Jupyter+conda/anaconda is almost always not installable and broken for most of its kernels (it was for Haskell), so I ended up piping GHCi to a websocket server that I found on GitHub and a minimal HTML/JS frontend to talk to it: a reorderable list of textareas.
Then I added some security in the form of passing and verifying long random strings (via files, URL arguments and the user, that one is annoying). GHCi is capable of shelling out, so letting every browser tab on my computer talking to it is not ideal.
AFAIK Jupyter does not have any advanced features like dependency tracking between cells, or tracking variables, data, etc. So if you want Haskell in the browser, I recommend my approach, it is small, fast, the UI is configurable, and the whole process is hackable with hooks and transformations at any place.
Yeah it's a bummer. It seems that notebooks that support these sort of "reactive" workflows are custom built around that model. Marimo, Pluto.jl, and observable are mostly language specific. Creating one would be non trivial.
Do you have your approach documented (tutorial style) anywhere?
