Coding Python Use Modern Libraries

Modern Python library choices

Default to current, well-maintained libraries over dated stdlib or legacy packages when the better option is clear. Install third-party deps via uv (a tool with uvx, or a script with PEP-723 inline metadata run by uv run). Libraries on this list are pre-approved: just use them and let uv fetch them, no need to ask. These are general public defaults, not absolutes: a project's own conventions win.

Adding an entry

Before adding a library here, vet it: trustworthy (reputable maintainer or community, not a typo-squat), common (widely adopted), and modern (actively maintained, current releases). Each new row must carry: a short description of what it is for (in the Use cell), the older library/libraries it replaces (in the Avoid cell), and - whenever the pick is not self-evident - why it is better (the parenthetical note in the Use cell). A swap that explains itself (pathlib.Path over os.path, pytest over unittest) needs no parenthetical; a contested or trade-off pick (isal vs deflate, one MySQL driver over another) always does. Keep rows one line; do not add anyone's private or in-house packages here.

If the new library overlaps in function with an existing entry, resolve it - never leave two rows silently competing for the same job:

• Both have a place: keep both and make the distinction explicit - state when to use which and why (e.g. one streaming vs one one-shot, sync vs async, simple vs full-featured).
• The new one supersedes the old: only when you are sure, replace the old row with the new one, move the now-obsolete library into the new row's Avoid cell, and say why it was replaced.

Quick reference

Use names the pick with a brief why; Avoid names what it replaces.

Structured data: pydantic at the edges, dataclasses inside

For any structured data in a Python app, parse untrusted/external input into pydantic models at every boundary, and use dataclasses for pure internal layers. Do NOT use attrs, hand-woven classes, or raw dicts for structured data. For the full end-to-end discipline (parse once at the boundary, typed models throughout, Enums for fixed values, minimal conversions), use the bitranox:coding-python-enforce-data-architecture-strict skill.

HTTP example (httpx2)

import httpx2 as httpx          # API-compatible with httpx
with httpx.Client(timeout=30.0, proxy="http://user:pass@host:port") as client:
    r = client.get("https://api.example.com/data")
    r.raise_for_status()
    data = r.json()

Use a Client for many requests (connection reuse), and AsyncClient under asyncio. Both support proxy= (one proxy), mounts= (per-scheme proxy transports), timeout=, and HTTP/2. There is no proxies= kwarg - httpx dropped it in 0.28 and httpx2 never had it, so passing it raises TypeError: Client.__init__() got an unexpected keyword argument 'proxies'.

httpx2 (github.com/pydantic/httpx2) is the legitimate Pydantic-org-stewarded successor to httpx and a drop-in replacement. Some security scanners flag it as a typosquat - that is a FALSE POSITIVE (verified: Pydantic org + Trusted Publisher + Sigstore). If a scanner re-flags it, re-verify the publisher independently and surface it to the user rather than auto-dismissing; do not silently swap it out.

Notes

• Prefer a library over an external command-line tool: httpx2 instead of shelling out to curl, stdlib/orjson instead of jq, re/stdlib instead of grep/sed. Libraries are the same on every OS; external commands may be missing or take different flags on Windows.
• Reach for stdlib when no third-party library is warranted (small glue, no hot path): the point is "best tool", not "most dependencies".
• Logging and CLI framework are deliberately left to each project's own conventions rather than prescribed here.
• subprocess raises a DIFFERENT exception type per OS for the same condition, so never branch on which one you got. A missing cwd raises FileNotFoundError on POSIX but OSError [WinError 267] "The directory name is invalid" on Windows, so a FileNotFoundError -> 127 / OSError -> 126 mapping returns two different exit codes for one condition - and the POSIX run stays green because POSIX happens to raise the type the code expected. Pre-check the path yourself (Path(cwd).exists() / .is_dir()) and decide the result explicitly, then run the missing-path test on Windows before trusting its exit code.
• On 3.14+, except A, B: without parentheses is VALID (PEP 758) and catches both. On 3.13 and earlier the same line is a SyntaxError under the old Python-2 except E, name: grammar, so it reads like an obvious bug. Check requires-python or the interpreter and probe with python -c before "fixing" it; on 3.14+ adding parentheses is a style choice, not a correction.