Welcome! typeshed is a community project that aims to work for a wide range of Python users and Python codebases. If you're trying a type checker on your Python code, your experience and what you can contribute are important to the project's success.
- Prepare your environment.
- Find out where to make your changes.
- Prepare your changes:
- Small fixes and additions can be submitted directly as pull requests, but contact us before starting significant work.
- Create your stubs, considering what to include and conforming to the coding style.
- Optionally format and check your stubs.
- Optionally run the tests.
- Submit your changes by opening a pull request.
- Make sure that all tests in CI are passing.
You can expect a reply within a few days, but please be patient when it takes a bit longer. For more details, read below.
Typeshed runs continuous integration (CI) on all pull requests. This means that if you file a pull request (PR), our full test suite -- including our linter, Flake8 -- is run on your PR. It also means that bots will automatically apply changes to your PR (using Black and Ruff) to fix any formatting issues. This frees you up to ignore all local setup on your side, focus on the code and rely on the CI to fix everything, or point you to the places that need fixing.
If you prefer to run the tests and formatting locally, it's possible too. Follow platform-specific instructions below. For more information about our available tests, see tests/README.md.
Whichever platform you're using, you will need a virtual environment. If you're not familiar with what it is and how it works, please refer to this documentation.
Note that some tests require extra setup steps to install the required dependencies.
On Linux and Mac OS, you will be able to run the full test suite on Python 3.9-3.12. To install the necessary requirements, run the following commands from a terminal window:
$ python3 -m venv .venv
$ source .venv/bin/activate
(.venv)$ pip install -U pip
(.venv)$ pip install -r requirements-tests.txt
Run the following commands from a Windows terminal to install all requirements:
> python -m venv .venv
> .venv\Scripts\activate
(.venv) > pip install -U pip
(.venv) > pip install -r "requirements-tests.txt"
To be able to run pytype tests, you'll also need to install it manually as it's currently excluded from the requirements file:
(.venv) > pip install -U pytype
The code is formatted using Black
.
Various other autofixes and lint rules are
also performed by Ruff
and
Flake8
,
with plugins flake8-pyi
,
and flake8-noqa
.
The repository is equipped with a pre-commit.ci configuration file. This means that you don't need to do anything yourself to run the code formatters or linters. When you push a commit, a bot will run those for you right away and add any autofixes to your PR. Anything that can't be autofixed will show up as a CI failure, hopefully with an error message that will make it clear what's gone wrong.
That being said, if you want to run the formatters and linters locally when you commit, you're free to do so. To use the same configuration as we use in CI, we recommend doing this via pre-commit:
(.venv)$ pre-commit run --all-files
The stdlib
directory contains stubs for modules in the
Python standard library — which
includes pure Python modules, dynamically loaded extension modules,
hard-linked extension modules, and the builtins. The VERSIONS
file lists
the versions of Python where the module is available.
We accept changes for future versions of Python after the first beta for that version was released. We drop support for a Python version three months after it reaches end-of-life. This means that we will no longer actively test the stubs against that version. After six months, we will remove the stubs for that version and start to use syntax and typing features not supported by that version.
We accept stubs for third-party packages into typeshed as long as:
- the package is publicly available on the Python Package Index;
- the package supports any Python version supported by typeshed; and
- the package does not ship with its own stubs or type annotations.
The fastest way to generate new stubs is to use scripts/create_baseline_stubs.py
(see below).
Stubs for third-party packages go into the stubs
directory. Each subdirectory
there represents a PyPI distribution, and contains the following:
METADATA.toml
, describing the package. See below for details.- Stubs (i.e.
*.pyi
files) for packages and modules that are shipped in the source distribution. - (Rarely) some docs specific to a given type stub package in
README
file.
When a third party stub is added or
modified, an updated version of the corresponding distribution will be
automatically uploaded to PyPI within a few hours.
Each time this happens the least significant
version level is incremented. For example, if stubs/foo/METADATA.toml
has
version = "x.y"
the package on PyPI will be updated from types-foo-x.y.n
to types-foo-x.y.n+1
.
Note: In its current implementation, typeshed cannot contain stubs for multiple versions of the same third-party library. Prefer to generate stubs for the latest version released on PyPI at the time of your stubbing.
The metadata file describes the stubs package using the TOML file format. Currently, the following keys are supported:
version
: The versions of the library that the stubs support. Two formats are supported:- A concrete version. This is especially suited for libraries that use Calendar Versioning.
- A version range ending in
.*
. This is suited for libraries that reflect API changes in the version number only, where the API-independent part is represented by the asterisk. In the case of Semantic Versioning, this version could look like this:2.7.*
. When the stubs are updated to a newer version of the library, the version of the stub should be bumped (note that previous versions are still available on PyPI).
requires
(optional): A list of other stub packages or packages with type information that are imported by the stubs in this package. Only packages generated by typeshed or required by the upstream package are allowed to be listed here, for security reasons. See this issue for more information about what external dependencies are allowed.extra_description
(optional): Can be used to add a custom description to the package's long description. It should be a multi-line string in Markdown format.stub_distribution
(optional): Distribution name to be uploaded to PyPI. This defaults totypes-<distribution>
and should only be set in special cases.upstream_repository
(recommended): The URL of the upstream repository.obsolete_since
(optional): This field is part of our process for removing obsolete third-party libraries. It contains the first version of the corresponding library that ships its ownpy.typed
file.no_longer_updated
(optional): This field is set totrue
before removing stubs for other reasons than the upstream library shipping with type information.upload
(optional): This field is set tofalse
to prevent automatic uploads to PyPI. This should only used in special cases, e.g. when the stubs break the upload.partial_stub
(optional): This field marks the type stub package as partial. This is for 3rd-party stubs that don't cover the entirety of the package's public API.requires_python
(optional): The minimum version of Python required to install the type stub package. It must be in the form>=3.*
. If omitted, the oldest Python version supported by typeshed is used.
In addition, we specify configuration for stubtest in the tool.stubtest
table.
This has the following keys:
skip
(default:false
): Whether stubtest should be run against this package. Please avoid setting this totrue
, and add a comment if you have to.ignore_missing_stub
: When set totrue
, this will add the--ignore_missing_stub
option to the stubtest call. See tests/README.md for more information. In most cases, this field should be identical topartial_stub
.stubtest_requirements
(default:[]
): A list of Python packages that need to be installed for stubtest to run successfully. These packages are installed in addition to the requirements in therequires
field.apt_dependencies
(default:[]
): A list of Ubuntu APT packages that need to be installed for stubtest to run successfully.brew_dependencies
(default:[]
): A list of MacOS Homebrew packages that need to be installed for stubtest to run successfullychoco_dependencies
(default:[]
): A list of Windows Chocolatey packages that need to be installed for stubtest to run successfullyplatforms
(default:["linux"]
): A list of OSes on which to run stubtest. Can containwin32
,linux
, anddarwin
values. If not specified, stubtest is run only onlinux
. Only add extra OSes to the test if there are platform-specific branches in a stubs package.
*_dependencies
are usually packages needed to pip install
the implementation
distribution.
The format of all METADATA.toml
files can be checked by running
python3 ./tests/check_typeshed_structure.py
.
If your change will be a significant amount of work to write, we highly recommend starting by opening an issue laying out what you want to do. That lets a conversation happen early in case other contributors disagree with what you'd like to do or have ideas that will help you do it.
Each Python module is represented by a .pyi "stub file". This is a syntactically valid Python file, where all methods are empty and type annotations are used to describe function signatures and variable types.
Typeshed follows the standard type system guidelines for stub content and coding style.
Stubs should include the complete interface (classes, functions, constants, etc.) of the module they cover, but it is not always clear exactly what is part of the interface.
The following should always be included:
- All objects listed in the module's documentation.
- All objects included in
__all__
(if present).
Other objects may be included if they are being used in practice
or if they are not prefixed with an underscore. This means
that typeshed will generally accept contributions that add missing
objects, even if they are undocumented. Undocumented objects should
be marked with a comment of the form # undocumented
.
When submitting new stubs, it is not necessary to annotate all arguments,
return types, and fields. Such items should either be left unannotated or
use _typeshed.Incomplete
if this is not possible:
from _typeshed import Incomplete
field: Incomplete # unannotated
def foo(x): ... # unannotated argument and return type
Incomplete
can also be used for partially known types:
def foo(x: Incomplete | None = None) -> list[Incomplete]: ...
Type stubs are meant to be external type annotations for a given library. While they are useful documentation in its own merit, they augment the project's concrete implementation, not the project's documentation. Whenever you find them disagreeing, model the type information after the actual implementation and file an issue on the project's tracker to fix their documentation.
Typeshed stubs should not include duplicated docstrings from the source code.
Typeshed includes scripts/create_baseline_stubs.py
.
It generates stubs automatically using a tool called
stubgen that comes with mypy.
To get started, fork typeshed, clone your fork, and then
create a virtualenv.
You can then install the library with pip
into the virtualenv and run the script below,
replacing $INSERT_LIBRARY_NAME_HERE
with the name of the library:
(.venv)$ pip install $INSERT_LIBRARY_NAME_HERE
(.venv)$ python3 scripts/create_baseline_stubs.py $INSERT_LIBRARY_NAME_HERE
When the script has finished running, it will print instructions telling you what to do next.
If it has been a while since you set up the virtualenv, make sure you have
the latest mypy (pip install -r requirements-tests.txt
) before running the script.
Since PEP 484 was accepted, there have been many other PEPs that added new features to the Python type system. In general, new features can be used in typeshed as soon as the PEP has been accepted and implemented and most type checkers support the new feature.
Supported features include:
- PEP 544 (Protocol)
- PEP 585 (builtin generics)
- PEP 586 (Literal)
- PEP 591 (Final/@final)
- PEP 589 (TypedDict)
- PEP 604 (
Foo | Bar
union syntax) - PEP 612 (ParamSpec)
- PEP 647 (TypeGuard): see #5406
- PEP 655 (
Required
andNotRequired
) - PEP 673 (
Self
) - PEP 675 (
LiteralString
) - PEP 702 (
@deprecated()
)
Features from the typing
module that are not present in all
supported Python versions must be imported from typing_extensions
instead in typeshed stubs. This currently affects:
TypeAlias
(new in Python 3.10)Concatenate
(new in Python 3.10)ParamSpec
(new in Python 3.10)TypeGuard
(new in Python 3.10)Self
(new in Python 3.11)Never
(new in Python 3.11)LiteralString
(new in Python 3.11)TypeVarTuple
andUnpack
(new in Python 3.11)Required
andNotRequired
(new in Python 3.11)Buffer
(new in Python 3.12; in thecollections.abc
module)@deprecated
(new in Python 3.13; in thewarnings
module)
Some type checkers implicitly promote the bytearray
and
memoryview
types to bytes
.
PEP 688 removes
this implicit promotion.
Typeshed stubs should be written assuming that these promotions
do not happen, so a parameter that accepts either bytes
or
bytearray
should be typed as bytes | bytearray
.
Often one of the aliases from _typeshed
, such as
_typeshed.ReadableBuffer
, can be used instead.
Even more excellent than a good bug report is a fix for a bug, or the implementation of a much-needed stub. We'd love to have your contributions.
We use the usual GitHub pull-request flow, which may be familiar to you if you've contributed to other projects on GitHub. For the mechanics, see Mypy's git and GitHub workflow help page, or GitHub's own documentation.
Anyone interested in type stubs may review your code. One of the maintainers will merge your pull request when they think it's ready. For every pull request, we aim to promptly either merge it or say why it's not yet ready; if you go a few days without a reply, please feel free to ping the thread by adding a new comment.
To get your pull request merged sooner, you should explain why you are making the change. For example, you can point to a code sample that is processed incorrectly by a type checker. It is also helpful to add links to online documentation or to the implementation of the code you are changing.
As the author of the pull request, it is your responsibility to make sure all CI tests pass and that any feedback is addressed. The typeshed maintainers will probably provide some help and may even push changes to your PR to fix any minor issues, but this is not always possible. If a PR lingers with unresolved problems for too long, we may close it (see below).
Also, do not squash your commits or use git commit --amend
after you have submitted a pull request, as this
erases context during review. We will squash commits when the pull request is merged.
This way, your pull request will appear as a single commit in our git history, even
if it consisted of several smaller commits.
Third-party stubs are generally removed from typeshed when one of the following criteria is met:
- The upstream package ships a
py.typed
file for at least six months, and the upstream type annotations are of a comparable standard to those in typeshed, or - the upstream package was declared or appears to be unmaintained, and retaining the stubs causes maintenance issues in typeshed.
If a package ships its own py.typed
file, please follow these steps:
- Open an issue with the earliest month of removal in the subject.
- A maintainer will add the "stubs: removal" label.
- Open a PR that sets the
obsolete_since
field in theMETADATA.toml
file to the first version of the package that shippedpy.typed
. - After at least six months, open a PR to remove the stubs.
If third-party stubs should be removed for other reasons, please follow these steps:
- Open an issue explaining why the stubs should be removed.
- A maintainer will add the "stubs: removal" label.
- Open a PR that sets the
no_longer_updated
field in theMETADATA.toml
file totrue
. - When a new version of the package was automatically uploaded to PyPI (which can take up to a day), open a PR to remove the stubs.
If feeling kindly, please update mypy for any stub obsoletions or removals.
The process for preparing and submitting changes also applies to maintainers. This ensures high quality contributions and keeps everybody on the same page. Avoid direct pushes to the repository.
When reviewing pull requests, follow these guidelines:
- Typing is hard. Try to be helpful and explain issues with the PR, especially to new contributors.
- When reviewing auto-generated stubs, just scan for red flags and obvious errors. Leave possible manual improvements for separate PRs.
- When reviewing large, hand-crafted PRs, you only need to look for red flags and general issues, and do a few spot checks.
- Review smaller, hand-crafted PRs thoroughly.
When merging pull requests, follow these guidelines:
- Always wait for tests to pass before merging PRs.
- Use "Squash and merge" to merge PRs.
- Make sure the commit message is meaningful. For example, remove irrelevant intermediate commit messages.
- The commit message for third-party stubs is used to generate the changelog. It should be valid Markdown, be comprehensive, read like a changelog entry, and assume that the reader has no access to the diff.
- Delete branches for merged PRs (by maintainers pushing to the main repo).
We sometimes use the "status: deferred" label to mark PRs and issues that we'd like to accept, but that are blocked by some external factor. Blockers can include:
- An unambiguous bug in a type checker (i.e., a case where the type checker is not implementing the typing spec).
- A dependency on a typing PEP that is still under consideration.
- A pending change in a related project, such as stub-uploader.
PRs should only be marked as "deferred" if there is a clear path towards getting the blocking issue resolved within a reasonable time frame. If a PR depends on a more amorphous change, such as a type system change that has not yet reached the PEP stage, it should instead be closed.
Maintainers who add the "deferred" label should state clearly what exactly the blocker is, usually with a link to an open issue in another project.
To keep the number of open PRs manageable, we may close PRs when they have been open for too long. Specifically, we close open PRs that either have failures in CI, serious merge conflicts or unaddressed feedback, and that have not seen any activity in three months.
We want to maintain a welcoming atmosphere for contributors, so use a friendly message when closing the PR. Example message:
Thanks for contributing! I'm closing this PR for now, because it still
<fails some tests OR has unresolved review feedback OR has a merge conflict>
after three months of inactivity. If you are still interested, please feel free to open
a new PR (or ping us to reopen this one).