Approvals & Questions by VoiceNEW

A voice request is gated exactly like a typed one. When the agent needs a decision, the prompt takes over the voice surface: the mic closes, the prompt is read out, and the mic reopens for your answer once it has finished speaking.

Nothing here guesses. An answer that is not clearly one thing is asked again, and an approval nobody answers is rejected.

Where prompts are read out

The ribbon is deliberate: voice started inside a chat never speaks prompts, because that chat already has its own approval and question cards a few centimetres away.

Which voice reads them

Approval and question prompts are read by the browser's built-in speech — or your phone's, on Mobile. They are never read by the text-to-speech provider voice you picked, even when one is active and speaking the replies. That covers the prompt, both re-asks and the timeout lines.

What Sophon says

An approval, at None, Low or Medium risk:

I need your approval. {description}. This is {risk} risk. Say yes to approve or no to reject.

At High or Critical risk the last sentence changes:

I need your approval. {description}. This is {risk} risk. Say 'yes, approve' to approve, or no to reject.

The description is the tool's own one-line description, and the risk word is the level. The prompt names the kind of action and its risk, not its arguments — read the card for those.

When an answer is unclear, the re-ask is one of:

Sorry, I didn't catch that. Say yes to approve or no to reject.
Sorry, I didn't catch that. This needs you to say 'yes, approve', or no to reject.

And when a request runs out of time:

The approval timed out, so I didn't go ahead.
I didn't get an answer in time, so I'm continuing without it.

How your answer is matched

A keyword counts when it is the whole answer, or when the answer starts with it followed by a space, a comma or a full stop.

On top of that:

• A reject intent always rejects, at every risk level. A plain "no" rejects straight away and is not asked again.
• The negation rule. An answer that starts as an approval but also contains any of no, not, nope, nah, don't, dont, cancel, stop, wait, deny, reject, negative, never is treated as unclear, so you are asked again. "Yes — wait, no" is never an approval.
• A plan's risk is its highest step's risk, so a plan containing one Critical step asks for the phrase.

Sophon Mobile applies the same rules, and answers approvals only for its own voice conversation.

Two re-asks, then a tap

An unclear answer is re-prompted at most twice. After that the mic closes and a hint appears in place of another attempt:

On the Voice page and the Command Bridge a fresh mic press reopens capture for one more attempt. The card stays on screen throughout.

Timeouts

The child timeout only extends a request that left its timeout at the default, and it only ever makes the request wait longer for you, never less.

The hint under the buttons

The Voice page's approval card always reads:

Tap a button, or say "yes" / "no"

It says that at every risk level, including High and Critical, where a bare "yes" will in fact be re-asked. Sophon Mobile's card is risk-aware — Tap a button, or say "yes, approve" or "no" for High and Critical, Tap a button, or say "yes" or "no" otherwise.

Answering by voice does not start a turn

Pressing the mic to answer a waiting prompt opens capture in answer mode: your audio is transcribed and handed back, and no agent turn is started. Your words are matched against the prompt, the decision or the answer is submitted, and only then does the agent carry on. That is why answering an approval out loud does not leave a stray message in the chat.

On the Command Bridge the single mic button becomes "Answer by voice" while a prompt is waiting, and a separate Stop voice button appears next to it. Stop turns voice off without answering; the prompt stays on its chat card and its timer keeps running.

Questions

Questions are read out on the Voice page and the Command Bridge. Mobile shows them on screen without reading them, and the chat ribbon never reads them.

How one is asked

Question 1 of 2. {question} Your options are: one, {A}; two, {B}. Say a number, or answer in your own words.

• The "Question N of M" lead appears only when there is more than one question. Questions are asked one at a time, and every answer is submitted together after the last one.
• The closing line above is the default, because questions accept free text unless they say otherwise. A fixed-choice question ends "Say the number of your choice." instead.
• A multi-select question adds "You can choose more than one." and takes every option you name.
• A question with no options at all ends with "Say your answer."

Answering by number

Spoken numbers cover one to four — one/first/1, two/second/2, three/third/3, four/fourth/4 — and only in a short answer of four words or fewer. "Two" picks option 2; "I think option two is best" is five words, so the number counts as an ordinary word and the answer is treated as free text.

"One" counts as the number 1 only as the first word, or directly after option, number, choice, and or or, because "the green one" is a pronoun.

Options five and up are read out with their numbers, but can only be chosen by name or by tapping. Wider number matching is roadmap.

Answering by name

Say the option's name. Symbols in a label are both read and matched as the words people say for them:

So "C plus plus" picks C++ and "C sharp" picks C#. If your speech recognizer writes the symbol instead of the words, the option will not match and you are asked again — tap it instead.

When two names match on a single-select question and one contains the other, the longer one wins: saying "Visual Studio Code" cannot help also saying "Visual Studio". Two genuinely rival names, or a spoken number naming a different option than the one you said, is ambiguous and is asked again.

Negating an option

If you negate an option, Sophon will not pick it. "Don't send" against Send and Archive never returns Send.

• On a question that accepts free text, your own words are passed to the agent as the answer instead.
• On a fixed-choice question, you are asked again. The cost is real: "Email, no rush" re-asks on a fixed-choice question even though you named an option.
• Negation words that are part of an option's own name do not count, so "never ask again" still answers an option called "Never ask again".

Where question cards come from

Plans and other tasks raise them. A voice turn does not raise one itself: the voice tool set has no ask tool, so a voice agent is set up to ask you conversationally instead.

Answer once, and it closes everywhere

When an approval is decided — by voice, on a chat card, in another tab, in the mobile app, or by replying in the channel it was delivered to — the resolution is broadcast with its outcome: approved, rejected, timed out or cancelled. Any other open chat card, Mobile card or voice prompt for it closes on its own. Questions close the same way when they are answered or cancelled.

The Dashboard's Approvals page is not on that broadcast. It catches up on its own poll, every 5 seconds.

Each approval is recorded exactly once. A timeout or a cancel is never recorded as an approval, and no client may assert one.

Outcomes

The live broadcast carries only four of these — approved, rejected, timed out or cancelled. An auto-approval is broadcast as approved, and an edit resolves the wait as an approval too; both keep their own label in the approval history.

Cancelled is new. A cancelled approval or question used to be recorded and announced as "timed out". It is now recorded as cancelled, voice does not read the timeout line for it, and the cards show Cancelled — including after a reload.

Only your own conversation

Voice reads out only prompts whose chat session, or escalation session, is the one the voice conversation is attached to. An approval raised by an unrelated session never interrupts you. Sophon Mobile applies the same scoping.

A plan that finishes after you approve it

If a plan pauses for a durable approval and finishes later, its result is spoken to a voice session still attached to that chat — but only when no voice turn of that session is in flight at that moment. If you are mid-conversation when the plan lands, the result goes to the chat thread only, so it never talks over the reply you are waiting for. It is in the chat thread in both cases.

Where to go next

• Conversations — how a voice turn runs, hands-free, and interrupting
• Spoken Replies — which voice speaks what, and when a reply is silent
• Surfaces — what each of the four surfaces does differently
• Approval Gates — the risk levels and thresholds behind these prompts
• Limits — what voice does not do yet