Because you can refer to the _Type from Python, that is the whole point of this exercise. And because you can refer to it from Python you can pass it to calldll.newcall and such.
I still fail to see why you need additional ParseTuple support in calldll.
Sorry, you're right. My example was for a return value, so we're talking Py_BuildValue here. But this situation is equivalent to a GrafPort argument, where PyArg_Parse would be used.
In cdc_call, there is a loop over all arguments, rather than a ParseTuple call. I don't see how this could change: all arguments are processed uniformly. Precisely how would you use O@ in there? Actually, it may be worthwhile to get rid of the PyArg_ParseTuple call in call_newcall also: for performance reasons, to soften the dependency on MAXARG, and to give better diagnostics in case of user errors. There is a loop over argconv, anyway; this loop could have run over args in the first place. All you might want to have is additionals slots in type objects; as Thomas explains, you can have that using just the 2.2 facilities. For the specific case of calldll, it seems that a generic mechanism would be harmful: You want to be absolutely sure that an object is convertible to a long *for the purposes of API calls*. So I'd even encourage to create a PyCallDll_RegisterTypeConverter function; extension types that want to support calldll should register a conventry and a rvconventry. That approach works for any Python version. Regards, Martin