Complete library typing and enforce mypy strict in CI #722

Plan and status

Agreed with Vincent Biret (@baywet) on 2026-10-05: small PRs, one at a time, no source-breaking changes. Each PR only adds or tightens annotations, and public names stay importable as they are today (__all__ lists every name a module exports now, so from x import * keeps working). When a package reaches zero errors, its [tool.mypy] gets strict = true, so CI keeps it there.

Baseline: mypy --strict reports 297 errors on main (37492f7, 1.14.2) with mypy 2.3.1. 95 of them are missing explicit exports in four __init__.py files. The numbers in brackets below are the errors each item covers once the items before it land.

• Declare the public exports in kiota_abstractions.serialization, kiota_abstractions.store, kiota_http.middleware and kiota_http.middleware.options, and add py.typed to kiota_bundle (297 -> 202, no runtime change) (fix: declare public exports for strict type checkers and ship py.typed in bundle #764)
• Small packages to zero, with strict = true in their [tool.mypy]: kiota_authentication_azure (3), kiota_bundle (2), kiota_serialization_multipart (1), kiota_serialization_form (1), kiota_serialization_text (2), kiota_serialization_json (4) (fix: pass mypy --strict in the small packages and read "false" as False in text #785)
• Default the RequestAdapter and ParsableFactory type parameters to the top of their range (object and the Parsable bound), so generated clients pass mypy --strict without regenerating, and strict = true for tests/validation (141 errors in the validation client, about 26,500 annotations in msgraph-sdk 1.64.0) (fix(abstractions): default the type params of RequestAdapter and ParsableFactory #787)
• kiota_abstractions.store (22)
• kiota_abstractions: request_information and the small modules (13)
• kiota_abstractions.serialization proxy factories (7), then strict = true for kiota_abstractions
• kiota_http.middleware.options (25)
• kiota_http.middleware: retry handler and pipeline (35)
• kiota_http.middleware: the other handlers (32)
• kiota_http.httpx_request_adapter (20) and the rest of kiota_http (4), then strict = true for kiota_http
• A CI check that builds the wheels, installs them and runs mypy --strict on a small consumer file using the public API

PR links are added here as each one opens.

Original request

Hi! I use LLMs to help write code, and passing strict type checks such as mypy --strict is one of my guardrails. A project I'm working on depends on this library, so more complete type information would save me a lot of troubleshooting and help catch mistakes earlier.

Is your feature request related to a problem? Please describe.

The library already has substantial type annotations and runs mypy in CI. However, strict checking still exposes gaps that affect consumers. On the checkout described below, ordinary mypy passes, while mypy --strict reports 298 errors across the eight library packages.

For example, checking this consumer code with mypy --strict example.py reports implicit-export errors for both imports, an untyped call to get_key(), and an Any return:

from kiota_abstractions.serialization import Parsable
from kiota_http.middleware.options import RetryHandlerOption


def retry_option_key() -> str:
    return RetryHandlerOption().get_key()

I also built the microsoft-kiota-bundle wheel locally and confirmed that it lacks py.typed. Mypy treats an installed import of kiota_bundle.default_request_adapter as untyped, despite the annotations in its source.

Describe the solution you'd like

Complete the existing typing support and enforce strict mypy checks for the library packages.

Suggested definition of done:

• All eight library source packages pass mypy --strict in CI with a pinned mypy version across the supported Python versions.
• Consumer checks against installed wheels verify public imports and representative inferred return types; all intended typed packages ship py.typed.
• Public types remain useful, with any necessary suppressions narrowly scoped and explained, rather than broadly replacing types with Any or disabling checks.
• Existing runtime tests pass and public behavior remains compatible.

I suggest keeping unit-test typing and adopting another checker outside the initial scope. Recurring generated-code annotation changes could be tracked separately with the generator work they require.

Describe alternatives you've considered

Local stubs or per-project mypy overrides could work around individual problems, but would leave consumers maintaining those workarounds. Completing the shared library annotations would benefit dependent projects directly.

Additional context

• Checkout: eef36d7e72197063f234b24407edad15ad873bea; local packages report version 1.12.1.
• Environment: Linux, Python 3.13.5, mypy 2.3.1; dependencies installed using uv, with all eight packages installed in editable mode.
• The library-only baseline is 298 errors in 101 checked source files. The generated validation client adds 345 errors in 86 files; those are separate from the library count.
• The largest recurring library categories are missing explicit exports, incomplete function annotations, untyped calls, and missing generic arguments.
• Related earlier typing discussions: Typing questions #355 and Inconsistent typing used in generated code and the libraries. #333, both closed.

Would you be open to using strict mypy compliance as a tracked enhancement for the library packages?