[feature] the card is for voice typing, and the record button listens

[YJack0000]

Two changes that landed together because they are the same correction from the same device session: the Live Activity was doing too much, and the thing the user actually looks at while dictating was faking it.

They are one PR rather than two because they were developed in one working tree and interlock — the removal and the additions touch overlapping surfaces, and an intermediate state would not have built.

The meeting card is gone

It shipped in 1.15 and was removed after use. The founder was told twice that the Dynamic Island and the lock-screen card are two presentations of one Live Activity and cannot be separated — so dropping the island for meetings also drops the lock-screen card that motivated the whole feature — and reaffirmed. The island's constant presence cost more than the lock-screen card was worth for a recording he is sitting in front of.

• MicActivityState.Mode is down to .dictation and .standby; title went with the meeting name.
• MeetingRecorder stops driving the card. Its start(token:app:), weak var host, and the Darwin observer — all added by [feature] the microphone says so on the lock screen, and in your hand #393 purely for the card's Stop button — are reverted to their pre-[feature] the microphone says so on the lock screen, and in your hand #393 form.
• MeetingControlChannel, StopMeetingRecordingIntent and their catalog strings are removed.
• The meeting haptics added in [fix] the island stops covering the phone, and stopping a recording is felt #395 stay. They were never part of the card, and they are what the founder asked for.

Starting a meeting now makes a live card disappear rather than turn red. That is the behaviour, not a bug: yieldMicrophone() ends the dictation and closes the window, so both of derive's inputs go nil. The answer is written at derive, because that is where somebody chasing the disappearance will land.

The dictation card loses ⏹ and ✕

Deciding what happens to a transcript is not a choice to make on a surface you are glancing at with the phone face-up on a table. A dictation card is read, not operated; .widgetURL still opens Parley, where both decisions are one tap away.

Standby keeps its one button

A first pass removed this too, as a literal reading of "no buttons". That was wrong and it is restored. ⏹ and ✕ settled the fate of words the card deliberately does not show. Ending standby settles nothing about content — standby is precisely the state in which nothing is being recorded. It closes a microphone that is open.

And it is the surface that most needs to offer it: every place that announces a window has always also been a way to end one (the Settings picker, the Record tab's bar, the keyboard's chip), and this is the fourth and the only one reachable from a locked screen — exactly where someone who has just noticed an unexplained orange dot is looking.

The record button now listens

KeyboardRootView pulsed the record button on a fixed scaleEffect(animating ? 1.25 : 1) loop that had nothing to do with what was being said. That is the canned visualiser this project's visual rules exist to prevent — driven by real amplitude, flat in silence — and the founder spotted it on the device immediately.

The keyboard could not have done better: no amplitude crossed the App Group. So:

• A seventh mailbox (dictation-level.json) carrying one normalised level plus a stamp. Not a field on Downlink: that mailbox's staleness is the liveness watchdog's only input, and a value that moves whether or not a word does would refresh it forever — which does not weaken the watchdog, it switches it off.
• 12 Hz, which is the rate the measurement already arrives at (AudioCapture taps 4096 frames ≈ 85 ms), so a faster mailbox would mostly republish unchanged readings into a backgrounded process.
• Reuses the RMS the in-app meter already computes, published from the main-actor hop that meter already pays for — not from the audio thread, where a file write costs buffers. Gated by the existing reportsLevel gate, which already exists to stop an idle window spending battery on a number nobody reads.
• Silence is a value the app writes, not one the keyboard infers: every path that stops reporting forces a final silent write, and a stale reading reads as silence too, so nothing is ever left frozen mid-swell.
• Smoothing is asymmetric — fast attack so the swell lands with the syllable, slow release so the gap between words is a settle rather than a strobe. The ripple is a lagged copy of the same smoothed value drawn further out, so a syllable pushes one ring and then the next; that is a wavefront without a timer. In silence there is no ripple at all — the view leaves the tree.

A haptic for the microphone being taken

Tapping iOS's own dictation key in the strip below the keyboard opens a recording session in another process, and Parley stops hearing anything. That key cannot be hidden — it is system-drawn over third-party keyboards from iPhone X onward and there is no public API to suppress it; the only switch is the user's global Settings › General › Keyboard › Enable Dictation.

So the least Parley can do is tell the hand, at the one moment the user is most likely to keep talking into nothing. A flat pair rather than a rise or a fall: every other beat in the file answers something the user did, and direction is how they say what. This one answers something done to them, so it has no direction to borrow.

Deliberately not done

No waveform on the card. It cannot be driven by real audio — that would need activity.update() landing several times a second from a backgrounded app, which is this feature's one unverified assumption and a battery cost even if it worked. What is left is a loop that moves whether or not anyone is speaking: a picture of listening rather than evidence of it, on the surface whose entire job is to be believable. The reasoning is written where someone would otherwise add one.

Verification

• swift test --package-path ios/ParleyKit → 421 tests, 0 failures (1 pre-existing skip). Six new wire-format tests for the level mailbox, including staleness, the unstamped file, NaN/clamping, and one pinning the level off the downlink.
• xcodegen generate + xcodebuild -scheme Parley -configuration **Release** -destination generic/platform=iOS → BUILD SUCCEEDED.
• Both locales present for every catalog key that has localizations.
• Not verified — say it plainly: nothing here has been seen or felt. No device, and a simulator cannot reach a real dictation. The swell depth, the ripple reading as a travelling wave rather than concentric breathing, the attack/release constants, and the new beat's distinguishability from the existing four are all arguments. The smoothing constants are the thing most likely to want a tweak on a real phone.

🤖 Generated with Claude Code

[github-actions]

✅ SonarQube Quality Gate passed — pathorsAI_parley

0 open issues on this PR.