Language Python

Python

Type hints

• Annotate all function signatures; Any is a smell unless at a genuine boundary.
• Use dataclass for data containers with behavior; TypedDict for dict-shaped data at boundaries.
• Prefer Protocol over ABC when duck typing suffices — no inheritance required.
• from __future__ import annotations for forward refs in 3.9 and earlier.

from dataclasses import dataclass, field

@dataclass
class Order:
    id: str
    items: list[str] = field(default_factory=list)
    total: float = 0.0

Errors and exceptions

• Use exceptions for exceptional paths, not flow control.
• Raise specific subclasses; catch the narrowest class you can handle.
• except Exception: is almost always wrong — at minimum log and re-raise.
• contextlib.suppress(SomeError) for intentional ignore; bare except: never.

try:
    result = load(path)
except FileNotFoundError:
    raise MissingConfigError(path) from None

Context managers

• with for any resource with a cleanup obligation: files, locks, DB connections.
• @contextlib.contextmanager for ad-hoc managers without a full class.
• Never hold a resource longer than the with block.

@contextlib.contextmanager
def managed_resource():
    r = acquire()
    try:
        yield r
    finally:
        release(r)

Generators and iterators

• Prefer generators over materializing full lists when you only iterate once.
• yield from to delegate to sub-generators.
• Reach for itertools before writing loops: chain, islice, groupby, product.

def read_chunks(path: Path, size: int = 4096):
    with open(path, "rb") as f:
        while chunk := f.read(size):  # walrus operator, 3.8+
            yield chunk

Concurrency

• GIL caveat: threads don't parallelize CPU-bound work — use ProcessPoolExecutor or multiprocessing.
• asyncio for IO-bound concurrency; asyncio.TaskGroup (3.11+) for structured fan-out.
• ThreadPoolExecutor for legacy sync IO or blocking C extensions.
• One event loop per process; never nest or mix loops.

async def fetch_all(urls: list[str]) -> list[str]:
    async with asyncio.TaskGroup() as tg:
        tasks = [tg.create_task(fetch(u)) for u in urls]
    return [t.result() for t in tasks]

Pattern matching (3.10+)

Use match for structural dispatch on data shapes; avoid it as a glorified if/elif chain.

match command:
    case {"action": "move", "direction": dir}:
        move(dir)
    case {"action": "quit"}:
        quit()
    case _:
        raise ValueError(f"unknown command: {command}")

Dependencies and packaging

• pyproject.toml is the standard — no setup.py in new projects.
• uv for fast installs; poetry for lockfile publishing workflows.
• Pin transitive deps via lockfile (uv.lock, poetry.lock) in applications; use version ranges in libraries.
• Always isolate with a virtualenv — never install into the system Python.

Tooling

• Imports: ruff check --select I --fix
• Format: ruff format / black .
• Lint: ruff check (+ mypy for types)
• Test: pytest (see Testing below)

Testing

• pytest over unittest — fixtures, parametrize, and plugins make it richer.
• @pytest.mark.parametrize instead of loops inside tests.
• unittest.mock.patch for external boundaries only; don't mock internals.
• pytest-asyncio for async tests; respx or httpx mock transport for HTTP clients.

@pytest.mark.parametrize("a, b, expected", [(1, 2, 3), (0, 0, 0)])
def test_add(a, b, expected):
    assert add(a, b) == expected

Performance

• Profile before optimizing: cProfile for CPU hotspots, tracemalloc for memory.
• py-spy samples live processes without code changes.
• C extensions (cffi, Cython) only after profiling confirms a Python bottleneck.
• Cache attribute lookups in tight loops: fn = obj.method outside the loop.

Avoid

• Mutable default arguments (def f(x=[]) — use None, assign inside).
• from module import * — pollutes namespace, breaks static analysis.
• global / nonlocal except in narrow closures.
• Broad try/except blocks that swallow errors silently.
• subprocess.run(shell=True) with user-controlled input — use the list form.
• Reimplementing what itertools, functools, or collections already provide.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: lugassawan
• Source: lugassawan/swe-workbench
• License: MIT

Install and usage instructions live in the source repository linked above.