Backchannel v0.3.8

Backchannel v0.3.8

Self-hosted models are now first-class. Every model an OpenAI-compatible
endpoint serves shows by its own name in the pickers, a workspace can register
more than one endpoint, and Privacy First keeps the analysis agents running
when they point at a server on your own machine or LAN.

Highlights

• Every self-hosted model, by name. An endpoint's served models each
  become a real registry entry (endpoint:<slug>:<model>), so a lab-hosted
  model shows under its endpoint's name next to Gemini and OpenAI instead of
  hiding behind one opaque placeholder. Ollama tags such as llama3.1:8b keep
  their colons intact.
• More than one endpoint. Register any number of OpenAI-compatible servers
  -- Ollama, LM Studio, vLLM, LiteLLM -- from the new Self-Hosted Models
  card in Admin -> API Keys. Connect & list models reads the server's model
  list so wire names are never typed by hand; each endpoint has its own test,
  enable, and remove controls and an On your network badge.
• Privacy First judges destination, not vendor. With Privacy First on, an
  endpoint on loopback, a private network, or a LAN hostname keeps the analysis
  agents running; a public host is still treated as cloud and blocked. A fully
  local setup -- local ONNX transcription plus a self-hosted model -- now runs
  with the switch on. Only interim live captions remain cloud-only.

Upgrading

No action required. A previously configured single OpenAI-compatible endpoint
is migrated automatically into a named endpoint on first launch, its key
carried over and its agents repointed. Env-var-only installs are unchanged.

Included

The showcase and demo data were also rebuilt on a self-contained fictional
scenario, so the public screenshots and sample workspace are reproducible end
to end.