copy.replace type inference regression: breaks Self-returning __replace__ with bound TypeVar #21672

Activity

1. trim21

   changed the title [-]copy.replace type inference regression: breaks Self-returning __replace__ with bound TypeVar[/-] [+]`copy.replace` type inference regression: breaks Self-returning __replace__ with bound TypeVar[/+] on Jul 3, 2026
2. A5rocks

   added

   bugmypy got something wrongmypy got something wrong

   topic-type-variables

   topic-self-typesTypes for selfTypes for self

   on Jul 3, 2026
3. trim21 commented on Aug 9, 2026

   

   trim21

   on Aug 9, 2026

   Last edited by trim21

   Author

   More actions

   relative: python/typeshed#15973 python/typeshed#15997
4. JPHutchins

   added a commit that references this issue on Oct 20, 2026

   test: Seeded keeps its own __post_init__ beside SeededByDefault's (#354)

   ...

   482f203
5. JPHutchins commented on Oct 21, 2026

   

   JPHutchins

   on Oct 21, 2026

   Last edited by JPHutchins

   More actions

   Warning

   LLM Disclosure

   This post was authored by claude-opus-5-5 on behalf of @JPHutchins. While working out JPHutchins/salix#359, JP asked me to report the upstream causes with evidence; this adds the union cases to this issue.

   The union siblings of this issue: in mypy 2.4.0, copy.replace on a union of dataclasses returns object, and so does a TypeVar bound to a union. That's a join, which is wider than this issue's bound. pyright 1.1.414 and ty 0.0.90 keep the argument's type in both cases.

   import copy
   from dataclasses import dataclass
   from typing import TypeVar
   
   
   @dataclass(frozen=True)
   class A:
       x: int
   
   
   @dataclass(frozen=True)
   class B:
       x: int
   
   
   T = TypeVar("T", bound="A | B")
   
   
   def bump(value: A | B) -> A | B:
       return copy.replace(value, x=1)  # error: got "object", expected "A | B"
   
   
   def bump_bound(value: T) -> T:
       return copy.replace(value, x=1)  # error: got "object", expected "T"

   Measurements and the mypy issues behind them

   • Measured with mypy --strict --python-version 3.14 and mypy 2.4.0's bundled typeshed, which includes Make copy.replace more strongly typed typeshed#14819. Each call reports Incompatible return value type (got "object", ...).
   • The join predates Unpacking an iterable with an __iter__ method that returns self leads to variable being silently inferred as Any #14819. I ran the signature from Refine the copy._SupportsReplace.__replace__ signature typeshed#14786 (covariant _SupportsReplace[_RT_co], **changes: Any) as a local stub: mypy 2.4.0 returns object for both functions there too.
   • mypy solves _RT_co from the union by join, which is abs(Union[int, Decimal]) is inferred as object #8601 for SupportsAbs. (🎁) Perform 'union math' on non overloaded calls #13493 ("union math" on non-overloaded calls) might fix it.
   • A second, never-matching copy.replace overload doesn't help when the whole union matches the first overload, as this identical-field A | B does. It does help when each member has a field the others lack: mypy then splits the union and returns it unwidened. The details, and the same rejection in pyright and ty, are in copy.replace rejects a union whose members each have a field the others lack (since #14819) typeshed#16582.
   • Downstream: replace: a union or union-bound TypeVar comes back as Struct (mypy), and a union whose members each have their own field is rejected (pyright, ty) JPHutchins/salix#359.