Agent Skills: Python Typing

Python Typing

Choose the strongest valid lane automatically. Do not ask the user to pick a typing philosophy.

Consult references/typing-policy.md for the full policy document.

Required Policy

• Forbid Any, broad object, and unchecked cast() in normal internal code
• Allow them only at explicit boundaries where unknown-shape data enters
• Isolate boundary code in dedicated validator, parser, adapter, or boundary modules
• Validate immediately and return strongly typed internal objects
• Do not let raw payloads cross into the typed core
• Allow narrow lint exceptions for Any only in approved boundary modules

Lane Selection

1. Python 3.10 Constrained or stdlib-only

• Use only compatibility-safe stdlib typing features
• Prefer dataclasses, TypedDict, Protocol, Literal, TypeGuard, NewType
• Validate with explicit runtime checks in dedicated boundary wrappers
• No third-party type assumptions

2. Python 3.11+ stdlib-only

• Use modern stdlib typing features supported by the interpreter
• Use Self, assert_type, and reveal_type where useful during refactoring
• TypedDict with NotRequired

3. Python 3.11+ with Pydantic

• Use Pydantic models for ingress contracts
• Prefer strict mode where coercion would hide upstream errors
• Use TypeAdapter for annotated types that do not need full models
• See references/pydantic-boundaries.md

4. Python 3.11+ with Hypothesis

• Property-test boundaries, validators, parsers, and invariants
• Prefer from_type() where practical
• See references/hypothesis-boundaries.md

5. Python 3.12+

• Use type statement for explicit type aliases: type JSONValue = str | int | ...
• Use PEP 695 generic parameter syntax for new generic helpers

6. Python 3.13+

• Use TypeIs for clearer custom narrowing helpers (replaces TypeGuard where bidirectional narrowing needed)
• Use ReadOnly in TypedDict fields that must not mutate after validation

7. Python 3.14+

• Keep annotation-reading infrastructure compatible with deferred evaluation (PEP 649)
• Use annotationlib.get_annotations() in infrastructure that inspects annotations at runtime

Boundary Implementation Standard

Use dedicated wrappers named like:

• parse_*
• validate_*
• decode_*
• coerce_*
• *_from_raw

Boundary modules should return typed objects only.

Example: stdlib-only boundary

from typing import TypedDict, NotRequired
from dataclasses import dataclass

class _RawIncoming(TypedDict):
    user_id: int
    email: str
    metadata: NotRequired[dict[str, str]]

@dataclass(frozen=True, slots=True)
class IncomingPayload:
    user_id: int
    email: str
    metadata: dict[str, str]

def parse_incoming(data: _RawIncoming) -> IncomingPayload:
    return IncomingPayload(
        user_id=data["user_id"],
        email=data["email"],
        metadata=data.get("metadata", {}),
    )

Example: Pydantic boundary

from pydantic import BaseModel, TypeAdapter

class IncomingPayload(BaseModel):
    user_id: int
    email: str
    metadata: dict[str, str] = {}
    model_config = {"strict": True}

def parse_incoming(data: dict[str, object]) -> IncomingPayload:
    return IncomingPayload.model_validate(data)

References

• references/typing-policy.md — full boundary validation policy
• references/pydantic-boundaries.md — Pydantic model and TypeAdapter patterns
• references/hypothesis-boundaries.md — property-based testing for validators