My main question is how this compares to nom which has long been a solid choice for parser combinators in Rust. But no mention in the readme?

Error recovery in nom is left as a very obtuse exercise to the reader. Custom error reporting is difficult at best. That stuff is supposed to be better in chumsky; I don’t know if it actually is.

However, for my own parser which is currently written in nom, my current plan is to port it over to tree-sitter. Its error recovery is completely automatic, and a fair sight better than anything I have time to do by hand.

How do tree-sitter's ergonomics compare to these other two?

The essential difference between nom/chomsky and tree-sitter is that the former are libraries for constructing parsers out of smaller parsers, whereas tree-sitter takes a grammar specification and produces a parser. This may seem small at first, but is a massive difference in practice.

As far as ergonomics go, that's a rather subjective question. On the surface, the parser combinator libraries seem easier to use. They integrate well with the the host language, so you can stay in the same environment. But this comes with a caveat: parser combinators are a functional programming pattern, and Rust is only kind of a functional language, if you treat it juuuuust right. This will make itself known when your program isn't quite right; I've seen type errors that take up an entire terminal window or more. It's also very difficult to decompose a parser into functions. In the best case, you need to write your functions to be generic over type constraints that are subtle and hard to write. (again, if you get this wrong, the errors are overwhelming) I often give up and just copy the code. I have at times believed that some of these types are impossible to write down in a program (and can only exist in the type inferencer), but I don't know if that's actually true.

deep breath

Tree-sitter's user interface is rather different. You write your grammar in a javascript internal dsl, which gets run and produces a json file, and then a code generator reads that and produces C source code (I think the codegen is now written in rust). This is a much more roundabout way of getting to a parser, but it's worth it because: (1) tree-sitter was designed for parsing programming languages while nom very clearly was not, and (2) the parsers it generates are REALLY GOOD. Tree-sitter knows operator precedence, where nom cannot do this natively (there's a PR open for the next version: https://github.com/Geal/nom/pull/1362) Tree-sitter's parsing algorithm (GLR) is tolerant to recursion patterns that will send a parser combinator library off into the weeds, unless it uses special transformations to accommodate them.

It might sound like I'm shitting on nom here, but that's not the goal. It's a fantastic piece of work, and I've gotten a lot of value from it. But it's not for parsing programming languages. Reach for nom when you want to parse a binary file or protocol.

As for chumsky: the fact that it's a parser combinator library in Rust means that it's going to be subject to a lot of the same issues as nom, fundamentally. That's why I'm targeting tree-sitter next.

There's no reason tree-sitter grammars couldn't be written in an internal DSL, perhaps in parser-combinator style (https://github.com/engelberg/instaparse does this). That could smooth over a lot of the rough edges.