v0.2.1 (early release)

Install or update:

• macOS / Linux: curl -fsSL https://enowx.ai/install.sh | sh
• Windows (PowerShell): irm https://enowx.ai/install.ps1 | iex

Each archive has a .sha256 checksum beside it.

Interface work

• Everything that can be set is in Settings. New sections: General
  (browser preview, language server checks, background delegation,
  switching agent by itself, compaction and its threshold, model calls per
  turn, the shell timeout, the model for each tier), Display (sidebar,
  currency and rate) and TypeSafe. Each field is saved through the same
  checks as enx config set, and a bad value is refused with why.

• Structure before styling: ui-structure and ui-anatomy. The
  frontend, canvas and mobile agents now write a blueprint before any
  code: the screen's job and its one primary action, the regions its kind
  of page needs in order (homepage, dashboard, list, detail, settings,
  form, sign-in, pricing, docs, article, checkout) with why each is there,
  the skeleton in numbers (container, gutters, grid, spacing, what each
  breakpoint changes), and the components with their states. They build
  in a fixed order (tokens, layout primitives, components, sections, the
  page) and every control meets its contract in ui-anatomy: parts,
  variants, sizes, every state, keyboard.

• Reviewers judge structure before taste. Content touching the
  window's edge, sections on different left edges, mixed surfaces, cut-off
  phone layouts, selected states shown by a shade, invalid submits and
  missing loading, empty or error states rule out Good; a page not seen
  in a browser is never called good to look at.

• preview reports more: content touching the window's edge (no
  gutter), content cut off past it, and a theme toggle with text on it.

Agents and delegation

• The orchestrator and maestro can stop a delegation at work with
  stop_delegation and a reason; its report says who stopped it and why,
  and the list of delegations still running now gives their sessions.
• Each delegated sub-agent shows the model it runs on in the sidebar
  list and the log, so a glance shows what is answering each one.
• Stopping a sub-agent asks first. A right-click in the sidebar used to
  stop it at once; it now opens a confirm that defaults to keeping it, so a
  stray click stops nothing.

Models

• A model manager that tests before it assigns. A new model-manager
  skill, carried by the orchestrator and maestro, helps set a model per
  agent: enx models list shows the connected providers and their models,
  the model in use and the per-tier and per-agent ones; enx models test provider/model makes one tiny call and says whether it answered, how
  fast, and the tokens, failing with the provider's own error. The skill
  lists what you own, brainstorms a mapping, tests every candidate, then
  sets only the ones that answered.
• The footer and /model follow the active agent's model. Setting an
  agent to another model used to leave the footer on the old one until the
  next /model, and picking a model while an agent had its own changed the
  conversation's model, not the agent's. Both now track the agent that is
  answering.
• antares --version and --help work as commands (the CLI took a
  leading flag as an option of the default serve).