Visibility of the tool packages

[roman-mazur]

Originally started in Slack.

The documentation of the CUE tool package says that the package is visible only from the _tool.cue files. The motivation of hiding them from the main files is related to the hermetic nature of configurations written in CUE: accessing data from the OS or getting the user inputs breaks it.

However, from my personal experience, this visibility enforcement of the tool/ sub packages does not seem to be true for cue v0.4.3. Here's a quick example of it works:

cue eval

-- print.cue --
package sample

import "tool/cli"

Print: cli.Print & { text: "Hello" }
-- print_tool.cue --
package sample

import "tool/cli"

Print2: cli.Print & { text: "Hello 2" }

command: print: Print
command: print2: Print2

Executing cue eval in the package specified above will print a definition of Print.

And the most fascinating part of it is that I personally find it very convenient and helpful in the context of how cue eval works today. It does not read _tool files at all. So if you have a problem like an invalid disjunction defining your command in the _tool.cue files, the cue cmd may simply report that command as missing, which is quite confusing. I even developed a debugging practice like

-- my_tool.cue --
#SomeCommandPipeline: {
  badPrint: cli.Print & {text: "ok"} & {text: "not ok"}
}
command: #SomeCommandPipeline

-- my_debug.cue --
debug: #SomeCommandPipeline

Without my_debug.cue, I get something like "badPrint not found". With my_debug.cue, I get detailed feedback about the issue with both cue badPrint and cue eval.

Another benefit I get from having command pipelines defs in the main files is the easy way to inspect them. Doing cue eval on the package prints a structure of the command pipeline that I can review. It seems to be impossible today if I keep them in the tool files.

So I wanted to express an opinion that maybe it's not worth enforcing the limited visibility of the tool/ packages.
Mainly because they don't make configuration non-hermetic while cue eval and cue export are used. The configuration of the command pipelines, from my point of view, remains hermetic, and only the execution of the pipelines (invoking cue cmd) breaks this property.

Let me know if it makes sense and what you think about this question.