Backchannel v0.3.7

Backchannel v0.3.7

The analysis agents can now run against a self-hosted OpenAI-compatible
server instead of a cloud provider. Paired with local ONNX transcription,
a deployment can run end to end on your own hardware with no API key from
anyone.

Highlights

• Self-hosted text models. A new openai-compatible entry in the model
  registry targets any OpenAI-shaped chat server: Ollama
  (http://localhost:11434/v1), LM Studio (http://localhost:1234/v1),
  vLLM, or LiteLLM. Set its base URL and model id in Admin -> API Keys and
  select it per agent like any other model.
• No credential required. Local servers usually want none, so this
  provider does not require an API key and sends no authorization header
  at all -- an empty bearer token breaks some servers rather than being
  ignored. Set OPENAI_COMPATIBLE_API_KEY only if your endpoint expects one.
• Layered configuration, unchanged defaults. The base URL resolves from
  the llm.openai_compatible.base_url app setting, then the
  OPENAI_BASE_URL environment variable, then the previous default. The
  app setting applies only to this provider, so a saved Ollama URL can
  never redirect a genuine OpenAI model. With nothing configured, every
  existing deployment behaves exactly as before.

Known limitation

Privacy First still switches the analysis agents off. Its gate recognizes
only the local ONNX transcription models, and an OpenAI-compatible URL can
point anywhere, so honoring it there needs a decision about whether to
require loopback. A fully local setup is configured through the endpoint
setting today, not by enabling that switch.

Not in this release

The WebSocket audio-handler refactor (ordered diarization worker, gateway
isolation, segment lifecycle serialization) is complete and reviewed on its
branch but is held back pending live dual-track call validation. It will
ship in a following release.

Upgrading

No migration or configuration change is required. Existing installs are
unaffected until an OpenAI-compatible base URL is set.