why not the jupytext format?

[majidaldo]

#2124

From what i can tell, marimo is parsing code for variables to assign them to input and output. So why not just stick with ~jupytext? If io is going to be parsed anyways, why need a "format" for marimo?

[akshayka]

You mean the file format, with @app.cell decorators? It's to make notebook files lazy so that they can be imported without executing the entire notebook, while still making the code accessible for re-use (ie not locked away in an if name == main guard). We have limited support for notebook reusability/composition (see the App and Cell sections in our API docs: https://docs.marimo.io/api/index.html), with more support planned.

Learn more here: https://github.com/marimo-team/meps/blob/main/mep-0001.md

[akshayka]

it's all about typing: the ide just doesn't know about cell argument types.

Yes, I agree that's the biggest drawback of our file format. I would love if we had an editor agnostic (LSP?) plugin that inferred types for the cell arguments. Or, we can try to serialize type hints in function args. This gets messy with imported modules.

[dmadisetti]

We could record the type on tree traversal if provided. E.g.

a : int = 0
b =  None

untyped_response = a if b else 1

typed_response: str = a if b else 1

gives

@app.cell
def _(): # require all hints for typing
    a : int = 0
    b =  None
    return a, b

@app.cell
def _(a: int, b):
    untyped_response = a if b else 1
    return (untyped_response,)

@app.cell
def _(a: int, b) -> tuple[str]:
    typed_response: str = a if b else 1  # Don't bother to enforce poor typing, defer to linter, just capture what's put
    return (typed_response,)

---

I don't think resolving implicit typing is worth it

[akshayka]

Using user types was my first thought. But we already have "implicit" types through jedi/lsp, no?

The only thing that's troublesome to type is modules. Eg, if np is an arg to a cell. You'd want completion.

[dmadisetti]

I think a reasonable compromise is the setup cell for types and modules

[akshayka]

Yea i think that could be fine. With setup cell for modules, plus ideally inferred types (in a way that doesn't make saving slow), we'd have a pretty great editing experience!