Check GL08 on class __init__ constructors

[mattgebert]

I was looking for a way to skip GL08 checks on class constructors (__init__) throughout my code. This is particularly as the numpydoc recommendation for class strings documents the constructor:

Class docstring
Use the same sections as outlined above (all except Returns are applicable). The constructor (__init__) should also be documented here, the Parameters section of the docstring details the constructor’s parameters.

It seems that there ought to be a default check ignore for __init__ constructors if no docstring has been provided (and if a class docstring has been provided). Obviously if an (optional) docstring for __init__ has been created, then it ought to be checked.

This conditional checking doesn't seem possible with the default override_<code>, exclude toml tags.
Keen to hear others thoughts!

[mattgebert]

In case it wasn't clear, GL08 has the description "The object does not have a docstring".

[stefanv]

That makes sense to me.

[mattgebert]

Hey all (@larsoner, @stefanv),

I have re-opened the issue as the current implementation is not working with the AstValidator. It is my understanding that the AstValidator is the preferred validator class, used when running numpydoc linting and CI testing? (i.e. The modifications we made are not /will not be used in the majority of numpydoc validation).

When we improved the class constructor init GL08 reporting, we only wrote tests for the Validator class which works perfectly. This is because Validator uses the inspect library to find the code_obj, we can determine the parent module / class to check if the __init__ is documented in the constructor.

However, under the implementation of the abstract syntax tree (ast) validator (numpydoc.hooks.validate_docstrings.AstValidator) the code_obj attribute does not become defined, making the resolving of the parent class (that the constructor is defined on) difficult. I have done a preliminary check if I can still resolve the similar names required to run the Validator._load_obj without the code_obj attribute.

Here's the way we get currently (roughly) get the class reference from the constructor:

if hasattr(doc, "code_obj"):
    cls_name = doc.code_obj.__qualname__.split(".")[0]
    module_name = doc.code_obj.__module__
    cls = Validator._load_obj(f"{module_name}.{cls_name}")

Any ideas what we should do for the AstValidator? Unfortunately it seems that the ast nodes, including ast.FunctionDef, only have a body attribute to see sub-nodes, but there's no link to the parent node.

The AstValidator is given a _filename attribute and _name attribute. The latter is a concat list of the sub-module class nodes, which should allow us to track to the class structure. I've just tested the following to be working:

elif isinstance(doc, AstValidator):
    # Equivalent to the above, but for the AstValidator in validate_docstrings.py.
    names = doc._name.split(".")
    if len(names) > 2: # i.e. module.class.__init__
        nested_cls_names = names[1:-1] #class1,class2 etc.
        cls_name = names[-2]
        filename = doc.source_file_name #from the AstValidator object

        # Use AST to find the class node...
        with open(filename) as file:
            module_node = ast.parse(file.read(), filename)
        
        # Create a function to find the class node
        def find_class_node(node: ast.AST, cls_name) -> ast.ClassDef:
            for node in ast.walk(module_node):
                if isinstance(node, ast.ClassDef) and node.name == cls_name:
                    return node
            raise ValueError(f"Could not find class node {cls_name}")
        
        # Recursively find each subclass
        next_node = module_node
        for name in nested_cls_names:
            next_node = find_class_node(next_node, name)
        
        # Get the documentation.
        assert isinstance(next_node, ast.ClassDef)
        assert next_node.name == cls_name
        cls_doc = AstValidator(ast_node=next_node, filename=filename, obj_name=cls_name)
    # else:
    #     # Ignore module.__init__ edge case functions. 

This seems quite convoluted just to get the parent class for the constructor. What are your thoughts?

[mattgebert]

Further discussion on this, should all tests in test_validate also be written/run with AstValidator in mind?