console v2026.10.12.2 and webmail inbuxa-v2026.10.12-g6e1a1cb are the
first releases to carry their tarballs; their checksums are pinned, from
the releases' SHA256SUMS and checked against downloads (the webmail's two
architectures are identical inside, with no native code). The console's
and the webmail's images follow the same constants, so a container and a
host install of either are the same release.
Server 2026.10.12.1 (upstream 0.16.26, with inbuxa:FrontEnds) replaces
2026.10.12, checksums pinned the same way; still glibc 2.39. join's
refusal of an older server, and the README, name it.
The interface's components screen now also asks the whole plan, so a
choice the machine could make but the other choices rule out -- the
console as files on loopback, with no web server to serve them -- is
refused there, not first at the plan.
All e2e cases on the real downloads, no overrides: webmail-host 18/18,
all-host 17/17, join 24/24 (against the released server), install-host
43/43, install-local 26/26, install-public 20/20, proxy-snippets 13/13,
interface 22/22, status-export 16/16, topology 28/28.
--console host and --webmail host install them without containers, from
their releases' tarballs (inbuxa-admin #106, inbuxa-webmail #68), each
checked against a checksum pinned in internal/release.
- The console is files: unpacked into <dir>/console, root-owned and
readable, and pointed at the server the way its image's entrypoint does
(the api-base-url meta tag). The managed Caddy serves the directory,
mounted read-only, or the operator's web server does from the
configuration --proxy snippets writes; both with the image's rules (app
routes fall back to index.html, hashed assets cached for good, index.html
never). With --local or --proxy none nothing would serve it, and the
plan says so.
- The webmail is the /app its image runs, in <dir>/webmail, owned by root,
under inbuxa-webmail.service as its own user on the installer's own Node
(now 26, as its image) in /opt/inbuxa/node, on 127.0.0.1:8080, with the
container's environment in <dir>/webmail.env. It can't write its own
files, which is what its IMMUTABLE setting checks. The unit is
ProtectSystem=strict. A server in a container is reached at its loopback
port and sees it at the project network's gateway, which is what gets
exempted from the auto-ban.
- Both unpack beside the old copy and swap only when complete; a re-run
skips an unchanged release and restarts the webmail only when what it
reads at start changed. A webmail moved back into a container gives its
port back first. status and export report them as host.
With all three on the host and --proxy snippets, no container runtime is
installed. The release pins are empty until those releases exist, and the
host shapes say so; INBUXA_CONSOLE_TARBALL and INBUXA_WEBMAIL_TARBALL (and
CONSOLE_TARBALL=, WEBMAIL_TARBALL= in e2e/vm/run.sh) install a local
tarball meanwhile.
Also: the "using docker" line no longer prints when nothing runs in a
container; Verify takes the plan's options and checks a host console by
its files.
Tests: plan (console host needs a web server; host front ends wait for
their releases; the webmail wants the installer's own Node), compose
(console files, the webmail unit, Caddy on the host's network, the
gateway exemption, nginx serving the directory), apply (api-base-url
replaced or inserted, the webmail's environment, unpack's swap and
marker). e2e: webmail-host.sh 18/18 (signing in included), all-host.sh
17/17 (no runtime, behind nginx), both with tarballs taken out of the
current images as the release jobs will; install-local, install-host,
install-public, proxy-snippets, interface, status-export and topology
pass unchanged.