fastapi-injected

Yet another attempt to reuse FastAPI's dependency injection outside of request handlers.

This is an opinionated library: it takes the DI machinery you already know from FastAPI (Depends, generator dependencies with teardown, dependency caching) and makes it usable in plain async functions — background jobs, CLI commands, workers, scripts — without a Request in sight.

Installation

pip install fastapi-injected

Requires Python 3.12+.

Usage

Declare dependencies as regular classes and annotate fields with Dep[...]:

from dataclasses import dataclass
from typing import AsyncIterator

from fastapi_injected import Dep, DepFactory, Injected, inject


@dataclass
class Session:
    closed: bool = False


async def session_dep() -> AsyncIterator[Session]:
    session = Session()
    try:
        yield session
    finally:
        session.closed = True


@dataclass
class Repository:
    session: DepFactory[Session, session_dep]


@dataclass
class Service:
    repo: Dep[Repository]


@inject
async def handler(*, service: Dep[Service] = Injected) -> None:
    ...  # service is built and injected, session is closed on exit


await handler()

• Dep[T] — resolve T by calling it, same as FastAPI's Annotated[T, Depends()].
• DepFactory[T, factory] — resolve T via a factory, same as Annotated[T, Depends(factory)]. Generator factories get proper teardown.
• Arg[T] — keep a parameter out of the dependency graph even though its annotation is a dependency, so the caller supplies it. See Caller-supplied arguments.
• Injected — a sentinel default that exists purely to make type checkers happy: without it they would complain about a missing argument at call sites. At runtime the parameter is always filled in by @inject.

Injected parameters mix freely with regular ones — pass your own arguments as usual and the rest is injected:

@inject
async def add(a: int, b: int, *, service: Dep[Service] = Injected) -> int:
    ...


result = await add(1, 2)

Caller-supplied arguments

A parameter is injected only if it is written as a dependency — Dep[...], DepFactory[...], an Annotated[..., Depends(...)] of your own, or a Depends(...) default. Everything else is left for the caller, annotation untouched: it is never handed to pydantic, so any type works with no marker at all.

class Connection:  # not a pydantic-friendly type, and it does not have to be
    ...


@inject
async def handler(conn: Connection, *, service: Dep[Service] = Injected) -> None:
    ...


await handler(connection)

Arg[...] is for the opposite case — a parameter that is spelled as a dependency, in a function that wants it passed in instead:

from fastapi_injected import Arg

type ServiceDep = Annotated[Service, Depends(get_service)]


@inject
async def handler(service: Arg[ServiceDep], *, session: Dep[Session] = Injected) -> None:
    ...


await handler(service)

Arg[T] is a no-op for type checkers — the parameter stays typed as T — and at runtime it tells @inject to keep this parameter out of the dependency graph.

Writing Injected as the default of a parameter that is not a dependency raises NotADependencyError at decoration time, since nothing would ever fill it in.

Resolving a type directly

No decorator needed — resolve a dependency graph on demand:

from fastapi_injected import resolve

service = await resolve(Service)

Like @inject, resolve accepts new_scope=True to force a fresh scope instead of reusing the surrounding one.

Scopes and caching

By default every call to an injected function gets its own scope: dependencies are built, cached within the call, and torn down when it returns. Wrap several calls in push_inject_scope() to share one cache (and defer teardown to the end of the scope):

from fastapi_injected import push_inject_scope

async with push_inject_scope():
    a = await handler()  # dependencies built here
    b = await handler()  # same instances reused
# generator dependencies are torn down here

Use @inject(new_scope=True) to opt a function out of the surrounding scope and always get fresh dependencies. It builds them again, but stays in the request they are being built for, so request-bound dependencies keep resolving.

Outside of a request there is no request to resolve against, so the scope makes one up. It answers what a dependency usually reads — method, url, headers, path_params, client, state — with the values of a request nobody sent. The application is the one thing it cannot invent: pass it when a dependency reaches for request.app:

async with push_inject_scope(app=app):
    await handler()  # `request.app.state` resolves as it does in a route

Without it, reading request.app raises a KeyError naming what is missing rather than a bare 'app'.

Analysing a dependency is the expensive part of resolving one, so the result is cached — keyed by the dependency itself, not by whatever object carried it, so nothing that only passed through is kept alive. clear_dependant_cache() drops it, for long-lived processes and test suites that want the memory back.

Overriding dependencies

push_overrides swaps dependencies out for the duration of a with block — handy in tests, or anywhere you need to run the same code against a different implementation:

from fastapi_injected import push_overrides

with push_overrides({Session: Session(closed=True)}):
    await handler()  # gets the override instead of the real dependency

A key can be the dependency itself, or the annotation you wrote in the signature — Dep[Session] and DepFactory[Session, session_dep] both work and are normalized to the same underlying dependency:

with push_overrides({DepFactory[Session, session_dep]: my_session}):
    ...

Values are used as-is, but two wrappers make the intent explicit and cover the ambiguous cases:

• ValueOverride(value) — always inject value, even when it is itself callable.
• FactoryOverride(factory) — call factory to produce the value. Sync, async, and generator factories are all supported, with the same teardown semantics as regular dependencies.

from fastapi_injected import FactoryOverride, ValueOverride

with push_overrides(
    {
        Session: ValueOverride(fake_session),
        Repository: FactoryOverride(lambda: FakeRepository()),
    },
):
    ...

Overrides apply to the whole graph, not just top-level parameters — overriding a nested dependency changes what its dependents receive. Nested push_overrides blocks merge, with the innermost one winning.

An override block is a scope of its own: it caches what it builds, and that cache goes away with the block. The surrounding scope keeps what it had resolved before, and it stays visible — except for the overridden dependencies and anything built from them, which are resolved again under the override:

async with push_inject_scope():
    real = await handler()  # dependencies cached here

    with push_overrides({Session: fake}):
        await handler()  # rebuilt with `fake`, everything else reused from the cache

    await handler()  # back to the cached, real dependencies

Overrides can also be handed to the scope directly, which is the same thing in one call:

async with push_inject_scope({Session: fake}) as scope:
    await handler()

    with scope.override({Session: other}):  # nest freely
        await handler()

Dependencies that are objects

Dependants are cached by the callable that resolves them, so a dependency that is an object — a class holding configuration, a parametrized resolver — has to be hashable to get there. A plain dataclass is not, and a frozen one still refuses as soon as it holds a list or a dict.

MakeDataclass is a base class that makes its subclasses dataclasses with a hash that always answers: by fields when they can be hashed, by identity when they cannot.

from dataclasses import field
from fastapi_injected import MakeDataclass, resolve


class Settings(MakeDataclass):
    hosts: list[str] = field(default_factory=list)

    def __call__(self) -> list[str]:
        return self.hosts


await resolve(Settings(["a", "b"]))  # a plain dataclass would raise TypeError here

Every dataclasses.dataclass option is accepted as a class keyword and passed straight through:

class Config(MakeDataclass, frozen=True, kw_only=True):
    retries: int = 3

The same is true of what holds a dependency. FastAPI's Depends and Security are frozen dataclasses, so they hash by field — and Security(get_user, scopes=["items:read"]), the documented spelling, holds a list, which makes the whole annotation unhashable and fails as a bare TypeError from typing. fastapi_injected ships both as MakeDataclass subclasses, which is the same class in every way that matters to FastAPI, and one that can be cached:

from fastapi_injected import Security, resolve

type CurrentUser = Annotated[User, Security(get_user, scopes=["items:read"])]

await resolve(CurrentUser)

Scopes are kept as a tuple, so the same scopes written as another list stay one dependency. Everything the package builds internally is held this way too, so a dependency object that cannot be hashed by fields no longer decides whether the annotation around it can be.

When resolution fails

A dependency can fail to resolve for the same reasons it would in a route — a missing header, a query parameter that does not validate. DependencyResolutionError carries those errors in the shape pydantic produced them, and turns into the response FastAPI would have returned:

from fastapi_injected import DependencyResolutionError

try:
    await handler()
except DependencyResolutionError as exc:
    exc.errors  # [{'type': 'missing', 'loc': ('header', 'x-trace'), ...}]
    raise exc.as_validation_error() from exc  # a RequestValidationError, so a 422

It is a ValueError, which is what resolution used to raise, so code catching that keeps working.

Dependencies decided at runtime

A dependency is usually written as an annotation, but sometimes it is a value a caller already has, or one picked while the program runs. Given turns a value into a dependency that resolves to it:

from fastapi_injected import Given, resolve

doc = Doc(title="readme")

await resolve(Given(doc))  # the doc itself

Constants holding equal values are the same dependency, so they share a cache entry and can be overridden like any other. What Given returns is a marker held as a value — DepOf[R], as opposed to Dep[R], which is the same marker in annotation position.

bind_deps binds such markers to the leading parameters of a function, which is how a dependency graph gets built from markers that were not known when the function was written:

from fastapi_injected import bind_deps

async def describe(doc: Doc, role: Role) -> str:
    return f"{doc.title}:{role.name}"


bound = bind_deps(describe, Given(doc), RoleDep)

await resolve(bound)  # "readme:admin", with `role` resolved as usual

What a parameter is bound to can be written any way a dependency can be: a marker like Given(...), Dep[T] or DepFactory[T, factory], a Depends(...), or the callable itself — the same things resolve accepts.

Binding takes the parameters away, and a type checker sees what is left: bind_deps(describe, Given(doc)) is a function of (role), and binding both leaves one of no arguments at all. The bound function is a dependency like any other — resolve it, @inject it, override what it depends on. Binding more markers than the function has parameters is a TypeError rather than a silent truncation.

Dependencies as objects

When the markers belong together, MakeInjected makes a dataclass of them: fields annotated with DepOf hold the markers, and the call receives what they resolve to.

from fastapi_injected import Given, MakeInjected
from fastapi_injected.types import DepOf


class IsOwner(MakeInjected):
    doc: DepOf[Doc]
    role: DepOf[Role]
    loud: bool = False        # a plain field stays a plain field

    async def __call__(self, doc: Doc, role: Role) -> bool:
        return doc.owner == role.name


await resolve(IsOwner(doc=Given(doc), role=RoleDep))

It is a MakeDataclass, so instances built from equal markers are equal — and therefore one dependency, resolved once per scope and overridable as a whole. For callables that are not dataclasses, the two halves of the binding are available on their own: signature_with_deps(func, deps) builds the signature, and remap_dep_args turns the arguments it names back into positional ones.

Inspecting annotations

A few helpers are exported for code that needs to reason about Dep[...] annotations — building override maps, custom decorators, and the like:

• is_dep(tp) — whether tp is a Dep/DepFactory annotation.
• unwrap_dep_tp(tp) — the annotated type (Any for a bare Dep).
• unwrap_dep_dependency(tp) — the callable that resolves it: the factory for DepFactory[T, factory], the type itself for Dep[T].

FastAPI request integration

Inside a FastAPI app, @inject-ed functions and resolve can share the request's own dependency cache — the same instances FastAPI built for the handler. Register init_inject_scope as a dependency:

from fastapi import Depends, FastAPI
from fastapi_injected import Dep, init_inject_scope, resolve

app = FastAPI(dependencies=[Depends(init_inject_scope)])


@app.get("/")
async def route(service: Dep[Service]) -> str:
    same = await resolve(Service)  # same instance as `service`
    ...

Anything called from the handler — including @inject-ed helpers — resolves against the request's cache, so a per-request dependency like a DB session stays a single instance for the whole request.

What is public

Everything the package supports is importable from fastapi_injected itself, and that is the surface a release keeps:

fastapi_injected.types holds the typing vocabulary the signatures are written in — DepReturn, DepShape, DepDecl, AsyncFunc, Coro and friends — and is public too. Anything else, including every module whose name starts with an underscore, is machinery that can change in a patch release.

License

MIT