v2.2.0

Fixed

• Desktop window on Linux desktops other than KDE (Hyprland/Omarchy,
  GNOME, Sway): pywebview picked GTK while the tray is Qt, so no tray
  appeared, closing the window hid it for good, and on Omarchy GTK died with
  Error 71 (Protocol error) dispatching to Wayland display. The window now
  runs on the Qt backend on Linux
• Arch package: python-uvicorn does not exist on Arch (uvicorn), so
  pacman could not resolve the package; python-h2 and python-qtpy added,
  webkit2gtk (AUR-only) and gobject-introspection dropped
• Deploy workflow pointed at ~/apps/openvidia, which does not exist on the
  hub; it now updates ~/.local/share/openvidia
• The Shellcheck CI job named a script that no longer exists and passed only
  because shellcheck was missing on the runner
• Compaction summaries on a retired model. summary_model defaulted to
  meta/llama-3.1-8b-instruct, which NVIDIA no longer serves: every
  compaction spent its attempts on 404s and fell back to trimming. The
  default is now empty (the user's selected model), a 404 stops key rotation
  at once, and a summary model that 404s is retried once on the user's model
• HTTP/2 to the upstream can be turned off with OPENVIDIA_HTTP2=0 (#10)

Added

• Upstream health badge, per-model health status and auto-refreshing model
  catalog in the dashboard
• Model catalog workflow: weekly, fails when a model id the code ships is
  no longer in GET /v1/models
• Issue forms for bugs, model presets and feature requests
• Release workflow publishes the wheel and sdist to PyPI through Trusted
  Publishing (pipx install openvidia)

Changed

• Linux installs no longer pull PyGObject and pycairo (unused with Qt;
  PyGObject builds from source and broke pip installs without headers)
• The repository is English-only: remaining Italian comments, docs and two
  prompt strings sent to models translated
• CI runs on the soliso-autcir-openvidia runner; release actions are
  pinned to commit SHAs and the OIDC token that can publish to PyPI exists
  only in the PyPI job
• Per-model presets follow the October 2026 catalog: deepseek-v4-pro and
  glm-5.2 (retired) removed, deepseek-ai/deepseek-v4.1-flash added with
  the sampling values from its model card
• Dashboard "Popular" filter lists the current free-tier chat models
• Repository URLs point to autcir/openvidia