Unihttp

unihttp skill

unihttp is a library for building declarative, type-safe API clients. You define API endpoints as typed BaseMethod subclasses and bind them to a client backed by httpx, aiohttp, requests, niquests, or zapros.

Here is a list of best practices for different parts of a unihttp client. For deeper reference material, see:

• references/markers.md — every marker and how it maps to the request
• references/serializers.md — choosing and customizing adaptix / pydantic / msgspec
• references/backends.md — choosing an HTTP backend and sync vs async

Imports

Import from submodules, never from the top-level unihttp package

The top-level unihttp/__init__.py is intentionally empty — it re-exports nothing. Importing names from unihttp directly raises ImportError at runtime, even though some README snippets show it.

Wrong:

# Raises: ImportError: cannot import name 'BaseMethod' from 'unihttp'
from unihttp import BaseMethod, Path, Query, Body, bind_method

Correct:

from unihttp.method import BaseMethod
from unihttp.markers import Body, File, Form, Header, Path, Query
from unihttp.bind_method import bind_method
from unihttp.http import UploadFile  # for multipart uploads

The only packages that DO re-export a public API are the serializer subpackages and unihttp.http / unihttp.middlewares / unihttp.exceptions:

from unihttp.serializers.adaptix import DEFAULT_RETORT, AdaptixDumper, AdaptixLoader
from unihttp.clients.aiohttp import AiohttpAsyncClient
from unihttp.exceptions import ClientError, ServerError

Defining methods

Define endpoints as @dataclass subclasses of BaseMethod[ReturnType]

The return type goes in the generic parameter — that is what the client returns and what the response is deserialized into. Declare __url__ and __method__ as class attributes, and annotate fields with markers.

Wrong:

from unihttp.method import BaseMethod
from unihttp.markers import Path


class GetUser(BaseMethod):  # no return type — client returns `Any`
    url = "/users/{id}"     # wrong attribute name
    method = "GET"          # wrong attribute name
    id: int                 # no marker — adaptix cannot place it in the request

Correct:

from dataclasses import dataclass

from unihttp.method import BaseMethod
from unihttp.markers import Path


@dataclass
class GetUser(BaseMethod[User]):
    __url__ = "/users/{id}"
    __method__ = "GET"

    id: Path[int]

__returning__ is filled automatically from the generic argument, so client.get_user(id=1) is statically typed as User.

Every {placeholder} in __url__ must have a matching Path field

__url__.format(**path_data) fails with KeyError at request time if a placeholder has no corresponding Path field.

Wrong:

@dataclass
class GetUserPost(BaseMethod[Post]):
    __url__ = "/users/{user_id}/posts/{post_id}"
    __method__ = "GET"

    user_id: Path[int]
    # post_id is missing -> KeyError: 'post_id' when the request is built

Correct:

@dataclass
class GetUserPost(BaseMethod[Post]):
    __url__ = "/users/{user_id}/posts/{post_id}"
    __method__ = "GET"

    user_id: Path[int]
    post_id: Path[int]

Binding methods to a client

Prefer bind_method for the common case

bind_method is a descriptor that exposes the method on the client with the exact typed signature of its constructor — sync clients return the value, async clients return an awaitable. Reach for it whenever an endpoint needs no extra per-call logic.

Wrong (hand-written passthrough that duplicates the signature and loses typing):

class UserClient(AiohttpAsyncClient):
    async def get_user(self, id: int) -> User:
        return await self.call_method(GetUser(id=id))

Correct:

from unihttp.bind_method import bind_method


class UserClient(AiohttpAsyncClient):
    get_user = bind_method(GetUser)

Use call_method only when an endpoint needs real custom logic

When you must preprocess arguments, combine calls, or adapt names, write an explicit method and call call_method. Do not use it just to rename parameters that bind_method already handles.

Correct:

class UserClient(AiohttpAsyncClient):
    create_user = bind_method(CreateUser)  # plain endpoint -> bind_method

    async def register(self, email: str) -> User:
        # genuine extra logic -> call_method
        normalized = email.strip().lower()
        return await self.call_method(CreateUser(email=normalized))

Markers

Pick the marker that matches the wire location; never put a body field in the query

See references/markers.md for the full table. The common mistakes:

• JSON request body -> Body, not Query.
• application/x-www-form-urlencoded fields -> Form.
• Multipart file uploads -> File (wrap content in UploadFile to set filename and content type).
• Body cannot be combined with Form/File on the same method — the clients raise ValueError. Use Form for the scalar fields of a multipart request.

Wrong:

@dataclass
class CreateUser(BaseMethod[User]):
    __url__ = "/users"
    __method__ = "POST"

    name: Query[str]   # this belongs in the JSON body, not the query string
    email: Query[str]

Correct:

from unihttp.markers import Body


@dataclass
class CreateUser(BaseMethod[User]):
    __url__ = "/users"
    __method__ = "POST"

    name: Body[str]
    email: Body[str]

Serializers

Prefer adaptix + dataclasses by default

For plain dataclasses / TypedDicts, adaptix is the recommended default: fast, no runtime dependency inside your model classes, and field-level customization. Choose msgspec when raw speed matters and pydantic when you need its rich validation. See references/serializers.md.

Pass the serializer explicitly with AdaptixDumper / AdaptixLoader:

from unihttp.serializers.adaptix import DEFAULT_RETORT, AdaptixDumper, AdaptixLoader

client = UserClient(
    base_url="https://api.example.com",
    request_dumper=AdaptixDumper(DEFAULT_RETORT),
    response_loader=AdaptixLoader(DEFAULT_RETORT),
)

Map wire names with a Retort instead of renaming your Python fields

When an API uses camelCase (or any non-Pythonic naming), keep idiomatic snake_case fields and let adaptix translate them. Do not pollute your dataclasses with userName-style attributes.

Wrong:

@dataclass
class User:
    userName: str  # noqa: N815 — leaks the wire format into Python field names
    emailAddress: str  # noqa: N815

Correct:

from dataclasses import dataclass

from adaptix import name_mapping

from unihttp.serializers.adaptix import DEFAULT_RETORT, AdaptixDumper, AdaptixLoader


@dataclass
class User:
    user_name: str
    email_address: str


# Define the model first, then extend DEFAULT_RETORT (keyword-only `recipe`) so
# its built-in providers are kept and `name_mapping` can reference the class.
retort = DEFAULT_RETORT.extend(
    recipe=[name_mapping(User, map={"user_name": "userName", "email_address": "emailAddress"})],
)


client = UserClient(
    base_url="https://api.example.com",
    request_dumper=AdaptixDumper(retort),
    response_loader=AdaptixLoader(retort),
)

name_mapping also applies to BaseMethod subclasses — a method's Body and Query fields are renamed the same way, so map the method classes too, not only the response models. When many classes share the same snake↔camel fields, use one global name_mapping(map={...}) with no class predicate; it renames those fields everywhere (models, request bodies, query filters) in a single recipe entry.

Backends, sync and async

Match the client class to sync/async and to the installed backend

Async backends need an async client class and await; sync backends a sync one. Install the matching extra (unihttp[aiohttp], unihttp[requests], …). See references/backends.md for the full matrix.

Correct (async):

from unihttp.clients.aiohttp import AiohttpAsyncClient

class UserClient(AiohttpAsyncClient):
    get_user = bind_method(GetUser)

# bind_method returns an awaitable on async clients
user = await client.get_user(id=1)

Correct (sync):

from unihttp.clients.requests import RequestsSyncClient

class UserClient(RequestsSyncClient):
    get_user = bind_method(GetUser)

user = client.get_user(id=1)

Always close async clients — prefer async with

Async clients own an HTTP session that must be closed, or you leak connections. The base async client implements __aenter__/__aexit__, so use a context manager. Construct the client inside a running event loop.

Wrong:

client = UserClient(base_url=..., request_dumper=..., response_loader=...)
user = await client.get_user(id=1)
# session never closed -> "Unclosed client session" warnings, leaked sockets

Correct:

async with UserClient(
    base_url="https://api.example.com",
    request_dumper=AdaptixDumper(DEFAULT_RETORT),
    response_loader=AdaptixLoader(DEFAULT_RETORT),
) as client:
    user = await client.get_user(id=1)

Error handling

unihttp handles errors in layers — pick the narrowest one that fits.

Map status codes to exceptions with the error-mapper middleware, not ad-hoc ifs

For "turn HTTP status N into exception X" rules that apply across endpoints, use the error-mapper middleware instead of repeating status checks. Match the sync/async variant to your client.

Wrong:

class UserClient(AiohttpAsyncClient):
    async def get_user(self, id: int) -> User:
        user = await self.call_method(GetUser(id=id))
        # status handling buried in every method, duplicated everywhere
        return user

Correct:

from http import HTTPStatus

from unihttp.middlewares import AsyncErrorMapperMiddleware

client = UserClient(
    base_url="https://api.example.com",
    request_dumper=AdaptixDumper(DEFAULT_RETORT),
    response_loader=AdaptixLoader(DEFAULT_RETORT),
    middleware=[
        AsyncErrorMapperMiddleware({
            HTTPStatus.NOT_FOUND: UserNotFound,
            HTTPStatus.UNAUTHORIZED: SessionExpired,
        }),
    ],
)

Keys may be an int/HTTPStatus, a range, or a tuple of codes. An exception class is instantiated as ExcClass(message, response) (so it must accept those two positional args — a plain Exception subclass does). Pass lambda r: ... when you need custom construction, e.g. 429: lambda r: RateLimited(r.headers["Retry-After"]).

Override on_error (method) or handle_error (client) for targeted handling

Use a method's on_error to raise an endpoint-specific exception, and the client's handle_error for cross-cutting concerns like auth. on_error is a raise-or-pass hook: its return value is ignored, so return None does not make the call return None — it just lets the pipeline fall through and deserialize the error body. To turn a 404 into a None result, override make_response or wrap the call in a call_method method that catches the error.

@dataclass
class GetUser(BaseMethod[User]):
    __url__ = "/users/{id}"
    __method__ = "GET"
    id: Path[int]

    def on_error(self, response: HTTPResponse) -> None:
        if response.status_code == 404:
            raise UserNotFound(self.id)  # endpoint-specific error
        return super().on_error(response)

Validate "200 OK but error in body" with validate_response

Some APIs return 200 with {"ok": false}. Override validate_response (on the method or client) instead of leaking that check into business code.

def validate_response(self, response: HTTPResponse) -> None:
    if isinstance(response.data, dict) and response.data.get("ok") is False:
        raise ApiError(response.data.get("error", "unknown error"))

Read response headers/status by overriding make_response

bind_method returns only the deserialized body — headers, status code, and cookies are invisible to callers by default. When you need them (a Link pagination header, a rate-limit header, a Location), override the method's make_response; it receives the full HTTPResponse (.headers, .status_code, .cookies, .data). The return type then becomes your wrapper, not the bare body.

from unihttp.http.response import HTTPResponse
from unihttp.serialize import ResponseLoader


@dataclass
class ListIssues(BaseMethod[IssuePage]):  # IssuePage wraps items + next_link
    __url__ = "/repos/{owner}/{repo}/issues"
    __method__ = "GET"
    owner: Path[str]
    repo: Path[str]

    def make_response(
        self,
        response: HTTPResponse,
        response_loader: ResponseLoader,
    ) -> IssuePage:
        # The base method's parameter is `response_loader` and the client calls it
        # by keyword — keep that exact name or the override raises TypeError.
        items = response_loader.load(response.data, list[Issue])
        return IssuePage(items=items, next_link=parse_link(response.headers.get("Link")))

Middleware

Match middleware sync/async type to the client, and call next_handler

Sync clients take Middleware (sync handle); async clients take AsyncMiddleware (async handle). Always pass the request to next_handler and return its response, unless you deliberately short-circuit.

Correct (async client -> async middleware):

from unihttp.middlewares import AsyncHandler, AsyncMiddleware
from unihttp.http.request import HTTPRequest
from unihttp.http.response import HTTPResponse


class AuthMiddleware(AsyncMiddleware):
    def __init__(self, token: str) -> None:
        self._token = token

    async def handle(
        self,
        request: HTTPRequest,
        next_handler: AsyncHandler,
    ) -> HTTPResponse:
        request.header["Authorization"] = f"Bearer {self._token}"
        return await next_handler(request)

Type next_handler (AsyncHandler for async, Handler for sync — both from unihttp.middlewares); leaving it unannotated only passes mypy because the generated mypy.ini mutes no-untyped-def. Note the field name asymmetry: a request carries request.header (singular), a response uses response.headers (plural). For bearer/API-key auth, take the token as a client __init__ parameter and attach this middleware there. For transient failures, prefer the built-in RetryMiddleware / AsyncRetryMiddleware over a hand-rolled retry loop.

Custom JSON

Swap in a faster JSON codec via json_dumps / json_loads

import orjson

client = UserClient(
    base_url="https://api.example.com",
    request_dumper=AdaptixDumper(DEFAULT_RETORT),
    response_loader=AdaptixLoader(DEFAULT_RETORT),
    json_dumps=lambda obj: orjson.dumps(obj).decode(),
    json_loads=orjson.loads,
)