I gave build123d a try about a year ago. I really wanted it to work, but it has a lot of issues, mainly in documentation. I'm going off memory here, and it's been a while, so maybe some of these have been fixed. One of the biggest issues is one of the fundamental classes (I want to say "Part") is not documented at all. And it's essentially the most important class. I tried enumerating all of the methods on the class, but didn't make much progress. Fillets were promising, but it seems once you've got a complex edge from a few operations it quits working, or at least did for my part. You're supposed to be able to do something like b=Box(10,10,10), then access the width as b.width, but all of those properties were always zero.

OpenSCAD supposedly supports Python now (https://pythonscad.org/), but I was not able to get it to work at all. I've fallen back to just OpenSCAD, even though it has limitations, at least I'm familiar with them. I'm mostly just waiting for improvements to anything that'll make it better than OpenSCAD.

i got really carried away making a contoured split-keyboard in build123d, and affirm there are some gaps in docs and functionality -- eg. it was quite tricky to get the center of an object in the right form (vector, position, or location), and often zero otherwise. i think a lot of it boils down to an OOP hierarchy with a lot of inherited or generit methods that aren't applicable or impl'ed. OpenCascade causes its own geom issues too, and by the end i was fighting build123d just to slap absolute-position--based hacks over inadvertantly invalid geometry. the dev team's interest in llm-generated code does not inspire confidence that it will become more reliable.

but i found-out up-thread that libfive, which i was esp. interested in for more reliable geom, has a Rust-based successor in Fidget -- and there's even a parametric split-keyboard that can be contrasted with my (static, hack-ier, but single-file) build123d iteration. exited to follow suit!