v0.18.8: Live Parakeet fast by default + truthful settings
Live Parakeet transcription is fast by default. One selected engine now means one behavior. Previously, choosing Parakeet still left live transcription paying a full model load per utterance, because the warm-sidecar setting was hidden (no UI control existed), defaulted off, and the "Warm: yes" status label meant "ran once since launch" rather than "warm server resident." On a real meeting that compounded into 4 of 5 utterances dropped while every surface claimed healthy. The sidecar now auto-enables whenever Parakeet is your engine and the example-server binary resolves, every status surface reports the observed state with the remedy attached ("Sidecar: unavailable (auto); cold per-utterance runs. Install example-server to enable"), and a legacy parakeet_sidecar_enabled = false that earlier versions wrote into every config file is recognized as a serializer artifact and treated as auto. Forcing the sidecar off is now explicit: parakeet_sidecar_enabled = "off". (#295, #296)
minutes setup --parakeet checks the binary live mode needs. If example-server is missing, setup says so and prints the build recipe instead of letting live transcription silently run slow. (#296)
Dead settings removed. Desktop-context domain allow/deny lists were accepted by the settings API and documented, but enforced nowhere (browser capture is window-title-only in v1). Removed until real enforcement exists. (#297)
Install: brew install silverstein/tap/minutes or cargo install minutes-cli, or grab the DMG below. npx minutes-mcp for the MCP server + dashboard.