Are hacks necessary to get sbt to work correctly in automated workflows?

[nchammas]

I have seen in a few scripts -- e.g. that run on GitHub Actions -- people invoke sbt like this:

cat /dev/null | sbt scalafmtCheck

Is this necessary?

Various LLMs seem convinced that this is a good idea because it forces sbt to run in non-interactive mode. But sbt seems to work fine without the cat ... prefix, so I wonder if this is just a defunct pattern that used to be necessary long ago and is now just carried forward out of inertia.

As long as you are passing some arguments to sbt, I understand that it will run in batch mode and these hacks are just noise.

Is that correct, or am I missing something?

[eed3si9n]

Is that correct, or am I missing something?

Thanks for this question. Unfortunately, the magical incantation is useful in some situation. It depends who the target user is here, specifically when there's a typo or a mistake in build.sbt file. By default, sbt test etc (not just scalafmtCheck but any task/command) would helpfully block the system prompt asking if the user wants to restart.

$ sbt test
[info] welcome to sbt 1.12.11 (Azul Systems, Inc. Java 17.0.18)
[info] loading project definition from /private/tmp/aaa/project
[error] [/private/tmp/aaa/build.sbt]:1: '=' expected but eof found.
[warn] Project loading failed: (r)etry, (q)uit, (l)ast, or (i)gnore? (default: r) << waits here

Piping /dev/null will stop that from happening:

$ cat /dev/null | sbt test
[info] welcome to sbt 1.12.11 (Azul Systems, Inc. Java 17.0.18)
[info] loading project definition from /private/tmp/aaa/project
[error] [/private/tmp/aaa/build.sbt]:1: '=' expected but eof found.
[warn] Project loading failed: (r)etry, (q)uit, (l)ast, or (i)gnore? (default: r)
$ << you get the prompt back

This might be helpful if Claude etc is often generating some mistake and correcting itself by looking at the error. Without out it the subagent/subprocess would get stuck. There's a built-in option called --batch to do the samething:

$ sbt test --batch
[info] welcome to sbt 1.12.11 (Azul Systems, Inc. Java 17.0.18)
[info] loading project definition from /private/tmp/aaa/project
[error] [/private/tmp/aaa/build.sbt]:1: '=' expected but eof found.
[warn] Project loading failed: (r)etry, (q)uit, (l)ast, or (i)gnore? (default: r)
$

If enough people are using LLM with sbt, maybe we should make --batch behavior the default.

[nchammas]

Very interesting. Thank you for the detailed explanation.

In that case, it seems like good solution for CI users would be to set SBT_OPTS="--batch". That seems better than peppering cat /dev/null | ... or --batch all over the place (assuming one has multiple sbt invocations in their script).

What do you think?

Separately: Shouldn't it be possible for sbt to detect when it's running in an environment with no TTY and refrain from prompting the user, even if the build definition is incorrect?

[eed3si9n]

Yea, so my guess is cat /dev/null is for LLMs, not CIs.