push-to-talk viability at the desk

[ojfbot]

Question

Is browser-native voice good enough for a hands-free grill at the desk — push-to-talk or continuous, and does the agent speak back?

Type

prototype (HITL, /prototype)

Context

Operator ruling D1: voice-capable at the desk, not in the car. The cockpit stays bound to 127.0.0.1.

Prototype the Web Speech API in the cockpit's browser:

• push-to-talk vs continuous listening
• TTS for the agent's turn — and whether it is wanted at all
• barge-in (interrupting a spoken answer)
• reduced-motion and a11y; the rail already honours reduced-motion elsewhere

Blocked deliberately on the cadence prototype: voice pushes hard toward one-question-at-a-time and would prejudge that decision if run first.

Note the fallback if voice proves unworkable here: core's queued chat-side skill northstar-voice (rm:rm-l2-ojfbot#S3) was specced for exactly this and never built.

Blocked by

'does a one-thread grill survive a 372px rail' — modality must not prejudge cadence.

Resolution (filled at close)

—