typekeeper

Lightweight, dependency-free runtime argument validation for Python via a single decorator. See Changelog.md for version history.

Overview

typekeeper instruments your functions with comprehensive runtime checks through one decorator, @validate_args. With no code changes beyond the decorator, you get:

• Type checking against rich typing annotations.
• Warnings for mismatched or mutable default values.
• Inclusive numeric-range and length-range constraints via a tiny spec string, including nested-path targeting with : and wildcards.
• Recursive _validate() traversal of nested data structures.
• Precise source locations (file + line) for every warning, both at decoration time and call time.
• Global, per-decorator, and contextual on/off switches.

Features

• Type validation — recursively checks the common typing forms: built-in generics (list[int], dict[str, float]), Optional, Union / PEP 604 unions, Literal, Type[T] / type[T], Annotated, Final, NewType, TypeVar (constraints and bounds), Callable (callability and positional arity), TypedDict (structural dict check), Protocol (structural attribute check for both @runtime_checkable and plain protocols), Self, PEP 695 type aliases (TypeAliasType), and abstract base classes such as Sequence, Mapping, and Iterable (mutable variants correctly reject immutable values). Unresolved string forward references are treated as unchecked; ForwardRef objects are resolved against the typing namespace when possible.
• Default-value checks — warns when a default does not match its annotation, and detects mutable defaults (suppressible per decorator via ignore_defaults=True).
• Range & length specs — pass constraints; each spec is name=token1,token2,... where a token is either a single number N or an inclusive range min-max. Whitespace around the inner - is tolerated. Use : to traverse nested mappings and sequences, with integer segments selecting sequence elements (negative indices follow normal Python rules). Wildcards (* or a trailing :) match every element under that point; later specs override earlier ones for the same target. Missing mapping keys yield no matches (no spurious warnings).
• Recursive _validate() — every argument is walked. Any object with a callable _validate() is invoked; falsy returns or raised exceptions are reported with the path inside the original structure. Mappings (including UserDict / MappingProxyType) are recursed via collections.abc.Mapping. Iterators and generators are never consumed by validation.
• Context-aware warnings — messages embed both the decoration location (file:line) and the call site.
• Global control — set_arg_checks(False) disables decoration entirely; set_stop_on_error(True) turns every warning into a ValueError; suspended_arg_checks() is a context manager that switches checks off inside a with block (overhead drops to under half a microsecond per call). Suspension is thread-/task-local, so it never disables checks running concurrently in other threads.

Installation

pip install typekeeper

Development

pip install -e ".[dev]"
python -m pytest tests/ --cov=typekeeper

The test suite covers 100% of typekeeper/typekeeper.py and exercises a broad slice of typing features end-to-end.

Configuration

API Reference

validate_args

def validate_args(
    _func=None,
    constraints: str | None = None,
    *,
    ignore_defaults: bool = False,
    stop_on_error: bool | None = None,
) -> Callable: ...

• constraints — numeric/length spec string covering both numeric ranges and sequence-length ranges (see syntax below).
• ignore_defaults — skip the mutable-default warning if True.
• stop_on_error — per-decorator override of the global set_stop_on_error flag; warnings from this decorator raise ValueError regardless of the global setting.

Constraint syntax — a semicolon-separated list:

name=token1,token2,...

Each token is N (exact match) or min-max (inclusive range). Nested targets use : after the name: users:tables:headers=100, rows:0:score=0-100, data:*=1-5. Whitespace inside ranges is tolerated (1 - 3 parses identically to 1-3).

Inside path segments, the following backslash escapes are honoured:

A bare empty segment (e.g. trailing :) or * still means every entry at this level; specs whose first segment is empty (":x=1-3") emit a parse-time warning instead of being silently dropped.

Examples

Full runnable scripts live in examples/. Each script prints its own validation overhead via examples/_bench.py.

# Snapshot of basic.py
from typekeeper import validate_args

@validate_args()
def greet(name: str, times: int = 1) -> str:
    return " ".join([name] * times)

greet("Alice")
greet(42)
#  UserWarning: Arg 'name' to '<function greet ...>' mismatches <class 'str'>;
#  expected <class 'str'>, got <class 'int'>

Performance

Run tools/profile.py to reproduce these numbers. They are representative of a single CPython 3.14 run on a desktop x86-64 machine (50,000 iterations per scenario, after a warm-up burst):

Run it yourself:

python tools/profile.py --iterations 50000          # scenario benchmark
python tools/profile.py --cprofile                  # cProfile breakdown
python tools/profile.py --iterations 50000 --json results.json

Key takeaways:

• The decorator itself (suspended mode) costs well under 1 µs per call (typically 0.2 µs for fixed-arity signatures).
• Scalar arguments add roughly 10 µs per call for type checking.
• Per-element checking dominates the cost on collection-shaped signatures; expect ~5–10 µs / leaf in the worst case.
• Path-spec wildcards walk every leaf — prefer indexed paths (rows:0:v=...) when you only care about a few cells.
• The _recursive_validate() walk runs on every argument; if your hot path passes large structures and you don't need it, hoist the validated call out of the loop or wrap it in with suspended_arg_checks(): ....

Contributing

Contributions are welcome:

1. Fork the repository.
2. Create a feature branch.
3. Run python -m pytest tests/ --cov=typekeeper (target stays at 100%).
4. Submit a pull request.

License

MIT License.

Author

Parth Mittal — parth@privatepanda.co