herdr-voice 1.4.0

A fully local voice: Local OpenLive mode runs speech recognition, the model and the voice on your Mac, with no audio leaving it and no per-minute fees.

herdr plugin install brogrammerMW/herdr-voice   # new install, or reinstall to update
herdr-voice                                     # start talking (Grok by default)

What's new

• Local OpenLive mode. Whisper hears you and Kokoro speaks, both on the GPU inside a sandboxed helper that listens only on your Mac; Qwen 3.5 through Ollama drives the same 30 tools, confirmations and reports as the cloud models. It's the first entry in the orb's model menu. It's an advanced feature with a one-time setup from a source checkout (scripts/local-openlive-setup, needs Node 22+, pnpm 10+ and Ollama); until then the menu shows it as "(not set up)". Full guide: docs/local-openlive.md.
• Warm-up you can see. The orb is amber with the thinking bubble while the local models load, and blue once it's listening.
• Survives crashes. If the local speech helper dies, the voice relaunches it and reconnects in about 5 seconds.
• Setup survives updates. Local mode's models and inventory live in ~/.local/state/herdr-voice, outside the plugin folder that reinstalls replace. Existing caches are migrated once, automatically.
• Quieter turn-taking in local mode. Coughs, clicks and echo tails no longer become turns; sentences that trail off wait for you to finish.
• Honest refusals. When a confirmation is refused, the voice now says the action was not done, and keeps saying so if asked later.

Known limits

• Local mode replies about two seconds after you stop talking; Grok is faster. Choose it for privacy, offline use and cost.
• macOS with WebGPU only. No vision.

Full changes: v1.3.0...v1.4.0