_with("x"))
#> Error in `rows_distinct()`:
#> ! Cannot use tidyselect helpers for an undefined lazy tbl.
#> Caused by error:
#> ! object 'blah' not found
#> Run `rlang::last_trace()` to see where the error occurred.

Relatedly, the PR also allows the bypassing of tidyselect when tidyselect is not strictly necessary (e.g., when columns is a character vector). This resolves #528 by not forwarding the c(...) character vector to tidyselect at all.

agent1 <- create_agent(~undefined_df) |>
  col_exists(c("cyl", "vs", "am", "gear", "carb"))
get_agent_report(agent1, display_table = FALSE)[, 1:3]
#> # A tibble: 5 × 3
#>       i type       columns
#>   <int> <chr>      <chr>  
#> 1     1 col_exists cyl    
#> 2     2 col_exists vs     
#> 3     3 col_exists am     
#> 4     4 col_exists gear   
#> 5     5 col_exists carb

In sum, {tidyselect} support in columns is a lot more limited when the tbl cannot be materialized, but this is inevitable. This behavior already existed in v0.11 and it's just a genuine tradeoff users will have to make if they want an extreme version of laziness for their tbl.

Remaining consideration with rows_distinct()

Unfortunately, there is one regression with rows_distinct(), because in v0.12 we have the columns = everything() default, whereas in v0.11 we had the columns = NULL default. When the tbl cannot be materialized, rows_distinct() errors on v0.12 but passes through in v0.11.

More concretely put, this used to work in v0.11 but not anymore in v0.12:

agent <- create_agent(~x) %>% 
  rows_distinct()
x <- mtcars
interrogate(agent)

This behavior actually doesn't have anything to do with {tidyselect} though: rows_distinct() is exceptional in that columns = NULL becomes distinct({{ NULL }}) in interrogate(), and distinct(NULL) happens to have a special behavior of checking for row-level distinctness across all columns. In other words, NULL is never resolved to a column selection, which is also why rows_distinct() used to not show anything in the columns column of the agent report.

I have yet to add this behavior of rows_distinct() back in. Adding back support for this should be simple (we'd just intercept everything() and pass down NULL instead) - it'd just require being explicit about this exceptionalism of rows_distinct() in the code.

Also, just to reiterate, this extreme laziness will always work if columns is supplied as character vector or vars(); it's just the dynamic tidyselect expression that requires tbl to be materialize-able.

Let me know what you think!

[rich-iannone]

LGTM, thanks!

[rich-iannone]

Just wanted to say thanks and don't worry about the weirdness of rows_distinct()/rows_complete() (it could always be fixed up later).