Live Talk-Back (talk through your inbox)

What it is

The inbox walk reads each waiting session aloud and listens for one instruction. Live talk-back keeps that exact walk — the same sessions in the same order, the same commands, the same spoken yes — but lets you talk to it instead of speaking one fixed command at a time. While a session is being read out you can ask "what did it change in the login code?" and get an answer from that session's text, or say "reply with sounds good" and hear it read your reply back before it asks you to confirm the send.

Under the hood it uses a live speech-to-speech voice. Google Gemini is the default; OpenAI Realtime is a switch in Settings. It needs your own key for the provider you choose. The voice never acts by itself: it can only call the walk's own commands (reply, ask a follow-up, archive, snooze, skip, repeat, stop), and the app carries them out through the walk's usual checks.

Where to find it

• On desktop: the Talk it through button sits beside Go through my inbox in the inbox header. While a live walk runs it reads Stop live, and it shows Connecting… for the moment it takes to set up the voice. The keyboard shortcut is Alt+Shift+T (it starts a live walk, or stops the one that is running).
• On the phone: the same button appears in a bar just above the walk's own, on the inbox. Tap it to start; the phone asks for the microphone the first time.
• There is no command-line version: a live walk needs a live microphone and speaker, so it is reachable only from the app on desktop or on a paired phone.
• Settings → Voice Control → Live talk-back: choose the voice provider and model, set a daily spend limit, decide whether you can cut in, and see today's spend. The master on/off switch is the Live talk-back toggle in Settings → Lab.
• It ships hidden. Live talk-back is in development, exactly like the walk it rides on, so a normal install shows none of it: turn on Voice — Inbox Walk and Voice — Live talk-back in Settings → Lab (or set AMC_SHOW_VOICE_LIVE_ENGINE=1) to see the button, the shortcut and the Settings card. Nothing can be started, and no voice credential is handed out, while it is hidden.

How it behaves

• It is the same walk. It does not have its own inbox order or its own commands, and the ordinary Go through my inbox button, phrase and voice shortcut still start the ordinary walk exactly as before.
• Sending is gated. A reply or a follow-up question is read back to you, then the walk asks for a spoken yes. "No", silence and a timeout all mean no. Archive and snooze act immediately, as they do in the ordinary walk.
• You can cut in. Talk over the voice and it stops at once and listens. The microphone always uses echo cancellation, and cutting in is decided by the voice service's own signal, never by how loud the room is. Turn it off in Settings for strict take-turns.
• It does not hear itself. The microphone is shut while the walk's own voice is speaking, and for a moment after.
• Dictation wins. While you are dictating with the OS-wide hold key, the live walk gives up the microphone and never reads dictated words as a command.
• Words from a session are information, not instructions. Text that came from a session is handed to the voice as labelled data, and a command only counts when you actually spoke in that same moment, so text planted in a session cannot make the walk act.
• Your voice goes to the provider you pick. While a live walk runs, your microphone audio and the text of the session being read are sent to Google or OpenAI over a secure connection that uses a short-lived credential made from your own key. The app does not save the audio; what the provider keeps is governed by your account with them.
• Trouble never leaves you with a dead microphone. No key saved, a refused key, today's budget used up, no connection, or a microphone that cannot be opened: the walk says so once, out loud, and carries on with the ordinary walk. A connection that drops mid-walk is replaced quietly; only repeated failure falls back.
• It costs money by the second, so there is a daily limit. The default is $1.00 a day (set 0 for no limit). When you pass it you are told out loud once, a walk already under way keeps going, and new live walks that day use the ordinary walk.
• In this version the walk's own recap and confirmations ("Archived …", "Sent to …") are still spoken in the walk's usual voice; the live voice speaks the conversation. A phone must keep its screen on during a walk.

For agents

• Settings keys: voiceLiveEnabled (default false - in development, hidden for customers; the Lab toggle, and it also needs voiceWalkEnabled), voiceLiveProvider (gemini | openai), voiceLiveModel, voiceLiveDailyCapUSD (default 1, 0 = off), voiceLiveBargeIn (default true). Unreleased-feature id voice-live-engine (a child of voice-inbox-walk); every read - the screens and Main's credential mint - goes through the central unreleased-feature helper, never the raw key. Shortcut action id voiceLiveWalk (Alt+Shift+T).
• Main mints a single-use vendor credential per connection (voice-live:mint-credential), bills audio from the renderer's delta reports (voice-live:report-usage) and reports spend and key presence (voice-live:get-status); no vendor key ever reaches a renderer or a phone.
• The contract is voice-live-engine-contract.md; the map is inbox-voice-walker.md (its live talk-back section).

Related

Going through your inbox by voice (hands-free walk) — the walk this talks through, with its six commands and its spoken yes. Voice, TTS & Wake Word — the read-aloud, dictation and voice-command stack both build on.