Listing flags in stack.yaml gives "Invalid flag specification"

[dtaskoff]

I'd expect listing the flags in stack.yaml as in

flags:
  library1:
    flag-name1: true
  library2:
    flag-name2: false

to has the same effect as

stack build --flag library1:flag-name1 --flag library2:-flag-name2

However, the first one fails with "Invalid flag specification" and wants me to add library1 and library2 to the extra-deps field, while the second one succeeds.

EDIT: This happens with stack version 1.5.1 on Windows, and I'll try it on Linux too.

[mgsloan]

Yeah, I recall seeing this distinction in the code. With CLI flags, it does implicit upgrade of packages to being extra-deps. Whereas with the configuration, you need to be explicit about the upgrade of packages to extra-deps.

I think this is a reasonable thing to expect, it potentially makes the behavior of flags clearer to the user. Though, if implicit snapshots gets implemented - #3330 - it will probably make sense to drop this requirement.

That said, the error should definitely be better than Invalid flag specification. I see two possible resolutions:

1. Drop the requirement. One issue with this is that configurations which don't follow the requirement won't work with older stack.

2. Make the error message clearer. Leaning towards this.

[dtaskoff]

Isn't it better for the upgrade to be implicit? If it's not, every time you want to change the lts version in the resolver field, you'll have to check the versions of the packages in this concrete lts version and update the extra-deps field.

EDIT: As a user, I'd expect stack to automatically pick the matching lts versions of the mentioned packages.

[mgsloan]

Yeah, that's reasonable. At the minimum, should have the error message give the line to add to extra-deps.

[snoyberg]

I just noticed this issue in the Haskell Weekly newsletter. I may have accidentally implemented this already on master, thinking that the requested behavior was the current behavior in 1.5 (which apparently it isn't). So before anyone jumps in to implement, simply testing against master would be good.

[dtaskoff]

Yes, it's implemented in master.
So, I'm closing the issue : )

NOTE: Maybe someone should add this to the unreleased changes

[dtaskoff]

EDIT: Sorry for the messed up issue history.