[bug] Hands-free says "Listening…" and shows no words — the composer still binds voice.interim, which stopped carrying speech in #281

[serge-ivo]

Reported

"In hands-free mode, when it has just started and says Listening, it is not transcribing or showing
any text/dictation. Only after I spoke for a while."

The composer reads a variable that stopped carrying speech

packages/sdk/src/voice/use-voice.ts:151:

// `interim` is now ONLY the composer's transient NOTICE line (mic errors, the wrong-language
// …
const [interim, setInterim] = useState("");

#281 promoted the live utterance to a first-class Dictation object (voice.dictation — text +
dictating/transcribing/failed status, words never blanked by a status change) and repurposed
interim as the notice line. The console was not updated with it. InstanceDetail.tsx:946:

value={voice.interim || input}

and :953 still styles the box for live speech when it is set:

${voice.interim ? "border-accent text-accent italic" : voice.micOn ? "border-green" : "border-line"}

So the accent-italic "you are speaking" treatment now fires only on a mic error, and the input can
never display speech — because speech no longer goes there.

What the user sees

In hands-free the top box shows its placeholder ("Listening…", :951) and stays empty for the whole
utterance. The words are being captured correctly and rendered correctly — as the voice.dictation
bubble at :1162, which is the last child of the message thread, i.e. below the fold, at the
opposite end of the screen from the input the user is watching.

Nothing is lost and nothing is broken in the voice pipeline. The live transcript is simply drawn
somewhere the reporter was not looking, while the place they were looking asserts "Listening…" and
shows nothing.

Two surfaces, one of them dead

This is the part worth fixing rather than papering over: there are currently two live-transcript
surfaces in the chat tab, and one of them has been disconnected since #281 without being removed. A
reader of InstanceDetail.tsx cannot tell which is authoritative, and the styling actively implies
the dead one is.

Fix

Stop binding the composer to voice.interim. voice.dictation is the single live-speech surface and
should stay that way.

interim still needs a home — it is the only thing that surfaces mic errors and the wrong-language
warning today, so it cannot simply be dropped. A small banner near the controls, not the input's
value (which also makes the input readOnly while a notice shows, :952).

Open, not proven

The dead binding fully explains "the box never shows my words". It does not necessarily explain
"only after I spoke for a while" — that half may additionally be the Web Speech gate's start-up
latency, since the gate is constructed and start()ed per turn (use-voice.ts:275-343) and cannot
emit onInterim until the engine boots. Worth measuring once the binding is fixed and the words are
somewhere visible; it may disappear entirely as a reporting artefact.

Related: #281 (the dictation object), #319 (the stored live capture), #365 (the layout change that
supersedes this).

[serge-ivo]

Fixed in b99a381 (CI green).

What the composer binds now: the typed draft, and nothing else. resolveComposer({draft, dictation, notice}) (new packages/sdk/src/voice/composer.ts, exported from @proagentstore/sdk/hooks) returns {value, readOnly, notice} — value is always the draft, readOnly means an utterance is in flight, and the notice is handed back to be rendered as its own line.

I went with the fix as written in the ticket rather than re-pointing the composer at voice.dictation, and the code agrees with the ticket: the bubble keeps the words through transcription, keeps them when transcription FAILS, and carries the status chrome ("Transcribing…", the failure note, Dismiss) that a <textarea value=…> cannot express — so putting dictation back in the box would have recreated the two-surface problem with the weaker of the two surfaces styled as authoritative. #365's own comment lands in the same place ("exactly one surface showing them"), and its move of the composer below the thread puts the box directly under the bubble, so the words are now next to where the user is looking.

Three other things came out of it:

• interim is now notice. The name is what let this survive three releases: it stopped carrying speech at [bug] Spoken words vanish between end-of-turn and the reply — dictation should land in the thread as a pending message with a transcribing status #281 and went on reading like it did. The one caller that legitimately consumed it (shouldShowComposer) now reads notice.
• A notice no longer locks the composer. It set readOnly for the 4.5–6s it was on screen, so "⚠ Whisper failed" also stopped you typing the message yourself. Only an in-flight utterance locks now; a failed one deliberately does not.
• The Coder Co-pilot had the identical dead binding (agents/coder/web/src/CopilotView.tsx), plus a comment asserting the input was its live-transcript surface. Fixed the same way — it has the same dictation bubble in its thread.

composer.test.ts scans the console / admin / coder-web trees for a voice string bound into an input value and asserts it would catch the exact line that shipped — a pure resolver only holds the rule if every consumer goes through it, and this binding type-checks, renders and looks deliberate.

Not addressed, and still open as the ticket says: the "only after I spoke for a while" half. dictation is null until the first partial arrives, so in hands-free the gap between the mic opening and the first word is real — the status pill ("Listening…") is the only thing filling it. Worth a fresh measurement now the words are somewhere the reporter is looking; it may well have been the dead binding all along.