Assignment to TypedDict field does not handle bidirectional type inference properly

[A-UNDERSCORE-D]

Environment data

• Language Server version: 2021.7.7
• OS and version: linux x64
• Python version (and distribution if applicable, e.g. Anaconda): 3.9.4 -- pyenv and a normal venv based virtual environment
• python.analysis.indexing: undefined
• python.analysis.typeCheckingMode: basic

Expected behaviour

Pylance understands that TypedDict is a wrapper around dict at runtime and __call__ on any TypedDict is just a call to dict()
which will happily accept a dict as its argument.

Actual behaviour

Pylance complains that you cannot pass a dict to SomeTypedDict(), stating that instead one must use key/value args

Logs

Python Language Server Log

[BG(1)] getSemanticTokens delta previousResultId:1628246633795 at /home/ad/development/python/EDMarketConnector/test-pycodestyle.py ...
[BG(1)]   parsing: /home/ad/development/python/EDMarketConnector/test-pycodestyle.py (1ms)
[BG(1)]   binding: /home/ad/development/python/EDMarketConnector/test-pycodestyle.py (0ms)
[BG(1)] getSemanticTokens delta previousResultId:1628246633795 at /home/ad/development/python/EDMarketConnector/test-pycodestyle.py (7ms)
[FG] parsing: /home/ad/development/python/EDMarketConnector/test-pycodestyle.py (0ms)
[FG] binding: /home/ad/development/python/EDMarketConnector/test-pycodestyle.py (0ms)
Background analysis message: getDiagnosticsForRange
Background analysis message: getDiagnosticsForRange
Background analysis message: setFileOpened
Background analysis message: markFilesDirty

Code Snippet / Additional information

from typing import TypedDict


class Test(TypedDict):
    thing: bool
    stuff: str


x = Test({"thing": True, "stuff": 'test'})

print(x, type(x), dict({"test": "stuff"}))

[erictraut]

I think you mean __init__ rather than __call__. The latter is used only when calling an instance of the class, but in your example, you're calling the class itself (i.e. invoking its constructor), so __call__ is not involved.

PEP 589 is very clear that a TypedDict class should not allow non-keyword arguments passed to its constructor, so pylance is correct in reporting an error here.

The created TypedDict type object is not a real class object. Here are the only uses of the type a type checker is expected to allow: ... It can be used as a callable object with keyword arguments corresponding to the TypedDict items. Non-keyword arguments are not allowed.

Passing a dict to the constructor of a TypedDict class doesn't make sense because that's redundant. In your code example, you're allocating a dict object and then attempting to construct a second copy of that same dict. It's the equivalent of dict({"thing": True, "stuff": "test"}) rather than simply using {"thing": True, "stuff": "test"}.

Here's how a TypedDict type is meant to be used.

x: Test = {"thing": True, "stuff": "test"}

[A-UNDERSCORE-D]

Sorry, I misread the source for the metaclass. You are correct I wanted __init__.

Runtime behavior is however as I described (and thus against the PEP, but thats an argument for the core devs).

I wanted to use it due to I guess a different edgecase or misunderstanding in behaviour where:

from typing import Optional, TypedDict


class Thing(TypedDict):
    thing: bool
    stuff: str


class Thing2(TypedDict):
    other_thing: Optional[Thing]
    stuff: str


thing2: Thing2 = {"other_thing": None, "stuff": "a"}
thing2['other_thing'] = {"thing": False, "stuff": "a"}
# ^^^^^^^^^^^^^^^^ 
# Could not assign item in TypedDict
#   Type "dict[str, Unknown]" cannot be assigned to type "Thing | None"
#     "dict[str, Unknown]" is incompatible with "Thing"
#     Type cannot be assigned to type "None"
#   PylancereportTypedDictNotRequiredAccess

Thus my first solution would be to wrap it in Thing to make the type checker happy (though it should work anyway. I see no reason why it wouldnt?).
Perhaps I should open an issue regarding that behaviour instead?

In the code Im working on I had gone with creating a new var and annotating it appropriately, then using that in the assignment, but its an extra step that feels quite pointless.

[erictraut]

I agree that this code should work. I've fixed the underlying reason for this error message. This will be fixed in the next release of pylance.

[heejaechang]

fixed in 2021.8.1