Not sure if this is the right place to comment, but here's one thing that came out (when playing with the "algebra of graphics" stuff). I think the new transform pipeline should be designed in such a way that it could be overloaded either based on the plot type for all arguments (already possible) or on the argument types for all plots (not easy right now, only convert_arguments hack).

I think this is handled well by for example https://github.com/JuliaGPU/Adapt.jl (in a completely different setting). They split adapt in adapt_structure and adapt_storage, which you should overload for wrapper type or innermost type respectively.

Can you give me 2 concrete examples so that I can add them to the tests for the new pipeline?

Mhm, actually by exploring a bit more, I think it was a matter of finding the right method, and it already works well (I'll try to see if it's powerful enough to support grammar of graphics). I think I had a misconception on how this is done in Makie. So, one example (overload on the arguments) could be for the Measurements package.

julia> using Measurements: ±, value, uncertainty, Measurement

julia> using AbstractPlotting

julia> using AbstractPlotting: SceneLike, PlotFunc

julia> function AbstractPlotting.plot!(scn::SceneLike, P::PlotFunc, attr::Attributes, v::Vector{<:Measurement})
           val = value.(v)
           u = uncertainty.(v)
           plot!(scn, P, attr, val)
           plot!(scn, Band, attr, 1:length(v), val - u, val + u)
           return scn
       end

And now, if I define a custom recipe with @recipe, say

julia> using Colors

julia> @recipe(RedPlot, x) do scene
           default_theme(scene)
       end

julia> AbstractPlotting.plot!(plot::RedPlot) = lines!(plot, plot[:x], color = colorant"red")

I can already do redplot(rand(10) .± 0.1). From what I understand, in the refactor this type of approach will be the default, so we need to make sure that this example remains possible.

Changes in #426