First, the documentation says that syntax for 'O!' conversion is
object(type='name_of_Python_type'), but actually the type argument should be the name of C structure which represents Python type. I.e. object(type='list') is rejected, and object(type='PyList_Type') is accepted.

Second, from this declaration

/*[clinic]
module spam
spam.ham

    items: object(type='PyList_Type')
    /

[clinic]*/

Argument Clinic generates following parsing code:

    if (!PyArg_ParseTuple(args,
        "O!:ham",
        PyList_Type, &items))
        goto exit;

It is wrong, should be &PyList_Type.

So a documentation error and having to manually specify "&" at the front of your string means it's "broken"?

Nevertheless, I'll take a look at it.

You can't manually specify "&" at the front of the name.

Actually, the documentation (the "howto") states:

Note that object() must explicitly support each Python type you specify for the type argument. Currently it only supports str. It should be easy to add more, just edit Tools/clinic/clinic.py, search for O! in the text, and add more entries to the dict mapping types to strings just above it.

Maybe this is a bad API.  But it's not broken.

Attached is a new, simpler approach for supporting O!.  The object() converter now takes two arguments:
  * type, which is the type you want the parameter declared as
    (e.g. "PyUnicodeObject *")
  * subclass_of, which is the PyTypeObject you want to enforce the
    value is an instance of (e.g. "&PyUnicode_Type")
The old approach was kind of a lovely idea, but was too complicated, and it would have meant registering any new type (like third-party types).  This is less convenient, but simple.  And it would lend itself well to making a custom converter if you used it a lot.

I'll write some documentation for it now, but I wanted to post the code so I could get a review.

p.s. Why was is marked "crash" and "release blocker"?  That's very inaccurate.  And assigned to docs?  WTH?

It is marked "crash" because Argument Clinic generates code which crashes. It is assigned to docs because this feature documentation is misleading.

There are lots of ways you can crash Python by giving erroneous input to Argument Clinic.  Clinic has no visibility into the C type system, so it has no way of verifying whether or not the type objects you pass in are correct.  That's unfixable and not really interesting.

New changeset ddb5cd3e0860 by Larry Hastings in branch 'default':
Issue #20141: Improved Argument Clinic's support for the PyArg_Parse "O!"
http://hg.python.org/cpython/rev/ddb5cd3e0860

Argument Clinic's support for "O!" is now simpler and more flexible.  It's not as convenient as the previous API, but that API wasn't really sustainable.  I'm assuming this fixes your problem; if not please open a new issue.

stage: patch review -> resolved

assignee: docs@python -> larry

nosy: - docs@python
messages: + msg207532
stage: patch review