The Proxies page gains a Sort selector next to Source: default order,
lowest latency first or highest latency first. Nodes that are untested,
being tested or whose test failed stay at the end in either direction,
in their original order. compareLatency no longer reports two such
nodes as unequal, which made the sort depend on the engine. The choice
is remembered like the source and the view.
Select failed, in the batch bar, replaces the selection with the listed
nodes whose last test returned TIMEOUT or another error, so they can be
reviewed and removed with the existing Delete. Nothing is deleted
automatically. On a phone with a selection it moves into the more
actions menu.
Refs #2066
In Docker, subscription and RoutingA updates through the core dialled
the default gateway instead of 127.0.0.1, a leftover from when the core
ran outside v2rayA's container. The core now runs in v2rayA's own
network namespace and, unless port sharing is on, listens on 127.0.0.1
only. With host networking the gateway is the LAN router, and on a
bridge network the request only came back to the container when port
sharing was on and the port was published, so updates in the "proxy"
and "port rules" modes failed with connection refused.
Fixes#2139
The test builds of main pass unstable-<date>.r<count>.<commit> to nfpm
as the package version. dpkg refuses it because a Debian version must
start with a digit, and pacman reports the package as corrupted because
the version contains a hyphen; the install check added in #2123 now
fails on the first push to main. Map such versions to the pre-release
0.0.0-unstable.<date>.r<count>.<commit>, which every packager accepts
and which sorts below any release. Release versions are unchanged, and
so are the file names and the version the binaries report.
The pull request build set EXTRA_LDFLAGS as a prefix of the go build
command and expanded ${EXTRA_LDFLAGS} in the same command, which the
shell does before the prefix assignment takes effect, so the flags were
always empty. The release build sets them in the job environment and
was not affected. Use the matrix value directly.
The 2.5.14 apk carried "provides = v2raya", which apk reads as a second
package conflicting with v2raya itself, so it could not be installed on
any Alpine release; nothing in the build noticed, and no package format
was ever installed by CI. Each packaging job now installs its packages
in containers of its own architecture, under QEMU where the runner
cannot run it, and runs v2raya --version: the apk on every architecture
(loong64 from Alpine's own minirootfs, which Docker's alpine image
lacks), the deb wherever Debian publishes an image, the rpm on amd64
and arm64, the Arch package on amd64. A change under install/nfpm or
install/universal now triggers the test build that packages them.
loong64 builds with cgo for mattn/go-sqlite3, and zig targeted
loongarch64-linux-gnu, so since 2.5.14 both binaries were linked
dynamically against glibc (interpreter
/lib64/ld-linux-loongarch-lp64d.so.1). The Alpine loongarch64 package
installed but neither binary could start on musl. Target musl and link
statically instead, so the binaries run on glibc and musl systems
alike. -linkmode=external makes the core, which has no cgo package of
its own, go through zig as well; with the internal linker it cannot
resolve the libc symbols of a static build.
The image downloaded the bare v2raya and v2raya_core binaries and its
own copies of the rule data, so the Alpine package itself was never
installed by anything in the release; the 2.5.14 apk shipped unable to
install. The image now installs installer_alpine_<arch>_<version>.apk
from the release, which makes a broken apk fail the image build, and
takes its geoip.dat and geosite.dat from the package like every other
Linux package. LoyalsoldierSite.dat stays a copy of geosite.dat, as the
image has always shipped it. The script now stops at the first failed
command and on an architecture with no Alpine package.
axios 0.34 tightened its header types, and the request interceptor's
copy of config.headers into a new object no longer type-checks: the
spread carries the per-method header maps, which the index signature
now rejects. axios 0.x keeps headers as a plain object, so the
interceptor now sets Authorization and X-V2raya-Request-Id on it
directly.
Recent pull requests left the "Submitted and answered for by" line
filled in by an assistant or dropped, kept a commit that a later
commit reverted, and described changes their diff no longer had.
CONTRIBUTING.md also told contributors to run "the checks CI runs",
but CI runs neither go vet nor go test, and it named Go 1.26 while
core/ needs 1.27.
- CONTRIBUTING.md: Go 1.27; say which checks CI runs; squash a commit that undoes part of an earlier one; a bug fix
comes with a failing test or a reproduced failure; check the title,
description and messages against the diff; the sign-off line and the
boxes are the contributor's.
- AGENTS.md: ask before changing behaviour that may be intentional; a
check that ran nothing proves nothing; leave the sign-off line and
the human's boxes alone.
- Pull request templates: the sign-off line is required and written by
the contributor.
v2raya.org is parked and redirects to spam and malware pages (#2109).
The GUI still sent users there: the mixed-content banner offered a
"switch to alternate site" button that opened http://v.v2raya.org, and
gui/public/CNAME shipped in every build.
The banner now tells the user to open the GUI over http at the
service's own address instead, the CNAME is gone, and the About page
links the project website, daeuniverse.org.
Closes#2109
runtimeAssetDir called os.MkdirAll on $XDG_RUNTIME_DIR/v2raya with mode
0700. When v2rayA runs as root on a system where /run/user does not
exist yet (OpenRC with elogind before any login), that created
/run/user itself as 0700 root:root, and every other user then failed to
get a runtime directory.
The runtime directory belongs to the login manager, so use it only when
it already exists and otherwise put the asset links in the
configuration directory, as a service user without a session already
does. A regular file there is one the user placed, so the lookup now
uses it instead of replacing it with a link to the system copy.
Closes#2092
A VLESS link with security=tls carries its ECH config list in the ech
query parameter. The service dropped it on import, so the stored node,
the generated core config and the exported share link all lost it, and
the server editor dropped it again when a node was edited and saved.
Parse ech into the stored node, pass it to the core as the TLS
settings' echConfigList, and write it back to share links when the
node uses TLS, in the service and in the editor's codec. The service
reads ech from the raw query so that an unescaped "+" in the base64
list stays a "+" instead of becoming a space. The VLESS
encryption half of the original change landed in 3e374af9.
Closes#2093
With about 1,400 nodes the proxies page froze for a minute on every
visit: it mounted a card, a menu and a tooltip for every row at once,
and each render looked group membership and selection up with a linear
scan per row. The source filter also went back to "all" every time, so
the user waited for the full list before they could narrow it.
The page now mounts 60 rows and adds the next 60 as a sentinel below
the last row scrolls into view; a new search, source or view starts
again from the first page. Select-all, Ctrl+A, the latency test and the
batch actions still work on the whole filtered list. Membership is
indexed once per member list, so inGroup is a map lookup, each row
keeps the same groups array between renders, and selection is checked
against a set; batch add and remove no longer compare every pair.
The selected source is stored like the view. Before the first sync
only "all" and "local" exist, so a saved subscription source is
dropped only after the subscriptions have loaded and it is not among
them.
Closes#2105
TUN mode runs `ip -4 route add default dev tun0 table 2020` and the
matching `ip rule` commands. The image is Alpine, whose busybox `ip`
rejects a numeric table ID ("ip: invalid argument '2020' to 'table'"),
so TUN mode could not start in the official container. iproute2
provides an `ip` that accepts it.
Closes#2077
_install_zig.yml was declared as a reusable workflow but called from a
step, and a step-level uses only accepts an action, so the workflows
failed to validate. A job-level workflow call cannot help here either,
because Zig has to be on PATH in the same job as the build.
Move the install into .github/actions/install-zig, which pins the Zig
version itself and keeps both call sites on a single step.