I would not expect to find a "graph" attribute inside the G.graph dict. Nor should G.graph include any node or edge information. G.graph is intended for meta data about the graph rather than info about the nodes/edges/structure of the graph.

I agree, and it makes sense to me.

I mainly use pydot because this is the first answer when you Google "how to export networkx graph to dot", and I don't know agraph does the same thing.

I used the same example with agraph, and it gives the following:

import networkx as nx

G = nx.Graph(test="test")
nx.drawing.nx_agraph.write_dot(G, "test.dot")
G = nx.drawing.nx_agraph.read_dot("test.dot")

print(G.graph)
# {'test': 'test', 'graph': {'test': 'test'}, 'node': {}, 'edge': {}}

test.dot:

strict graph "" {
	graph [test=test];
}

which is also not very intuitive.

Making things match will probably take a while since I will also need to spend time learning the pydot interface. I will need to rewrite a lot of the test as well, which is quite a lot of work.

If the agraph interface should be the "standard", will deprecating the pydot read_dot and write_dot make sense? The agraph interface will be the only way to read and write a DOT file. The only argument I would make against deprecating the pydot interface would be that pydot does not require any system package installation.