URL netloc overrides reject bracketed IPv6 addresses

[adenzhou1350]

URL.netloc returns bytes for a bracketed IPv6 address, but passing those bytes back to the public constructor or copy_with raises InvalidURL.

Minimal offline reproduction on current master b5addb64f0161ff6bfe94c124ef76f6a1fba5254:

import httpx

url = httpx.URL("http://[::1]:8080/")  # works
print(url.netloc)  # b'[::1]:8080'

url.copy_with(netloc=url.netloc)
# httpx.InvalidURL: Invalid port: ':1]:8080'

# The constructor fails the same way:
httpx.URL(scheme="http", netloc=b"[::1]:8080")

I expected url.copy_with(netloc=url.netloc) to preserve the host and port, as it does for domain names and IPv4. A plain IPv6 URL string already parses correctly; no connection is made in this reproduction.

The netloc override in _urlparse.py splits at the first :, including colons inside IPv6 brackets. I tested a small change that splits bracketed hosts at ]: and leaves the non-bracketed path unchanged. Constructor/copy tests cover IPv6 with and without a port, default-port normalization, IPv4-mapped IPv6, domain/IPv4 controls, preserved userinfo and malformed inputs. With the same regression tests, the baseline has 14 failing IPv6 cases; the change passes all 122 URL tests.

Environment: Linux container, Python 3.10.19, HTTPX 0.28.1 at the commit above; pytest 8.4.1, anyio 4.15.1, httpcore 1.0.9, idna 3.20, Trio 0.31.0. ruff format --diff, mypy and Ruff checks pass. The full offline native suite reports 1,444 passed / 1 failed / 5 network cases deselected after the change. The remaining test_write_timeout[trio] async-generator cleanup warning is also present in the baseline and is not changed here. I have not tested external networking or claimed a complete CI pass.

I searched the existing Issues/PRs and Discussions for netloc / IPv6 copying but may have missed related work. Would this be suitable for a focused PR?

[Arthur031221]

netloc is not round-trippable through copy_with(netloc=...) for a bracketed IPv6 host in HTTPX 0.28.1. Use the parsed host and port fields when rebuilding the URL; HTTPX will add the IPv6 brackets during serialization.

$ python - <<'PY'
import httpx
u = httpx.URL('http://[::1]:8080/')
print(httpx.__version__, repr(u.netloc))
try:
    print(u.copy_with(netloc=u.netloc))
except httpx.InvalidURL as exc:
    print(type(exc).__name__, str(exc))
print(u.copy_with(host=u.host, port=u.port))
PY
0.28.1 b'[::1]:8080'
InvalidURL Invalid port: ':1]:8080'
http://[::1]:8080/

For code that needs a new authority, pass host and port separately to URL(...) or copy_with(...). This preserves the example's IPv6 address and port while the netloc parser issue is resolved. Your bracket-aware split addresses the reproduced failure; a regression test for copy_with(netloc=url.netloc) would cover the public round trip.

[adenzhou1350]

Thanks for confirming. The proposed patch covers both URL(netloc=...) and copy_with(netloc=...), including IPv6/domain controls and malformed inputs; the focused 122-test URL suite passes. I can send the focused PR if maintainers would like it.

[Arthur031221]

Thanks for the test summary. The discussion gives maintainers a concrete reproduction and the scope of the proposed PR.

If the earlier workaround answered your immediate usage question, marking that reply as the answer helps others with the same question find it.