Require an annotation if inferred list or dict item type is object

[JukkaL]

Mypy infers the type List[object] for x in this example:

x = [1, 'a']

This is often not what the user expects and can be confusing, as there tends to be an error when the program tries to do something with the list. Often the user would prefer the type to be List[Any] or List[Union[int, str]], but generally the desired item type could be something else, such as an ABC or protocol (once protocols are supported).

It's perhaps impractical to unambiguously infer the desired type, but it would be easy enough to detect cases like these and require an explicit annotation. For the above example, mypy would require an annotation for x (and potentially suggest List[Any] or List[Union[int, str]] as the potential type), similar to how mypy deals with x = [].

For user-defined generic types object might be a reasonable value for a type parameter, so we should probably restrict this to things where the object type is likely a problem. For example, we could generate the message for the types List[object] and Dict[<anything>, object] only. It's less clear if we should do this for dictionaries withobject keys or for sets.

[JukkaL]

#4548 has relevant discussion.

[JukkaL]

Marking as high priority since this seems to be a common source of confusion.

[jxcl]

Working on this.

[gvanrossum]

Should this require a flag to enable it? (Or perhaps allow a flag to disable it?)

[jxcl]

I'm in support of a flag to disable. It is a backwards incompatible change, but no heavy refactoring is required to fix existing code, only taking the inferred types that the mypy errors generate and applying them to the codebase.

Disabling by default would not fix the issue of new users being confused by the inference, since they would also presumably not know about the flag.

[JelleZijlstra]

#6805 has a similar example where we unexpectedly infer Dict[str, int] for essentially {"a": 1, "b": True}.

[ilevkivskyi]

It would be great to sometimes suggest using TypedDict, but such cases are very hard to identify.

[DevilXD]

Is this ever gonna be a thing?

[JukkaL]

We might want to also do this for sets (#8947).

Also, there will need to be a way to disable this, but not enabling this by default would be a bit pointless, as discussed above.

[ChetanKhanna]

Hi! I had been looking at this issue and the PR #6792 that's already opened for it. Since there is no activity on the PR for over a year, can I start a new one from where it left? I believe what's remaining is to fix broken test cases and add a few new ones? I'll try to make it work for sets and the optional flag as discussed above.

[JukkaL]

@ChetanKhanna Sure, feel free to continue working on it!

[ChetanKhanna]

I opened a PR for this: #9051