---

name: type-safety-mastery description: Master type safety patterns - fix type errors at source, never use type: ignore or any, prefer Pydantic models, use type stubs for external libraries

Type Safety Mastery

Fix type errors at their source—never use # type: ignore to bypass warnings. When ty reports an error:

• Define proper types (Pydantic models, Protocol)
• Add type annotations to function signatures when library stubs are incomplete

Never Use Type Ignores

Never use # type: ignore[return-value] or any specific type ignore comments. If a function's return type doesn't match:

• Fix the actual return type
• Use proper type annotations
• Refactor the code to match the declared type

Never Use Any Type

Never use any type. It defeats the purpose of type checking. Instead:

• Use specific types or Union types
• Use TypeVar for generic types
• Use Protocol for duck-typed interfaces
• Use proper type annotations even if it requires more work

Prefer Optional[T]

Prefer Optional[T] over T | None:

• Use Optional[str] instead of str | None
• It is more explicit and readable
• Consistent with the codebase style

Minimize Cast and Runtime Workarounds

Minimize use of cast and runtime workarounds:

• Avoid using cast(Type, val) to silence type errors when possible
• Do not use getattr(obj, 'attr') or setattr(obj, 'attr', val) to bypass missing type definitions
• Prefer fixing the root cause: Update the type stubs in typings/ instead
• When cast is unavoidable (for example, dynamically resolved methods), add a short comment explaining why

Pydantic Over TypedDict

Use Pydantic models instead of TypedDict for data structures:

• Pydantic provides runtime validation
• Better consistency with the rest of the codebase
• Clear error messages when validation fails
• Better IDE support and autocomplete

Always prefer BaseModel over Dict/TypedDict. If you catch yourself annotating a variable as dict[str, X] (or using TypedDict) inside application code, extract a dedicated Pydantic model instead and expose helper methods (get_size(), set_size(), etc.) so callers never need .get() accessors.

Bad example:

Copyfrom typing import TypedDict

class LoRAData(TypedDict):
    id: int
    name: str
    weight: float

Good example:

Copyfrom pydantic import BaseModel, Field

class LoRAData(BaseModel):
    id: int = Field(..., description='Database ID')
    name: str = Field(..., description='Display name')
    weight: float = Field(..., ge=0.0, le=2.0, description='Weight/strength')

Type Aliases

Type aliases for complex types: Create type aliases for frequently used complex types to improve readability.

Bad example:

Copydef list_files(
    self, id: str,
    repo_info: Optional[Union[ModelInfo, DatasetInfo, SpaceInfo]] = None
) -> List[str]:
    pass

Good example:

Copy# Define type alias once at module level
RepoInfo = Union[ModelInfo, DatasetInfo, SpaceInfo]

def list_files(self, id: str, repo_info: Optional[RepoInfo] = None) -> List[str]:
    pass

Type Stubs for External Libraries

Use type stubs (.pyi files) for external library types:

• Create stub files in typings/{package_name}/ instead of runtime wrapper classes
• Update existing stubs when you encounter missing attributes or methods
• Stubs provide type hints without runtime overhead
• Follow PEP 561 conventions (.pyi extension)
• Configure stub path in pyproject.toml under [tool.ty.src]
• Never use runtime assertions (assert isinstance(...)) to force types
• Never use TYPE_CHECKING. If you have to use it, then you did it wrong. Go back and find the root cause, then fix it

Use public interfaces by default (lock, set_state()) and reserve underscores for truly private implementation details.