fix: Allow attribute access on union types containing Any #557
Conversation
This commit adds test cases for four reported issues: 1. Issue facebook#547: Dict narrowing after None checks - Already fixed in current version - Tests confirm proper type narrowing for Optional[Dict] after is not None checks - Tests cover attribute chains, early returns, and nested narrowing 2. Issue facebook#548: Pydantic Field overload resolution - Tests document expected behavior with positional arguments - Issue stems from Pydantic's use of sentinel values in type stubs - Workaround: use keyword arguments (default=value) 3. Issue facebook#549: False positives on union types with Any - Added type_contains_any() helper to check for Any in unions - When Any is present in a union, allow any attribute access without errors - Partial fix: works for explicit unions, Dict[str, Any] needs separate handling 4. Issue facebook#550: Init helper methods - Tests document the limitation that helper methods aren't recognized - Full fix requires flow analysis to track methods called from __init__ These tests improve code coverage and document both working features and known limitations, providing a foundation for future improvements.
When Any is present in a union type (e.g., A < /dev/null | Any), any attribute access should be allowed without generating errors for the non-Any members. This commit adds type_contains_any() helper that recursively checks if a type contains Any, and modifies attribute access/assignment/deletion to return Any without errors when true. Fixes facebook#549 (partially - explicit unions work, Dict[str, Any] needs separate handling)
|
The change implemented in this PR looks incorrect to me from the standpoint of type theory and the Python typing spec. I think the PR is based on an incorrect or incomplete understanding of unions and gradual types. For more details about how these work in the Python type system, refer to this section of the typing spec. When accessing an attribute or a method on a union, that method or attribute must be accessible on all subtypes within the union. If one of those subtypes is You can see that pyrefly's current behavior matches that of other type checkers including pyright and mypy. |
|
For the reasons @erictraut mentioned, I don't think we can merge this PR. I'm going to close this, since TypedDicts appear to work as expected and we can track the plan to handle regular dictionaries in #324 |
Summary
This PR fixes false positive errors when accessing attributes on union types that contain
Any.Issue #549: False Positive "missing-attribute" on Union Types
When a union type contains
Any(e.g.,A < /dev/null | Any), Pyrefly was incorrectly reporting missing attribute errors for the non-Any members of the union. According to gradual typing principles,Anyshould allow any operation.Changes
type_contains_any()helper method that recursively checks if a type containsAnyAnyAnywithout generating errorsTests
Added comprehensive tests in
issue_549_any_union.rs:A | Anyunions (no errors)A | Anyunions (no errors)A | Bunions withoutAny(still errors as expected)Limitations
This fixes the issue for explicit unions like
A | Any. The case ofDict[str, Any]subscript access expanding to unions needs separate handling and is not addressed in this PR.Partially fixes #549