Extending APIs¤

Griffe has an extension system that can be used to enhance or customize the data that Griffe collects. Extensions are written in Python.

Using extensions¤

Extensions can be specified both on the command-line (in the terminal), and programmatically (in Python).

On the command-line¤

On the command-line, you can specify extensions to use with the -e, --extensions option. This option accepts a single positional argument which can take two forms:

a comma-separated list of extensions
a JSON list of extensions

Extensions can accept options: the comma-separated list does not allow to specify options, while the JSON list does. See examples below.

With both forms, each extension refers to one of these three things:

the name of a built-in extension's module, for example dynamic_docstrings (this is just an example, this built-in extension does not exist)
the Python dotted-path to a module containing one or more extensions, or to an extension directly, for example package.module and package.module.ThisExtension
the file path to a Python script, and an optional extension name, separated by a colon, for example scripts/griffe_exts.py and scripts/griffe_exts.py:ThisExtension

The specified extension modules can contain more than one extension: Griffe will pick up and load every extension declared or imported within the modules. If options are specified for a module that contains multiple extensions, the same options will be passed to all the extensions, so extension writers must make sure that all extensions within a single module accept the same options. If they don't, Griffe will abort with an error.

To specify options in the JSON form, use a dictionary instead of a string: the dictionary's only key is the extension identifier (built-in name, Python path, file path) and its value is a dictionary of options.

Some examples:

griffe dump griffe -e pydantic,scripts/exts.py:DynamicDocstrings,griffe_attrs

griffe check --search src griffe -e '[
  {"pydantic": {"schema": true}},
  {
    "scripts/exts.py:DynamicDocstrings": {
      "paths": ["mypkg.mymod.myobj"]
    }
  },
  "griffe_attrs"
]'

In the above two examples, pydantic would be a built-in extension, scripts/exts.py:DynamicDocstrings the file path plus name of a local extension, and griffe_attrs the name of a third-party package that exposes one or more extensions.

Programmatically¤

Within Python code, extensions can be specified with the extensions parameter of the GriffeLoader class or load function.

The parameter accepts an instance of the Extensions class. Such an instance is created with the help of the load_extensions function, which itself accepts a list of strings, dictionaries, extension classes and extension instances.

Strings and dictionaries are used the same way as on the command-line. Extension instances are used as such, and extension classes are instantiated without any options.

Example:

import griffe

from mypackage.extensions import ThisExtension, ThisOtherExtension

extensions = griffe.load_extensions(
    {"pydantic": {"schema": true}},
    {"scripts/exts.py:DynamicDocstrings": {"paths": ["mypkg.mymod.myobj"]}},
    "griffe_attrs",
    ThisExtension(option="value"),
    ThisOtherExtension,
)

data = griffe.load("mypackage", extensions=extensions)

In MkDocs¤

MkDocs and its mkdocstrings plugin can be configured to use Griffe extensions:

mkdocs.yml

plugins:
- mkdocstrings:
    handlers:
      python:
        options:
          extensions:
          - pydantic: {schema: true}
          - scripts/exts.py:DynamicDocstrings:
              paths: [mypkg.mymod.myobj]
          - griffe_attrs

The extensions key accepts a list that is passed to the load_extensions function. See how to use extensions programmatically to learn more.

Writing extensions¤

In the next section we give a bit of context on how Griffe works, to show how extensions can integrate into the data collection process. Feel free to skip to the Events and hooks section or the Full example section if you'd prefer to see concrete examples first.

How it works¤

To extract information from your Python sources, Griffe tries to build Abstract Syntax Trees by parsing the sources with ast utilities.

If the source code is not available (the modules are built-in or compiled), Griffe imports the modules and builds object trees instead.

Griffe then follows the Visitor pattern to walk the tree and extract information. For ASTs, Griffe uses its Visitor agent and for object trees, it uses its Inspector agent.

Sometimes during the walk through the tree (depth-first order), both the visitor and inspector agents will trigger events. These events can be hooked on by extensions to alter or enhance Griffe's behavior. Some hooks will be passed just the current node being visited, others will be passed both the node and an instance of an Object subclass, such as a Module, a Class, a Function, an Attribute, or a Type Alias. Extensions will therefore be able to modify these instances.

The following flow chart shows an example of an AST visit. The tree is simplified: actual trees have a lot more nodes like if/elif/else nodes, try/except/else/finally nodes, and many more.

flowchart TB
M(Module definition) --- C(Class definition) & F(Function definition)
C --- m(Function definition) & A(Variable assignment)

The following flow chart shows an example of an object tree inspection. The tree is simplified as well: many more types of objects are handled.

flowchart TB
M(Module) --- C(Class) & F(Function)
C --- m(Method) & A(Attribute)

For a more concrete example, let say that we visit (or inspect) an AST (or object tree) for a given module, and that this module contains a single class, which itself contains a single method:

the agent (visitor or inspector) will walk through the tree by starting with the module node
it will instantiate a Module, then walk through its members, continuing with the class node
it will instantiate a Class, then walk through its members, continuing with the function node
it will instantiate a Function
then it will go back up and finish walking since there are no more nodes to walk through

Every time the agent enters a node, creates an object instance, or finishes handling members of an object, it will trigger an event.

The flow of events is drawn in the following flowchart:

flowchart TB
visit_mod{{enter module node}}
event_mod_node{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_node'><b><code style='color: var(--md-accent-fg-color)'>on_node</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_module_node'><b><code style='color: var(--md-accent-fg-color)'>on_module_node</code></b></a> event"}}
create_mod{{create module instance}}
event_mod_instance{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_instance'><b><code style='color: var(--md-accent-fg-color)'>on_instance</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_module_instance'><b><code style='color: var(--md-accent-fg-color)'>on_module_instance</code></b></a> event"}}
visit_mod_members{{visit module members}}
visit_cls{{enter class node}}
event_cls_node{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_node'><b><code style='color: var(--md-accent-fg-color)'>on_node</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_class_node'><b><code style='color: var(--md-accent-fg-color)'>on_class_node</code></b></a> event"}}
create_cls{{create class instance}}
event_cls_instance{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_instance'><b><code style='color: var(--md-accent-fg-color)'>on_instance</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_class_instance'><b><code style='color: var(--md-accent-fg-color)'>on_class_instance</code></b></a> event"}}
visit_cls_members{{visit class members}}
visit_func{{enter func node}}
event_func_node{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_node'><b><code style='color: var(--md-accent-fg-color)'>on_node</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_function_node'><b><code style='color: var(--md-accent-fg-color)'>on_function_node</code></b></a> event"}}
create_func{{create function instance}}
event_func_instance{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_instance'><b><code style='color: var(--md-accent-fg-color)'>on_instance</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_function_instance'><b><code style='color: var(--md-accent-fg-color)'>on_function_instance</code></b></a> event"}}
event_cls_members{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_members'><b><code style='color: var(--md-accent-fg-color)'>on_members</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_class_members'><b><code style='color: var(--md-accent-fg-color)'>on_class_members</code></b></a> event"}}
event_mod_members{{"<a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_members'><b><code style='color: var(--md-accent-fg-color)'>on_members</code></b></a> event<br><a href='/griffe/reference/griffe/extensions/#griffe.Extension.on_module_members'><b><code style='color: var(--md-accent-fg-color)'>on_module_members</code></b></a> event"}}

start{start} --> visit_mod
visit_mod --> event_mod_node
event_mod_node --> create_mod
create_mod --> event_mod_instance
event_mod_instance --> visit_mod_members
visit_mod_members --1--> visit_cls
visit_cls --> event_cls_node
event_cls_node --> create_cls
create_cls --> event_cls_instance
event_cls_instance --> visit_cls_members
visit_cls_members --1--> visit_func
visit_func --> event_func_node
event_func_node --> create_func
create_func --> event_func_instance
event_func_instance --> visit_cls_members
visit_cls_members --2--> event_cls_members
event_cls_members --> visit_mod_members
visit_mod_members --2--> event_mod_members
event_mod_members --> finish{finish}

class event_mod_node event
class event_mod_instance event
class event_cls_node event
class event_cls_instance event
class event_func_node event
class event_func_instance event
class event_cls_members event
class event_mod_members event
classDef event stroke:#3cc,stroke-width:2

Hopefully this flowchart gives you a pretty good idea of what happens when Griffe collects data from a Python module. The next section will explain in more details the different events that are triggered, and how to hook onto them in your extensions.

Events and hooks¤

There are two kinds of events in Griffe: load events and analysis events. Load events are scoped to the Griffe loader. Analysis events are scoped to the visitor and inspector agents (triggered during static and dynamic analysis).

Load events¤

There are two load events:

on_package_loaded: The "on package loaded" event is triggered when the loader has finished loading a package entirely, i.e. when all its submodules were scanned and loaded. This event can be hooked by extensions which require the whole package to be loaded, to be able to navigate the object tree without raising lookup errors or alias resolution errors.
on_wildcard_expansion: The "on wildcard expansion" event is triggered for each alias that is created by expanding wildcard imports (from ... import *).

Analysis events¤

There are 3 generic analysis events:

on_node: The "on node" events are triggered when the agent (visitor or inspector) starts handling a node in the tree (AST or object tree).
on_instance: The "on instance" events are triggered when the agent just created an instance of Module, Class, Function, Attribute, or Type Alias, and added it as a member of its parent. The "on instance" event is not triggered when an Alias is created.
on_members: The "on members" events are triggered when the agent just finished handling all the members of an object. Functions, attributes and type aliases do not have members, so there are no "on members" events for these kinds.

There are also specific analysis events for each object kind:

And a special event for aliases:

on_alias: The "on alias" event is triggered when an Alias was just created and added as a member of its parent object.

Hooks are methods that are called when a particular event is triggered. To target a specific event, the hook must be named after it.

Extensions are classes that inherit from Griffe's Extension base class and define some hooks as methods:

import ast
import griffe


class MyExtension(griffe.Extension):
    def on_instance(
        self,
        node: ast.AST | griffe.ObjectNode,
        obj: griffe.Object,
        agent: griffe.Visitor | griffe.Inspector,
        **kwargs,
    ) -> None:
        """Do something with `node` and/or `obj`."""

Hooks are always defined as methods of a class inheriting from Extension, never as standalone functions. IDEs should autocomplete the signature when you start typing def followed by a hook name.

Since hooks are declared in a class, feel free to also declare state variables (or any other variable) in the __init__ method:

import ast
from griffe import Extension, Object, ObjectNode


class MyExtension(Extension):
    def __init__(self) -> None:
        super().__init__()
        self.state_thingy = "initial stuff"
        self.list_of_things = []

    def on_instance(
        self,
        node: ast.AST | griffe.ObjectNode,
        obj: griffe.Object,
        agent: griffe.Visitor | griffe.Inspector,
        **kwargs,
    ) -> None:
        """Do something with `node` and/or `obj`."""

Static/dynamic support¤

Extensions can support both static and dynamic analysis of modules. If a module is scanned statically, your extension hooks will receive AST nodes (from the ast module of the standard library). If the module is scanned dynamically, your extension hooks will receive object nodes. Similarly, your hooks will receive a reference to the analysis agent that calls them, either a Visitor or an Inspector.

To support static analysis, dynamic analysis, or both, you can therefore check the type of the received node or agent:

import ast
import griffe


class MyExtension(griffe.Extension):
    def on_instance(
        self,
        node: ast.AST | griffe.ObjectNode,
        obj: griffe.Object,
        agent: griffe.Visitor | griffe.Inspector,
        **kwargs,
    ) -> None:
        """Do something with `node` and/or `obj`."""
        if isinstance(node, ast.AST):
            ...  # Apply logic for static analysis.
        else:
            ...  # Apply logic for dynamic analysis.

import ast
import griffe


class MyExtension(Extension):
    def on_instance(
        self,
        node: ast.AST | griffe.ObjectNode,
        obj: griffe.Object,
        agent: griffe.Visitor | griffe.Inspector,
        **kwargs,
    ) -> None:
        """Do something with `node` and/or `obj`."""
        if isinstance(agent, griffe.Visitor):
            ...  # Apply logic for static analysis.
        else:
            ...  # Apply logic for dynamic analysis.

The preferred method is to check the type of the received node rather than the agent.

Since hooks also receive instantiated modules, classes, functions, attributes and type aliases, most of the time you will not need to use the node argument other than for checking its type and deciding what to do based on the result. And since we always add **kwargs to the hooks' signatures, you can drop any parameter you don't use from the signature:

import griffe


class MyExtension(Extension):
    def on_instance(self, obj: griffe.Object, **kwargs) -> None:
        """Do something with `obj`."""
        ...

Visiting trees¤

Extensions provide basic functionality to help you visit trees:

visit: call self.visit(node) to start visiting an abstract syntax tree.
generic_visit: call self.generic_visit(node) to visit each subnode of a given node.
inspect: call self.inspect(node) to start visiting an object tree. Nodes contain references to the runtime objects, see ObjectNode.
generic_inspect: call self.generic_inspect(node) to visit each subnode of a given node.

Calling self.visit(node) or self.inspect(node) will do nothing unless you actually implement methods that handle specific types of nodes:

for ASTs, methods must be named visit_<node_type> where <node_type> is replaced with the lowercase name of the node's class. For example, to allow visiting ClassDef nodes, you must implement the visit_classdef method:

import ast
from griffe import Extension


class MyExtension(Extension):
    def visit_classdef(node: ast.ClassDef) -> None:
        # Do something with the node...
        ...
        # ...then visit the subnodes
        # (it only makes sense if you implement other methods
        # such as visit_functiondef or visit_assign for example).
        self.generic_visit(node)

See the list of existing AST classes to learn what method you can implement.

for object trees, methods must be named inspect_<node_type>, where <node_type> is replaced with the string value of the node's kind. The different kinds are listed in the ObjectKind enumeration. For example, to allow inspecting coroutine nodes, you must implement the inspect_coroutine method:

from griffe import Extension, ObjectNode


class MyExtension(Extension):
    def inspect_coroutine(node: ObjectNode) -> None:
        # Do something with the node...
        ...
        # ...then visit the subnodes if it makes sense.
        self.generic_inspect(node)

Triggering other extensions¤

If your extension creates new objects, you might want to trigger the other enabled extensions on these object instances. To do this you can use agent.extensions.call:

import ast
import griffe


class MyExtension(griffe.Extension):
    def on_node(self, node: ast.AST | griffe.ObjectNode, agent: griffe.Visitor | griffe.Inspector, **kwargs) -> None:
        # New object created for whatever reason.
        function = griffe.Function(...)

        # Trigger other extensions.
        agent.extensions.call("on_function_instance", node=node, agent=agent, func=function, **kwargs)

Extra data¤

All Griffe objects (modules, classes, functions, attributes, type aliases) can store additional (meta)data in their extra attribute. This attribute is a dictionary of dictionaries. The first layer is used as namespacing: each extension writes into its own namespace, or integrates with other projects by reading/writing in their namespaces, according to what they support and document.

import griffe

self_namespace = "my_extension"


class MyExtension(griffe.Extension):
    def on_instance(self, obj: griffe.Object, **kwargs) -> None:
        obj.extra[self_namespace]["some_key"] = "some_value"

For example, mkdocstrings-python looks into the mkdocstrings namespace for a template key. Extensions can therefore provide a custom template value by writing into extra["mkdocstrings"]["template"]:

import griffe

self_namespace = "my_extension"
mkdocstrings_namespace = "mkdocstrings"


class MyExtension(griffe.Extension):
    def on_class_instance(self, cls: griffe.Class, **kwargs) -> None:
        obj.extra[mkdocstrings_namespace]["template"] = "my_custom_template"

Options¤

Extensions can be made to support options. These options can then be passed from the command-line using JSON, from Python directly, or from other tools like MkDocs, in mkdocs.yml.

import griffe


class MyExtension(griffe.Extension):
    def __init__(self, option1: str, option2: bool = False) -> None:
        super().__init__()
        self.option1 = option1
        self.option2 = option2

    def on_attribute_instance(self, attr: griffe.Attribute, **kwargs) -> None:
        if self.option2:
            ...  # Do something.

Logging¤

To better integrate with Griffe and other tools in the ecosystem (notably MkDocs), use Griffe loggers to log messages:

import griffe

logger = griffe.get_logger(__name__)


class MyExtension(griffe.Extension):
    def on_module_members(self, mod: griffe.Module, **kwargs) -> None:
        logger.info("Doing some work on module %s and its members", mod.path)

Full example¤

The following example shows how one could write a "dynamic docstrings" extension that dynamically imports objects that declare their docstrings dynamically, to improve support for such docstrings. The extension is configurable to run only on user-selected objects.

Package structure (or just write your extension in a local script):

📁 ./
├──  pyproject.toml
└── 📁 src/
    └── 📁 dynamic_docstrings/
        ├──  __init__.py
        └──  extension.py

./src/dynamic_docstrings/extension.py

import ast
import inspect
import griffe

logger = griffe.get_logger(__name__)


class DynamicDocstrings(griffe.Extension):
    def __init__(self, object_paths: list[str] | None = None) -> None:
        self.object_paths = object_paths

    def on_instance(
        self,
        node: ast.AST | griffe.ObjectNode,
        obj: griffe.Object,
        agent: griffe.Visitor | griffe.Inspector,
        **kwargs,
    ) -> None:
        if isinstance(node, griffe.ObjectNode):
            return  # Skip runtime objects, their docstrings are already right.

        if self.object_paths and obj.path not in self.object_paths:
            return  # Skip objects that were not selected.

        # Import object to get its evaluated docstring.
        try:
            runtime_obj = griffe.dynamic_import(obj.path)
            docstring = runtime_obj.__doc__
        except ImportError:
            logger.debug(f"Could not get dynamic docstring for {obj.path}")
            return
        except AttributeError:
            logger.debug(f"Object {obj.path} does not have a __doc__ attribute")
            return

        # Update the object instance with the evaluated docstring.
        docstring = inspect.cleandoc(docstring)
        if obj.docstring:
            obj.docstring.value = docstring
        else:
            obj.docstring = griffe.Docstring(
                docstring,
                parent=obj,
                docstring_parser=agent.docstring_parser,
                docstring_options=agent.docstring_options,
            )

You can then expose this extension in the top-level module of your package:

./src/dynamic_docstrings/__init__.py

from dynamic_docstrings.extension import DynamicDocstrings

__all__ = ["DynamicDocstrings"]

This will allow users to load and use this extension by referring to it as dynamic_docstrings (your Python package name).

See how to use extensions to learn more about how to load and use your new extension.

AST nodes¤

Add

alias

And

AnnAssign

arg

arguments

Assert

Assign

AsyncFor

AsyncFunctionDef

AsyncWith

Attribute

AugAssign

Await

BinOp

BitAnd

BitOr

BitXor

BoolOp

Break

Bytes¹

Call

ClassDef

Compare

comprehension

Constant

Continue

Del

Delete

Dict

DictComp

Div

Ellipsis¹

Eq

ExceptHandler

Expr

Expression¹

ExtSlice²

FloorDiv

For

FormattedValue

FunctionDef

GeneratorExp

Global

Gt

GtE

If

IfExp

Import

ImportFrom

In

Index²

Interactive³

Invert

Is

IsNot

JoinedStr

keyword

Lambda

List

ListComp

Load

LShift

Lt

LtE

Match

MatchAs

match_case

MatchClass

MatchMapping

MatchOr

MatchSequence

MatchSingleton

MatchStar

MatchValue

MatMult

Mod

Module³

Mult

Name

NameConstant¹

NamedExpr

Nonlocal

Not

NotEq

NotIn

Num¹

Or

ParamSpec

Pass

pattern³

Pow

Print⁴

Raise

Return

RShift

Set

SetComp

Slice

Starred

Store

Str¹

Sub

Subscript

Try

TryExcept⁵

TryFinally⁶

Tuple

TypeAlias

TypeVar

TypeVarTuple

UAdd

UnaryOp

USub

While

With

withitem

Yield

YieldFrom

Next steps¤

Extensions are a powerful mechanism to customize or enhance the data loaded by Griffe. But sometimes, all you need to do to improve the data is to make Griffe happy by following a few conventions. We therefore invite you to read our recommendations on public APIs, Python code best practices and docstrings.

Deprecated since Python 3.8. ↩↩↩↩↩↩
Deprecated since Python 3.9. ↩↩
Not documented. ↩↩↩
print became a builtin (instead of a keyword) in Python 3. ↩
Now ExceptHandler, in the handlers attribute of Try nodes. ↩
Now a list of expressions in the finalbody attribute of Try nodes. ↩