Eugene Plexus v0.1.0-alpha.6

The sixth public alpha. It remains a prerelease: not a stable release, and
not a claim of production readiness. Start with the
installation guide.

alpha.5 updates from the console: Nodes → Versions → Update on each
machine, or run the installer again. See Moving to alpha.6.
From alpha.6 that page is under Machines.

Changes since alpha.5

Workbench, Eugene's own chat app

• Install Workbench from the console (Apps), on any machine in your
  install. It is a chat app for the models your machines serve, with web
  search when you turn it on, attachments, and answers that keep going if you
  close the tab.
• You sign in with Eugene. With nobody added, Workbench asks for Eugene's
  passphrase. Add people under People in the console and each signs in
  with their own name and password, and sees only their own chats.
• On a Windows service or a Linux system install, each app runs in an
  operating-system account of its own, so it cannot read Eugene's keys,
  passphrase or another app's key. Apps can send their logs to Eugene, and
  they appear on the machine's Logs page.
• The story of the first chat a person had in it, unedited:
  It looked itself up.

Web search, run by your install

• Add a search account under Backends: a SearXNG you run (free), or a
  Brave Search key. With both, the free one is tried first.
• Models search the web through any of the three doors: chat
  (web_search_options), Responses (web_search) and Anthropic Messages
  (web_search_20250305). Claude Code's WebSearch and Codex's live search
  are run by your install's search account.
• The search runs here; only the search terms go to the search service.
  A client key says whether it may search.
• image_generation runs on /v1/responses with an image model you have.

Settings built for your machine

• Build settings for this machine, on a model's page, measures the model
  on this machine's card and offers a choice between faster replies and a
  longer context, in tokens a second and tokens of context. Choose an accuracy
  level; the result is saved as a profile, and launching it uses exactly what
  was measured. It asks before stopping any model that is running.
• Mixture-of-experts models are scored for what they are: their experts
  can stay in system memory while the rest runs on the card. An 8 GB card is
  now offered a 30B mixture-of-experts model, measured at 46 tokens a second
  on such a card, where it was offered nothing that size before. The starter
  set gains that class, and after a Low build the builder can offer a smaller
  file of the same model.
• A running model says which profile it was launched from.

Clearer settings and screens

• The console is organised as Library · Backends · Gateway · Machines, and
  every setting is on one Settings page by topic, with a search box.
• Every setting shows the value in effect: unset, default, inherited and
  unknown are said as such, and a setting that waits for a restart says so.
• Updates follow released versions by default (Settings → Updates). A
  machine that never chose keeps what it followed.

Other changes

• MLX on Apple silicon is no longer experimental. It is checked on
  GitHub's macOS runners, which expose Metal; it has not been run on a
  physical Mac by us.
• A tool call a local model would get wrong is repaired: a named
  tool_choice on llama.cpp is answered through structured output, which took
  the measured models from 0 of 6 to 28 of 28.

Fixes

• Installing an app from another machine's console no longer loops you
  through sign-in.
• A profile with Flash attention on no longer fails to start. llama.cpp
  needs --flash-attn on, and a bare --flash-attn swallowed the next flag.
• Two runtimes of one model can be told apart: each shows its profile.
• A reply that searched no longer reads twice in Workbench. A model told
  to search can write an answer before searching; that text is now folded
  under Written before searching, and only the answer after the search goes
  back to the model.
• Searches on a Brave free-plan account are spaced out, so a second search
  in the same turn is no longer refused, and a used-up monthly quota is said
  as such.

Known issues

• The Flash attention box shows off when unticked, but llama.cpp then
  decides for itself and usually turns it on
  (agent #6).
• A confirmation inside a table row (for example The engine stops; the model
  files stay.) can widen the page so its buttons need a sideways scroll
  (ui #14).
• Brave's rate-limit headers are read as Brave documents them; they have not
  been seen from a live Brave account by us.
• A recovery checkpoint leaves out an app's data on a Linux system install
  (Workbench's chats are kept by systemd outside Eugene's folder), and has not
  been checked with an app installed on Windows
  (specs #11).

Verification and limits

• The exact pins passed Windows and Linux CI, the installer suites, and
  container acceptance before the image was published.
• The packaged installers passed a clean install on Windows and in WSL, as
  recorded in the release record.
• Upgrading alpha.5 in place on a Linux system install in WSL, as recorded
  there.
• Workbench, sign-in and app accounts each have an acceptance run in CI, on
  GitHub's Windows and Ubuntu runners for app accounts; web search was run
  against Claude Code, Codex, Chrome and a real SearXNG. See the records
  linked from the support matrix.
• Still owed:
  • an Intel Arc card;
  • two physical cards in one machine;
  • models on a passed-through NVIDIA card in the container;
  • a physical Mac;
  • modest-GPU capacity;
  • native Linux unattended boot;
  • the Windows reboot-before-sign-in checks;
  • moderated sessions with people new to Eugene.
• Use a trusted LAN or VPN, and finish first-run setup before enabling remote
  access.

Moving to alpha.6

• From alpha.5 or alpha.4: Nodes → Versions → Update on each machine,
  from any console (the page is under Machines once a machine runs
  alpha.6). Or run the installer again:
  • Windows:
    irm https://eugeneplexus.com/releases/v0.1.0-alpha.6/install.ps1 | iex
  • Linux or macOS:
    curl -fsSL https://eugeneplexus.com/releases/v0.1.0-alpha.6/install.sh | sh
• From alpha.3: run the installer again, the command above. It keeps your
  passphrase, models, client keys and joined machines.
• The container: select ghcr.io/eugene-plexus/control-plane:v0.1.0-alpha.6
  and keep the same data folder.
• From alpha.2: a fresh install, as the
  alpha.3 notes describe.

A recovery checkpoint made on alpha.5 restores alpha.5.

Distribution

The GitHub prerelease
provides install.ps1, install.sh, manifest.json and SHA256SUMS.

• The tag is annotated and immutable.
• The manifest records the specs commit and all seven component commits,
  the built UI export included. That export's join commands name the alpha.6
  installer, so a worker installs the version its root runs.
• Workbench installs from the agent's catalogue at a commit pinned in the
  agent, so the version an install offers is fixed by the release.

The Linux amd64 image is built from the tag and published only after container
acceptance passes. It passed 29 container checks with zero failures before publication. Image:
ghcr.io/eugene-plexus/control-plane:v0.1.0-alpha.6, immutable reference
ghcr.io/eugene-plexus/control-plane@sha256:89ba518c1b42e9255a9716890908d20aad9110a79785ef8c40c9c6fdb789c7f1.
Earlier alphas' tags, assets, containers and website installer URLs remain
available.

These pins fix Eugene's source, not every upstream dependency: Python packages,
uv, engine downloads and model catalogue data can change.