ruff/crates/ty_python_semantic/resources/mdtest/import/cyclic.md
David Peter 142c1bc760
[ty] Recognize submodules in self-referential imports (#18005)
## Summary

Fix the lookup of `submodule`s in cases where the `parent` module has a
self-referential import like `from parent import submodule`. This allows
us to infer proper types for many symbols where we previously inferred
`Never`. This leads to many new false (and true) positives across the
ecosystem because the fact that we previously inferred `Never` shadowed
a lot of problems. For example, we inferred `Never` for `os.path`, which
is why we now see a lot of new diagnostics related to `os.path.abspath`
and similar.

```py
import os

reveal_type(os.path)  # previously: Never, now: <module 'os.path'>
```

closes https://github.com/astral-sh/ty/issues/261
closes https://github.com/astral-sh/ty/issues/307

## Ecosystem analysis

```
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━┓
┃ Diagnostic ID                 ┃ Severity ┃ Removed ┃ Added ┃ Net Change ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━┩
│ call-non-callable             │ error    │       1 │     5 │         +4 │
│ call-possibly-unbound-method  │ warning  │       6 │    26 │        +20 │
│ invalid-argument-type         │ error    │      26 │    94 │        +68 │
│ invalid-assignment            │ error    │      18 │    46 │        +28 │
│ invalid-context-manager       │ error    │       9 │     4 │         -5 │
│ invalid-raise                 │ error    │       1 │     1 │          0 │
│ invalid-return-type           │ error    │       3 │    20 │        +17 │
│ invalid-super-argument        │ error    │       4 │     0 │         -4 │
│ invalid-type-form             │ error    │     573 │     0 │       -573 │
│ missing-argument              │ error    │       2 │    10 │         +8 │
│ no-matching-overload          │ error    │       0 │   715 │       +715 │
│ non-subscriptable             │ error    │       0 │    35 │        +35 │
│ not-iterable                  │ error    │       6 │     7 │         +1 │
│ possibly-unbound-attribute    │ warning  │      14 │    31 │        +17 │
│ possibly-unbound-import       │ warning  │      13 │     0 │        -13 │
│ possibly-unresolved-reference │ warning  │       0 │     8 │         +8 │
│ redundant-cast                │ warning  │       1 │     0 │         -1 │
│ too-many-positional-arguments │ error    │       2 │     0 │         -2 │
│ unknown-argument              │ error    │       2 │     0 │         -2 │
│ unresolved-attribute          │ error    │     583 │   304 │       -279 │
│ unresolved-import             │ error    │       0 │    96 │        +96 │
│ unsupported-operator          │ error    │       0 │    17 │        +17 │
│ unused-ignore-comment         │ warning  │      29 │     2 │        -27 │
├───────────────────────────────┼──────────┼─────────┼───────┼────────────┤
│ TOTAL                         │          │    1293 │  1421 │       +128 │
└───────────────────────────────┴──────────┴─────────┴───────┴────────────┘

Analysis complete. Found 23 unique diagnostic IDs.
Total diagnostics removed: 1293
Total diagnostics added: 1421
Net change: +128
```

* We see a lot of new errors (`no-matching-overload`) related to
`os.path.dirname` and other `os.path` operations because we infer `str |
None` for `__file__`, but many projects use something like
`os.path.dirname(__file__)`.
* We also see many new `unresolved-attribute` errors related to the fact
that we now infer proper module types for some imports (e.g. `import
kornia.augmentation as K`), but we don't allow implicit imports (e.g.
accessing `K.auto.operations` without also importing `K.auto`). See
https://github.com/astral-sh/ty/issues/133.
* Many false positive `invalid-type-form` are removed because we now
infer the correct type for some type expression instead of `Never`,
which is not valid in a type annotation/expression context.

## Test Plan

Added new Markdown tests
2025-05-13 16:59:11 +02:00

1.4 KiB

Cyclic imports

Regression tests

Issue 261

See: https://github.com/astral-sh/ty/issues/261

main.py:

from foo import bar

reveal_type(bar)  # revealed: <module 'foo.bar'>

foo/__init__.py:

from foo import bar

__all__ = ["bar"]

foo/bar/__init__.py:

# empty

Issue 113

See: https://github.com/astral-sh/ty/issues/113

main.py:

from pkg.sub import A

# TODO: This should be `<class 'A'>`
reveal_type(A)  # revealed: Never

pkg/outer.py:

class A: ...

pkg/sub/__init__.py:

from ..outer import *
from .inner import *

pkg/sub/inner.py:

from pkg.sub import A

Actual cycle

The following example fails at runtime. Ideally, we would emit a diagnostic here. For now, we only make sure that this does not lead to a module resolution cycle.

main.py:

from module import x

reveal_type(x)  # revealed: Unknown

module.py:

# error: [unresolved-import]
from module import x

Normal self-referential import

Some modules like sys in typeshed import themselves. Here, we make sure that this does not lead to cycles or unresolved imports.

module/__init__.py:

import module  # self-referential import

from module.sub import x

module/sub.py:

x: int = 1

main.py:

from module import x

reveal_type(x)  # revealed: int