I just noticed that it appears flag parsing happens before executing .ghci. And sure enough, looking at the source, GHCi.UI.runGHCi executes the dot files, which is definitely after all the dynflags parsing stuff in Main. This means that if your .ghci has a -Wall, you can't turn off warnings from the command line. You have to prepend ":set " to the flags and make a local ./.ghci, or pass -ghci-script. I assume this has how it's already been, and I'm surprised it's taken me this long to notice this! Is there a good reason to do it this way, or is it just the way it's always been and no one complained? It's surprising to me because I think most programs do it the other way. E.g. when you start your shell, settings on the command line will override your .zshrc or whatever. zsh for one doesn't seem to clearly document this, but experimenting seems to indicate that is so. I can put all my flags in a -ghci-script without too much trouble, it's just a surprising default.