Voice mode is the feature everyone thinks is easy: transcribe, send it to the model, read the answer aloud. Build that and you get a two-minute spoken essay with headings read as sentences, a microphone that hears its own reply, and a five-second silence nobody can explain. This quest walks the version that actually works, built in cwkPippa from the last week of September 2026 and still growing, across ten tracks. Voice is a property of a single turn, so switching between talking and typing never rewrites the system prompt. Speech comes in through Scribe, batch and realtime, with a token the client can hold but a key it never sees, a Send now button for streets where silence never comes, a check of everything a talk leaves behind, and a transcriber hint drawn from the talk instead of the roster. The brain composes for the ear from the first token: cards for anything that has to be seen, a short line before every tool, and a reasoning level Dad sets for voice turns, with a deeper one when he asks for depth, chosen by measurement, and Lightning, one tap that sends a spoken turn to a different brain. Bellows turns each speakable unit into one whole synthesis and streams it from the first chunk; a chat answer is read whole again, by Dad's ruling after a day of breaks, with first sound waiting for a cushion sized so the voice never stops mid-answer. The hands-free loop gives every finished utterance's words a named place to go instead of dropping them, keeps his recording with them even when the server is briefly away, pauses without quitting, and says why when it stops; it also gains a camera she can ask for and a watch that takes one turn per press. Barge-in rests on echo cancellation, a loudness detector, a grace window born from a laugh, a pitch check that tells her own voice from his, and an honest estimate of how far you heard. Then Whose Voice: every dictated turn's recording is kept and goes to a speaker ear before Dad's message is stored, Mom's voice is learned from Dad's corrections, a noisy-place mode peels a recording into layers, and a voice the ear knows is not Dad's can no longer cut her off, though a check that can't answer in time still lets the cut through. And when the talk is broadcast, a third fact on the turn puts her on air: a screen fit to film, a face made for the camera, effects she writes herself, a morning prep for a broadcast that never announces itself, a replay on the broadcast's own clock, a pre-flight check the machine makes, and the same screen without the show. Along the way: the voice itself, from a stock ElevenLabs voice to a v3 clone to a fresh Eleven v4 clone swapped in a single day, and the latency floor measured instead of guessed, once at a desk and again over LTE in Málaga, where spoken turns gained a lane of their own. Conceptual open-sourcing: the architecture is the deliverable, never a repository to clone.