Commit Graph
9 Commits
Author SHA1 Message Date
jcoffey-dev c78847440d Merge pull request 'Local AI README: no internal host name, and the port the webmail shares' (#2) from packaging/local-ai-readme into main
Reviewed-on: #2
2026-09-24 14:57:27 +00:00
jcoffey-dev 3cedee4202 Local AI README: no internal host name, and the port the webmail shares
Drops the internal machine name from the measurement line: this repo is
public. Says that 127.0.0.1:8080 is the mail namespace's own loopback, and
that a model run outside a namespace on a host the installer set up needs
another port, since the container install's webmail already binds the
host's 127.0.0.1:8080.
2026-09-24 07:55:57 -07:00
jcoffey-dev d4351beb78 Merge pull request 'Local AI: the model service production runs, and how to build its model' (#1) from packaging/local-ai into main
Reviewed-on: #1
2026-09-24 14:55:32 +00:00
jcoffey-dev 393b9d897f Local AI: the model service production runs, and how to build its model
packaging/local-ai/ holds what runs the AI spam classifier's model on
production's mail host, for the installer's bare-metal mode to use later
(nothing in the installer uses it yet):

- inbuxa-llm.service: llama.cpp's server with Qwen3 4B Instruct 2507
  (Q4_K_M), in the mail network namespace on 127.0.0.1:8080 only, bounded
  to 4 cores and 8 GB. Identical to host1's unit apart from the license
  header.
- build-model.sh: builds the model from Qwen's official weights, checked
  against Hugging Face's checksums, with llama.cpp b11160's converter and
  quantizer, and compares the result with production's SHA-256. Qwen
  publishes no GGUF of this model, so the file that runs is one we made.
  Two builds from the same weights came out byte-identical.
- README.md: what runs and where it came from, a by-hand install, and how
  to undo it.
2026-09-24 07:54:09 -07:00
jcoffey-dev e55ff6b2df Offer only what this machine can deliver
The installer knew one distribution: Debian, with docker, on a new enough
release. Everything else it would have offered and then failed at.

Three facts now decide what the matrix offers, and each is read from the
machine rather than assumed:

- The container runtime. Docker where the machine has one, podman on the Red
  Hat family, which ships no docker at all. Both go through the same compose
  plugin: compose speaks the Docker API and podman serves it, so there is one
  compose file and one deployment path, not two. Telling a Fedora operator to
  add Docker's own repository to a machine that already has a container
  runtime would have been the wrong trade.

- glibc. The server binary is downloaded, not built here, and it is linked
  against 2.39. Rocky 9 (2.34), Debian 12 and Ubuntu 22.04 cannot run it, so
  the host shape is refused there with the version found and the container
  shape named as the answer -- rather than installing a file that cannot
  start.

- The operating system itself. This compiles for macOS and Windows because Go
  compiles anything, and on either it would read no os-release, find no
  systemd, and describe a machine that does not exist. It now says what it is
  and exits.

Two bugs the other distributions found, both of which Debian could not have:

- The survey reported the first thing in the way and stopped, so on Fedora it
  asked to start podman.socket, and then -- having done it -- asked for the
  compose plugin. Needs are named now, not described, and reported together.

- apply used the survey taken before dependencies were installed, so on a
  machine that had no runtime at all it installed podman and then reached for
  docker. It re-surveys after resolving, and stops if containers still are
  not usable.

The lab takes DISTRO now: debian13, debian12, ubuntu2404, fedora, rocky9,
arch, each with its own disk and ssh port so several can be up at once. The
cases no longer say "docker" either. install-local passes on Debian 13,
Fedora 43 and Rocky 9 -- 20 checks each, ending with a sign-in to the webmail
the installer put there.
2026-09-22 22:22:45 -07:00
jcoffey-dev 8725d8c11b Obtain certificates, and wait for the one that matters
The public shape now works end to end: real ports, Caddy in front, and both
programs that need certificates getting them from the same CA -- Caddy for
the front ends over TLS-ALPN-01, the mail server for its own names over
HTTP-01, which Caddy forwards on port 80.

Proved in the lab against Pebble, with a DNS stub answering every name with
the machine's own address, so no public name or public CA is involved:
twenty checks, ending with IMAPS and submissions presenting a certificate
for the mail host that verifies against the CA, and the webmail sending
sign-in to the server as the first-party client the server registered.

Two things the test found, both of which would have shipped:

- The proxy fronted four of the server's five names. The server puts
  ua-auto-config in its own certificate too, so its challenge was never
  forwarded, one name failed, and the whole order failed with it -- leaving
  the mail ports on a self-signed certificate while everything else looked
  healthy. The list now matches what the server asks for.

- Nothing waited for the certificate. An order that fails is not retried on
  its own and a restart does not start a new one, so the install declared
  itself finished over a self-signed certificate. It now waits, asks again
  every 45 seconds, and reports the issuer -- or says plainly that the
  server will keep trying once the domain resolves here, which is the
  ordinary case on a first install.

--acme-directory and --acme-ca-root are what let a private CA be used: the
root is added to the server image's own bundle and given to Caddy, because
neither sees the other's trust store.
2026-09-22 19:11:52 -07:00
jcoffey-dev 7141e565ea Install the suite, for real, in containers
The plan now happens. "inbuxa install --local --domain example.test
--install-deps --yes" on a machine with nothing on it ends with a mail
server, a console and a webmail running, an administrator and a first
mailbox created, and the records the domain needs written out.

The sequence is the one ihasmail-oneshot worked out against a running
server, which is why its JMAP client and its Docker handling came across
nearly whole: bring the server up in bootstrap mode with a credential that
lives in an override file for that step only, complete bootstrap, bring the
rest up without it -- so no recovery credential outlives the setup -- exempt
the front ends from the auto-ban, restart for the settings that need it,
create the first account, and write down the password nothing else holds.

New here: three services rather than two. The console is static files that
learn their server's address at start, and the webmail is given the
first-party OAuth client secret that the server is given too.

Twenty checks in the lab, from a bare Debian 13. The two worth having are
the ones that catch an install that looks fine and is not: nothing in the
running server carries a recovery admin any more, and the account the
installer created can sign in to the webmail it installed.

Two bugs the lab caught, both of which would have shipped:

- the private addresses were worked out on a copy of the stack, so the
  server was told the webmail speaks from "", and refused it.
- the console image rewrites index.html when it starts, so a read-only root
  filesystem left it restarting forever. The webmail keeps read_only; the
  console cannot have it until that rewrite moves.
2026-09-22 18:26:31 -07:00
jcoffey-dev cc77f62c01 Offer to install what is missing, rather than refusing over it
Refusing because Docker is absent is not help. It is a chore handed back to
the operator, who will then install it however the first search result says
to -- which is how machines end up with a third-party apt repository and a
signing key nobody chose.

So the installer offers. It says what it would do, in the same words the
plan uses, and does it only when told: --install-deps during a run, or
"inbuxa deps --install" on its own for someone preparing a machine before
they have a domain to give it.

What it installs, and from where, is the part worth arguing about:

- the Docker daemon: the distribution's own package. One package from an
  archive the machine already trusts beats a new source.
- the Compose plugin: Debian's docker.io ships no compose v2 at all, so this
  is Docker's official static build, pinned by version with its checksum in
  the source beside it.
- Node: the official tarball into /opt/inbuxa/node, deliberately off PATH so
  it runs the webmail's unit and nothing else.

Every download is checked before it is put in place; a checksum that does
not match stops the step rather than warning and carrying on.

Tested from a bare Debian 13 in the lab: 25 checks, ending with docker and
compose answering, node 22 where the unit will look for it, and the
installer agreeing that nothing is missing any more. The last of those
failed the first time -- the survey looked for node on PATH only, and so
could not see the one it had just installed.
2026-09-22 18:06:40 -07:00
jcoffey-dev f97fd12556 Survey the machine, and say what an install would do to it
The first two pieces of the installer, and deliberately the two that change
nothing: what this machine is, and what would happen to it.

The survey is the floor under every later choice -- a component can only
offer a container shape if Docker answers for this user, or a host shape on
a systemd machine with what that shape needs. It reports what it could not
establish as unknown rather than guessing: an unprivileged probe of port 25
means "I may not bind this", not "something is listening", and reporting
the first as the second is a lie an operator would act on.

The plan is one list, and it will have three readers: --dry-run prints it,
the interface will show it before anything happens, and apply will walk it.
What you are shown is what runs.

Tested on a fresh Debian 13 in qemu, which has neither Docker nor Node and
so exercises every unavailable shape: 23 checks, including that nothing was
created by any of it. The lab that machine runs in is in e2e/vm.
2026-09-22 17:56:20 -07:00