Danny Yang

Fix suppression matching for f/t-strings Summary: `# pyrefly: ignore` comments placed above or on the closing line of a multi-line f/t-string now suppress errors inside the string. Extracts f-string ranges from the AST at the `Errors` level and plumbs them into `ErrorCollector::collect_into` to check suppression at the f-string's start and end lines in addition to the error's own line. Now, if the error suppression comment applies to either the first or last line of the f-string, all errors in the f-string are suppressed. Reviewed By: stroxler Differential Revision: D95964858 fbshipit-source-id: 29c693b3010db66e1f90a17dec2e674590ab4579

port the benchmarks from type_coverage_py repo Summary: This is ported directly from https://github.com/lolpack/type_coverage_py, with a few minor modifications. These are intended to be PUBLIC benchmarks that anyone can run. Changes made: 1. add --runs flag to type checker benchmark 2. fix macos memory measurements to use /usr/bin/time. previously it was completely wrong and only recorded the memory if a type checker used more memory than all type checkers that ran before it, which means in practice only pyright and ty had results 3. integrate the lsp daily_runner.py file (which handles cloning and orchestration) into the lsp_benchmark script directly For now, it should be run using python directly, not via buck Further improvements to come in later diffs Reviewed By: grievejia Differential Revision: D96997879 fbshipit-source-id: 8f78a15e57d5fb19e5539b57e446547b1ad721af

Preserve literal member type in enums Summary: Instead of promoting away the enum member's literal type, we preserve the unpromoted literal & change VisitMut to skip it. Then, in alt/class/enums.rs we modify some logic to decide when to promote away the literal & when to keep it. The special enum attr resolution is weirdly complex because it deals with a bunch of different conditions like django and mixins. I'm not a big fan of how complex it is but it seems necessary. closes https://github.com/facebook/pyrefly/issues/2741 Reviewed By: grievejia Differential Revision: D96157102 fbshipit-source-id: 69ae410dc1afcec9e3a4c290f7724a6560bdb38f

Fix __slots__ error on cross-module class imports Summary: When a class with `__slots__` was imported from another module, and the importing module also defined a class with `__slots__`, pyrefly could report a spurious "not declared in `__slots__`" error on valid attribute accesses. This happened because `ClassDefIndex` is a per-file counter, so looking up an imported class's `__slots__` binding in the current module's binding table could match a completely different class that happened to share the same index. Move the `__slots__` name extraction from `attr.rs` to class metadata computation, which always runs in the class's own module where `self.bindings()` correctly resolves the `ClassDefIndex`. The extracted slot names are stored in a new `SlotsInfo` field on `ClassMetadata`, and`extract_local_slot_names` now reads from metadata instead of bindings. fixes https://github.com/facebook/pyrefly/issues/2950 Reviewed By: rchen152, stroxler Differential Revision: D99442589 fbshipit-source-id: af82359e439011cdfd6bd7d2e5fbc1d33db1f04b

attempt to fix panic in LSP Summary: for https://github.com/facebook/pyrefly/issues/2784 There is no codebase to repro this so it's only an attempted fix, I'll ask the users to try it after next release Reviewed By: rchen152 Differential Revision: D96791064 fbshipit-source-id: 1fd2df6ab401d0fcd34336b80a8e4446288171ec

stubgen: add overloads, type constructors, old-style aliases, and PEP 695 generics (#2721) Summary: Port several features from the AST-only stubgen (PR #2721) into the inference-based implementation: - **Overload handling**: Skip non-overloaded implementations when `overload` variants exist, keeping only the typed overload signatures in stubs. - **TypeVar/NamedTuple/TypedDict preservation**: Recognize assignments like `T = TypeVar('T')`, `ParamSpec`, `TypeVarTuple`, `NewType`, `NamedTuple`, and `TypedDict` calls and emit them verbatim instead of trying to infer types. - **Old-style type alias detection**: Preserve assignments with subscript RHS (`Vector = List[float]`) or union pipe RHS (`X = int | str`) verbatim as type aliases. - **PEP 695 generics**: Extract and emit `type_params` for both classes (`class Stack[T]:`) and functions (`def first[T](...)`). Also adds snapshot tests for each new feature (overloads, typevar, type_alias_old_style, generics). Pull request resolved: #2721 Reviewed By: lolpack Differential Revision: D95793377 fbshipit-source-id: a48aa2d57278dd3b839c9d9c6ce9bbcda93d0d64

stubgen: emit synthesized `__init__` for dataclasses (#3586) Summary: Pull Request resolved: https://github.com/facebook/pyrefly/pull/3586 closes https://github.com/facebook/pyrefly/issues/3221 Dataclasses (and similar patterns like NamedTuples) generate their `__init__` at type-check time rather than declaring it in source. Previously, stubgen would omit the `__init__` entirely for these classes, producing stubs that lost the constructor signature. Look up `KeyClassSynthesizedFields` in the answers to find the synthesized `__init__` type (a `Type::Function` with the full `Callable` signature), then convert its `Param` list into stub parameters. Only emitted when no explicit `__init__` exists in the source. Reviewed By: rchen152 Differential Revision: D106327984 fbshipit-source-id: f7476513b3234393945e6737dfca32cea56f8297

Narrow individual variables in tuple match subjects Summary: When the match subject is a tuple expression (e.g., `match x, y:`), individual variables were not being narrowed by corresponding sub-patterns because `expr_to_subjects` returns empty for tuple expressions, leaving no narrowing subject for the MatchSequence handler. Fix: extract per-element narrowing subjects from the tuple elements and pass them to the MatchSequence handler, which uses them directly instead of faceting a single parent subject. fixes #2863 Reviewed By: rchen152 Differential Revision: D98297415 fbshipit-source-id: 838e52b1e716d618193fe2b9f297bcd4d649d6c9

migrate pyrefly report test suite to snapshot testing Summary: The unit tests were too unwieldy Reviewed By: grievejia Differential Revision: D95792654 fbshipit-source-id: 6b2480fa869704ca026bff943ca613e2171fb425

improve typed dict error messages Summary: fixes https://github.com/facebook/pyrefly/issues/3210 1. make invalid assignments report bad-assignment rather than bad-typed-dict-key, the key exists but the value is wrong 2. print out anonymous typed dict errors differently 3. for anonymous typed dict index errors, emit bad-index rather than bad-typed-dict-key, since anonymous typed dict is an implementation detail Reviewed By: grievejia Differential Revision: D105733104 fbshipit-source-id: 796a15ea7c98133c2ddf54d96bdba03bbbdab589

Add stubgen extraction logic Summary: Add the core extraction that walks a type-checked module's AST and builds stub declarations. Uses pyrefly's binding/answer system to resolve inferred types for functions, classes, and variables. Reviewed By: rchen152 Differential Revision: D95791732 fbshipit-source-id: da16334d228306af479bc2b0291e7149a07929d0

Add integration tests for stubgen with checked-in fixtures (#2808) Summary: Pull Request resolved: https://github.com/facebook/pyrefly/pull/2808 Add snapshot-based integration tests for the stubgen command. Each test case has a checked-in input.py and expected.pyi pair under pyrefly/lib/test/stubgen/. Test categories: - functions: params, defaults, async, pos-only, kw-only, varargs - classes: empty class, inheritance, methods, attributes - variables: annotated module-level variables with values - imports: preserved import statements - mixed: type aliases, classes, and functions together The test harness supports an update mode via STUBGEN_TEST_WRITE_PATH env var for regenerating expected output after intentional changes. Reviewed By: rchen152 Differential Revision: D95791749 fbshipit-source-id: 0ec29cfe5371533979b0a90c350ba7e7d0accb33

Fix provide-type for operator results and type alias qualification Summary: The provide-type endpoint had two bugs where it returned incorrect types: 1. For operator expressions like `+pos`, the endpoint returned the dunder method signature (`(self: Pos) -> Literal[False]`) instead of the expression's result type (`Literal[False]`). The fix is to use `get_result_type_at` instead of `type_from_expression_at` for the provide_type API 2. Type alias names in function signatures were not module-qualified (showed `TA` instead of `bar.TA`). The solution is to store the module in the union display_name field, which shows the original type alias name for unions. fixes https://github.com/facebook/pyrefly/issues/2975 Reviewed By: rchen152 Differential Revision: D99215882 fbshipit-source-id: 802bb97ec815efe49167459dd9b144f024655a86

Add --comment-location flag for configurable suppression comment placement Summary: Add a flag that puts trailing suppression comments rather than putting them on the line before. This flag only applies to `pyrefly suppress`, not `pyrefly check --suppress-errors` which defaults to the old behavior. Default behavior does not change. Google said they would help maintain it. fixes #2648 Reviewed By: rchen152 Differential Revision: D98297121 fbshipit-source-id: e5dbd0d96c53e9a5111baf235a9cf0d67b6eb481

Add backslash continuation support for suppression comments Summary: We can't put a suppression comment after the `\` in a backslash continuation, nor can we put a suppression comment in between lines. We should apply the same pattern here as we do for multi-line strings, where the suppression above the first line applies to the whole expression. The logic handles backslashes appearing in multiline strings, in comments, etc. fixes https://github.com/facebook/pyrefly/issues/2935 Reviewed By: maggiemoss Differential Revision: D98363019 fbshipit-source-id: 7d05d84a46afd9428f19d04fff3e72c55633bd9e

Exclude private module symbols & attributes Summary: Typestats does not report private class methods or attributes (those with a single-underscore prefix like _method or _attr). Pyrefly was reporting all class members regardless of visibility, causing ~2,100 extra symbols in the pandas comparison (87% of the 3,131 pyrefly_only diffs). Filter out class methods and instance attrs whose name starts with _ but not __ (dunders like __init__ are kept). Module-level private names are unaffected since they go through the exports filter. Update inherited_attrs test to cover both public methods (kept) and private methods (filtered). Reviewed By: rchen152 Differential Revision: D99628474 fbshipit-source-id: 8376e18f8d3759737d2b2a62811d064f9b48c1af

Add stubgen data model and emitter Summary: Add the stub data model (ModuleStub, StubItem, StubFunction, StubClass, StubVariable, etc.) and the emitter that generates .pyi text from it. The data types represent the intermediate representation of a stub file, and the emitter handles formatting concerns like spacing, indentation, parameter rendering, and the Incomplete import. Reviewed By: rchen152 Differential Revision: D95791729 fbshipit-source-id: c6d9debc04fc630426bfd3e7093b5f51d9aac9ef

deduplicate identical TypeVars when merging control flow branches Summary: When try/except (or other branching) defines the same TypeVar in multiple branches with different qualified names but identical name, constraint, etc, we create a union of type vars, which cannot be used as a type form. This fix checks that the type vars are structurally identical & have the same name, and merges them. fixes https://github.com/facebook/pyrefly/issues/3402 Reviewed By: rchen152 Differential Revision: D106520566 fbshipit-source-id: af3996d3859da6e3a606040b92c1eec81b68d819

Implement PEP 800 disjoint bases for subclasses (#1592) (#2726) Summary: Pull Request resolved: https://github.com/facebook/pyrefly/pull/2726 The `disjoint_base()` function previously only checked if a class itself was decorated with `disjoint_base`. It did not walk the MRO to find disjoint base ancestors. This meant that subclasses of disjoint bases (e.g. `class A(str)` and `class B(int)`) were not recognized as disjoint — their intersection produced `A & B` instead of `Never`. Fix: Walk the MRO in `disjoint_base()` to find the most specific ancestor marked `disjoint_base`. Changed return type from `&Class` to owned `Class` (cheap — wraps `Arc`) to avoid lifetime issues with the local MRO. fixes https://github.com/facebook/pyrefly/issues/1592 Reviewed By: rchen152 Differential Revision: D95800401 fbshipit-source-id: 9f8907d37ef444d9d6c55f18ef24fb4ab203f1c2

fix false positive w/ multiple match subjects #2168 Summary: Previously, we were only doing some narrow ops if the subject was a single variable. This diff makes it unconditional fixes https://github.com/facebook/pyrefly/issues/2168 Reviewed By: grievejia Differential Revision: D97624241 fbshipit-source-id: 60bbbe57a5faad8c50bea0ef136e7ac0052bec4c