I've discussed this in some part with various people, so I thought I'd get the thoughts down in writing. Currently, script interpreter usage of Stack has (IMO) three limitations:

• It's slower than just using runghc directly. Some of that is inherent: it needs to check for the presence of GHC and libraries, for example, which runghc does not. Nonetheless, there is more overhead than necessary. From my understanding, settings file checking is a significant component of this (and also likely something we don't want to impact the behavior of script usage).

• Scripts violate some reproducibility rules, because:

  • They don't require that a resolver be set
  • They don't require that all packages be listed
  • They can be impacted by other packages installed

  These trade-offs all make sense when considering the runghc feature in general, but are less appetizing when applied to reproducible scripts.

• There are multiple options that should almost always be set but aren't enforced: --install-ghc, a --resolver as implied above.

Proposal: let's add a new Stack command: stack script. It would have the following semantics:

• Require that a --resolver be set
• Hide all packages from the package database except those specified with --package. (Perhaps make an exception for wired in packages, or at the very least base.)
• Automatically imply --install-ghc
• Ignore all settings files

As an alternative (or addition) to the above, we could consider:

• Making these behaviors options available to runghc (e.g., --no-parse-settings-file)
• Don't actually add a script command, but add support to the script interpreter bits of Stack to recognize a script command and do the right thing.