Inbox Voice Walker (go through my inbox)

What it is

During a walk, each waiting session is read aloud from its spoken recap, and then you have one turn to say what to do with it:

You say	What happens	Confirmed first?
"reply with X"	Sends X to that session	Yes — say yes out loud
"ask it Q"	Asks that session your question	Yes — same spoken yes
"archive"	Archives that session	No — immediate
"snooze"	Snoozes that session	No — immediate
"skip"	Leaves it and moves on	No
"repeat"	Reads the same session again	No
"stop"	Ends the walk	No

A session with no spoken recap is simply read from its last message instead, so the walk never goes silent. A session that leaves the inbox mid-walk is skipped, not revisited.

Where to find it

• On desktop: the small sound-wave icon in the inbox header — hover it to see Go through my inbox — or just say the phrase. While a walk runs the same icon becomes a stop square with a short count beside it (for example "2/5"); hovering it reads Stop · 2 of 5. It is an icon rather than a labelled button so the header row keeps room for the inbox title.
• On the phone: the same control on the mobile inbox. A tap starts the walk; from then on it is hands-free.
• The walk is in development, so it ships hidden: turn on Voice — Inbox Walk in Settings → Lab (or set AMC_SHOW_VOICE_INBOX_WALK=1) to see it. Once on, it only ever starts when you say the phrase or press the button — never on its own.
• It borrows three other voice switches, so check them if the walk is silent or deaf: voice input (voiceEnabled - without it the walk hears nothing), spoken replies (ttsEnabled plus a voice key - without it nothing is read aloud), and Ask a live session (voiceL4Enabled - without it "ask it ..." is refused; replies and archive do not need it).

How it behaves

• It reads in inbox order and never re-sorts, so the sequence matches what you see on screen.
• One turn per session: it reads, listens once, acts, and steps on. If it does not catch a command it asks once more and then moves on rather than stalling.
• Sending is gated, and you hear what is being sent. Before a reply or a follow-up question goes, Omniscio reads it back - Send "ship it today" to Alpha job? - and waits for a yes (a long message is read as its opening, and says so). Only a plain yes counts: "no", silence, a timeout, or a yes buried in other talk ("okay, let me think about it") all mean no. Archive and snooze are immediate because both are reversible.
• What you said is what is sent - the same words, capitals and punctuation, with nothing added.
• You hear the answer to a follow-up. When the agent has answered, Omniscio reads it out between sessions; at the end it waits a little for answers still being worked, then tells you which ones have not come back yet. If a question could not reach an agent (the session is not running, or it is still busy with your last question) it says so instead of claiming it asked.
• It never asks you to look. Every outcome is confirmed in a short spoken sentence naming the session.
• It never hears itself. The microphone is only open between spoken lines.
• Hold-to-talk dictation wins. While you are dictating with the OS-wide hold key, the walk gives up the microphone entirely — a dictated sentence can never be read as a walk command.
• On the phone: the walk needs the screen on; if the phone sleeps or you switch away it either picks up where it left off or tells you it stopped. Everything is spoken on the phone itself - the computer it is paired with stays silent - and your questions and their answers never touch the computer's own voice.

For agents

• Settings key voiceWalkEnabled (default false - in development, hidden); unreleased-feature id voice-inbox-walk; every renderer read goes through isUnreleasedFeatureVisibleInRenderer, never the raw key. Voice action voice:walk-start with the phrase "go through my inbox".
• Another voice engine can speak the whole walk (every line, each session's recap or last message) by registering a speaker with registerWalkSpeaker (walk-speaker-port.ts); until one does, the walk speaks with the app's own text-to-speech. A speaker resolves only once its speech has ended.
• The walker's contract is inbox-voice-walker-contract.md; its map is inbox-voice-walker.md.

Related

Voice, TTS & Wake Word (hands-free Omniscio) — the read-aloud, dictation and voice-command stack the walk builds on.