Repository navigation
Passing these=d to attr.s skips attributes defined in class body #621
Description
@attr.s(these=dict(a=attr.ib()))
class test:
b: int = attr.ib(default=0)
c = test(5)
print(c.b)
This fails even if passing auto_attribs=True, albeit in a slightly different way.
The obvious workaround is to only put attributes in these, but that is harder for code analysis tools to understand (it lacks type annotations, even if the tools are attrs aware).
The second workaround would be to put the attributes from these into the class body, but that doesn't work if the attribute is inherited or describe a data descriptor (like getter/setters).
The final workaround is to duplicate the definition into both these and the class body, but that creates an opportunity for the two to get out of sync.
Anyway, ideally, attrs would add anything from the class body which is an instance of _CountingAttr, (and anything from the annotations dict if auto_attribs is set). If that is not practical, then a warning or exception should be raised rather than having weird runtime behavior.
Activity
wbolster commentedon Mar 21, 2022
taking one step back, what is the use case for using both styles at the same time?
perkinslr commentedon Mar 21, 2022
I believe I described the use case already, but to reiterate, passing a long list of attributes via the
theseparameters is ugly, but if you want to include an inherited data descriptor in yourinit/repr,eqand so on, you cannot put that annotation in the class body (maybe you can with theauto_attribs, but I don't think so).rawsun007 commentedon Aug 21, 2026
Still reproduces on main (0.1.dev20). The mechanism is the intersection of two separately documented behaviours, which may be why it has stayed confusing.
Two lines cause it, both intentional on their own:
_make.py,_transform_attrs:if these is not None: ca_list = list(these.items()). The class body is never consulted, exactly as the docstring says ("If these is passed, use that and don't look for them on the class"), sobis never collected as a field._make.py:self._delete_attribs = not bool(these). Withthesepassed, class-body attributes are not deleted either, per the 18.1.0 note ("If these is passed, no attributes are deleted from the class body").So
bis neither collected nor cleaned up, and the_CountingAttrsentinel that would normally be consumed and removed survives as a plain class attribute.test(5).bthen returns attrs' private placeholder rather than0or anAttributeError.The two documented behaviours are each reasonable; the undocumented part is the result, which is a private object reaching user code with no error. If the intent is to keep both, a note under
thesesaying thatattr.ib()in the class body will be left in place as an unconsumed sentinel would have answered this in 2020. If leaking_CountingAttris not intended, the narrow fix is to raise whentheseis passed and the class body still contains_CountingAttrinstances, which is detectable at exactly the point in (1) where the body is skipped.Not sending a patch since which of those you want is your call.
Disclosure: I investigated this with Claude Code. I ran the reproducer and read both call sites myself; the line references above are from current main.