This repository provides:

• PEP 561 typing stubs packages for the Trio project packages:
  • trio (trio-stubs)
  • outcome (outcome-stubs)
  • async_generator (async_generator-stubs)
• A package trio_typing containing types that Trio programs often want to refer to (Nursery, AsyncGenerator[Y, S], TaskStatus[T]) and a mypy plugin that smooths over some limitations in the basic type hints.

To type-check code that uses trio-typing, you need CPython 3.5.2 or later. (Mypy requires 3.5.2+, and its dependency typed-ast doesn't support PyPy.) We test on Linux using the latest releases from the 3.5, 3.6, and 3.7 branches, as well as 3.8-dev nightly. We're not knowingly doing anything OS-specific, so other OSes should work too.

You can run code that uses trio-typing on any platform supported by Trio, includng PyPy and CPython 3.5.0 and 3.5.1.

Type checkers other than Mypy are not supported, but might work. Experience reports and patches to add support are welcome.

Install with:

pip install -U trio-typing

Enable the plugin in your mypy.ini (optional, but recommended):

[mypy]
plugins = trio_typing.plugin

Start running mypy on your Trio code! You may want to import some typing names from trio_typing, like Nursery and TaskStatus; see below for more details.

The stubs packages provide types for all public non-deprecated APIs of trio, outcome, and async_generator, as of the release date of the corresponding trio-typing distribution. You don't need to explicitly configure these; just say import trio (for example) and mypy will know to look in trio-stubs for the type information.

The trio_typing package provides:

• Names for two important types that Trio keeps anonymous: Nursery and TaskStatus[T] (where T is the type of the value the task provides to be returned from nursery.start()). These are implemented as ABCs, and the actual private types inside Trio (like trio._core._run.Nursery) are registered as virtual subclasses of them. So, you can't instantiate the trio_typing types, but isinstance(nursery, trio_typing.Nursery) where nursery is a Trio nursery object does return True.
• A backport of typing.AsyncGenerator[YieldT, SendT] to Python 3.5. (YieldT is the type of values yielded by the generator, and SendT is the type of values it accepts as an argument to asend().) This is an abstract class describing the async generator interface: AsyncIterator plus asend, athrow, aclose, and the ag_* introspection attributes. On 3.6+, trio_typing.AsyncGenerator is just a reexport of typing.AsyncGenerator.
• CompatAsyncGenerator[YieldT, SendT, ReturnT], a name for the otherwise-anonymous concrete async generator type returned by @async_generator functions. It is a subtype of AsyncGenerator[YieldT, SendT] and provides the same methods. (Native async generators don't have a ReturnT; it is only relevant in determining the return type of await async_generator.yield_from_().)
• A few types that are only useful with the mypy plugin: YieldType[T], SendType[T], ArgsForCallable, and the decorator @takes_callable_and_args.

The trio_typing.plugin mypy plugin provides:

• Argument type checking for functions decorated with @asynccontextmanager (either the one in async_generator or the one in 3.7+ contextlib) and @async_generator

• Inference of more specific trio.open_file() and trio.Path.open() return types based on constant mode and buffering arguments, so await trio.open_file("foo", "rb", 0) returns an unbuffered async file object in binary mode and await trio.open_file("bar") returns an async file object in text mode

• Signature checking for task_status.started() with no arguments, so it raises an error if the task_status object is not of type TaskStatus[None]

• Boilerplate reduction for functions that take parameters (fn, *args) and ultimately invoke fn(*args): just write

  @trio_typing.takes_callable_and_args
  def start_soon(
      async_fn: Callable[[trio_typing.ArgsForCallable], Awaitable[T]],
      *args: ArgsForCallable,
      other_keywords: str = are_ok_too,
  ):
      # your implementation here
  

  start_soon(async_fn, *args) will raise an error if async_fn(*args) would do so. You can also make the callable take some non-splatted arguments; the *args get inserted at whatever position in the argument list you write ArgsForCallable.

  Note: due to mypy limitations, we only support a maximum of 5 positional arguments, and keyword arguments can't be passed in this way; nursery.start_soon(functools.partial(...)) will pass the type checker but won't be able to actually check the argument types.

• Mostly-full support for type checking @async_generator functions. You write the decorated function as if it returned a union of its actual return type, its yield type wrapped in YieldType[], and its send type wrapped in SendType[]:

  from trio_typing import YieldType, SendType
  @async_generator
  async def sleep_and_sqrt() -> Union[None, SendType[int], YieldType[float]]:
      next_yield = 0.0
      while True:
          amount = await yield_(next_yield)  # amount is an int
          if amount < 0:
              return None
          await trio.sleep(amount)
          next_yield = math.sqrt(amount)
  
  # prints: CompatAsyncGenerator[float, int, None]
  reveal_type(sleep_and_sqrt())
  

  Calls to yield_ and yield_from_ inside an @async_generator function are type-checked based on these declarations. If you leave off either the yield type or send type, the missing one is assumed to be None; if you leave off both (writing just async def sleep_and_sqrt() -> None:, like you would if you weren't using the plugin), they're both assumed to be Any.

  Note the explicit return None; mypy won't accept return or falling off the end of the function, unless you run it with --no-warn-no-return.

• Calls to variadic Trio functions like trio.run(), nursery.start_soon(), and so on, only can type-check up to five positional arguments. (This number could be increased easily, but only at the cost of slower typechecking for everyone; mypy's current architecture requires that we generate overload sets initially for every arity we want to be able to use.) You can work around this with a # type: ignore comment.
• outcome.capture() and outcome.acapture() currently don't typecheck their arguments at all.

trio-typing comes with a fairly extensive testsuite; it doesn't test all the mechanical parts of the stubs, but does exercise most of the interesting plugin behavior. You can run it after installing, with:

pytest -p trio_typing._tests.datadriven --pyargs trio_typing

Your choice of MIT or Apache 2.0.