Skip to content

Switchboard 0.4.2: voice server memory and idle CPU fixes #164

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
sameeeeeeep merged 1 commit into from
Oct 2, 2026
Merged

Switchboard 0.4.2: voice server memory and idle CPU fixes #164

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 15 additions & 1 deletion examples/god/tts/god-tts-server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,9 +112,23 @@ def _normalize_wav(src: Path, dst: Path, start: float = 0.0, dur: float | None =
# standard MLX-under-a-server pattern; without it /speak fails with "no Stream(gpu, 0)".)
_gpu = ThreadPoolExecutor(max_workers=1, thread_name_prefix="god-tts-gpu")

# MEMORY: MLX keeps every freed Metal buffer in its allocator cache, and the default cache limit is the
# whole recommended GPU working set (~5.7GB on an 8GB M1). Each synth allocates fresh activations, so the
# cache ratcheted up to ~5GB of DEAD buffers held forever by an idle voice server — the machine swapped
# (measured 2026-10-02: 5.25GB phys_footprint, 5.0GB of it IOAccelerator). The model's live weights are a
# small fraction of that. Cap the cache and hand freed buffers back after every GPU job; the compiled
# kernels (the expensive ~22s cold-compile keep-warm protects) live elsewhere and are unaffected.
CACHE_LIMIT_MB = int(os.environ.get("GOD_TTS_CACHE_MB", "128"))
mx.set_cache_limit(CACHE_LIMIT_MB * 1024 * 1024)


def _on_gpu(fn, *a, **k):
return _gpu.submit(fn, *a, **k).result()
def job():
try:
return fn(*a, **k)
finally:
mx.clear_cache() # return this job's freed activations to the OS (runs on the MLX thread)
return _gpu.submit(job).result()


def _model_get():
Expand Down
4 changes: 2 additions & 2 deletions packages/menubar/Info.plist
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
<key>CFBundlePackageType</key>
<string>APPL</string>
<key>CFBundleShortVersionString</key>
<string>0.4.1</string>
<string>0.4.2</string>
<key>CFBundleURLTypes</key>
<array>
<dict>
Expand All @@ -26,7 +26,7 @@
</dict>
</array>
<key>CFBundleVersion</key>
<string>43</string>
<string>44</string>
<key>LSMinimumSystemVersion</key>
<string>13.0</string>
<key>LSUIElement</key>
Expand Down
3 changes: 2 additions & 1 deletion packages/menubar/LooseNotch.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -150,7 +150,8 @@ struct CreatureView: View {
let shape = Capsule()
ZStack {
shape.fill(Color.page)
NotchField(accent: tint, working: model.working, animated: model.running, eyes: true,
// Docked = panel ordered out but this view still alive; its 24 fps TimelineView kept ticking unseen.
NotchField(accent: tint, working: model.working, animated: model.running && model.creatureLoose, eyes: true,
eyeColor: model.cameraOn ? Color.senseCamera : (model.listening ? Color.senseAudio : nil),
gaze: { loose.creatureCenter() })
.padding(.horizontal, 6).padding(.vertical, 3)
Expand Down
83 changes: 70 additions & 13 deletions packages/menubar/RelayMenuBar.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1278,8 +1278,11 @@ struct Panel: View {
// USE B: the hero liveness BEACON — a real 7×5 operator lamp field replacing the flat dot.
// The phase is carried by the PATTERN (thinking idle · working busy · listening signed-out),
// one accent only (lime, or the sanctioned danger when signed-out; faint + still when offline).
// Animate ONLY while the panel is on screen: an ordered-out NSHostingView keeps its
// TimelineView ticking, and this 35-view HStack/VStack grid re-laid-out the WHOLE hidden
// panel 24×/s (~22% idle CPU, measured 2026-10-02). Same fix as PanelDotField's `paused`.
DotMatrix(pattern: heroPattern, accent: heroBeaconAccent, cols: 7, rows: 5, dot: 2.6, gap: 2.4,
animated: model.running)
animated: model.running && model.panelVisible)
.frame(width: 34, alignment: .leading)
Text(heroTitle).font(.display).foregroundColor(heroColor).lineLimit(1)
}
Expand Down Expand Up @@ -1812,8 +1815,13 @@ struct Panel: View {
Text(sub).font(.hanken(10.5)).foregroundColor(.inkFaint)
}
Spacer(minLength: 6)
// Refresh once a second so "detected" decays back to "used Ns ago" on its own.
TimelineView(.periodic(from: Date(), by: 1)) { ctx in testPill(pulse: pulse, now: ctx.date) }
// Refresh once a second so "detected" decays back to "used Ns ago" on its own — only while the
// panel is on screen (Settings stays open across hides; a hidden 1 Hz tick re-laid-out the panel).
if model.panelVisible {
TimelineView(.periodic(from: Date(), by: 1)) { ctx in testPill(pulse: pulse, now: ctx.date) }
} else {
testPill(pulse: pulse, now: Date())
}
trailing()
}.padding(.vertical, 7)
}
Expand Down Expand Up @@ -2064,7 +2072,9 @@ struct Panel: View {
} else {
HStack(spacing: 5) {
Circle().fill(Color.lime).frame(width: 5, height: 5)
.opacity(breathe ? 1 : 0.3).animation(.easeInOut(duration: 0.8).repeatForever(autoreverses: true), value: breathe)
// Breathe only while the panel is on screen (a repeatForever keeps the hidden panel's
// display cycle running); `breathe` follows panelVisible, nil animation = stop instantly.
.opacity(breathe ? 1 : 0.3).animation(model.panelVisible ? .easeInOut(duration: 0.8).repeatForever(autoreverses: true) : nil, value: breathe)
Text("try it").font(.hanken(10, .semibold)).foregroundColor(.lime)
}.padding(.horizontal, 9).padding(.vertical, 5).background(Capsule().stroke(Color.lime.opacity(0.4), lineWidth: 1))
}
Expand Down Expand Up @@ -2283,7 +2293,8 @@ struct Panel: View {
)
.clipShape(NotchDropShape()) // no stroke — the black shape blends into the notch, no grey line
.ignoresSafeArea()
.onAppear { breathe = true }
.onAppear { breathe = model.panelVisible }
.onChange(of: model.panelVisible) { v in breathe = v }
.onChange(of: showSettings) { open in if open { onboard.note(.settings) } } // tour step 2
}
}
Expand Down Expand Up @@ -2482,13 +2493,18 @@ struct OrbView: View {
var onOpen: () -> Void
@State private var hovering = false
@State private var teamPulse = false
private func setTeamPulse(_ on: Bool) {
if on { withAnimation(.easeInOut(duration: 2.6).repeatForever(autoreverses: true)) { teamPulse = true } }
else { var t = Transaction(); t.animation = nil; withTransaction(t) { teamPulse = false } }
}
var body: some View {
// Health tint: lime = running + signed-in · red = running, signed-out · faint = daemon down.
let tint = model.running ? (model.signedIn ? Color.lime : Color.danger) : Color.inkFaint
let shape = NotchDropShape(ear: 9, botR: 10)
ZStack {
shape.fill(Color.page) // the black notch body — a notch on any Mac
NotchField(accent: tint, working: model.working, animated: model.running, eyes: (model.journalOn || model.cameraOn || model.listening) && !model.creatureLoose,
// While the creature is loose the orb is alpha-0 but still ordered front — don't animate an invisible field.
NotchField(accent: tint, working: model.working, animated: model.running && !model.creatureLoose, eyes: (model.journalOn || model.cameraOn || model.listening) && !model.creatureLoose,
eyeColor: model.cameraOn ? Color.senseCamera : (model.listening ? Color.senseAudio : nil),
gaze: { LooseNotch.shared.notchCenter() })
.padding(.horizontal, 6).padding(.top, 1).padding(.bottom, 4)
Expand Down Expand Up @@ -2527,7 +2543,10 @@ struct OrbView: View {
.onChanged { v in LooseNotch.shared.pullFromNotch(translation: v.translation, ended: false) }
.onEnded { v in LooseNotch.shared.pullFromNotch(translation: v.translation, ended: true) })
.onTapGesture { onOpen() } // a deliberate CLICK opens the panel
.onAppear { withAnimation(.easeInOut(duration: 2.6).repeatForever(autoreverses: true)) { teamPulse = true } }
// The team heartbeat runs ONLY while a team is live — a repeatForever started unconditionally at
// launch kept the always-on orb's display cycle busy for every user, team or not.
.onAppear { setTeamPulse(model.teamActive) }
.onChange(of: model.teamActive) { on in setTeamPulse(on) }
.help("Click to open Switchboard")
}
}
Expand Down Expand Up @@ -5201,6 +5220,7 @@ struct ActionConsentDrop: View {
IgnitionController.shared.install() // arms ~/.relay/ignite: the full-screen dot-matrix ignition (onboarding Frame 0) → chains into the tour
installUpdateCheck() // daily GitHub release check → ONE notch card when a newer build ships
installOpenWrappTrigger() // ~/.relay/open-wrapp.json → open that wrapp in the native bridged window (CLI threads can launch surfaces)
refreshInstalledVoiceServer() // an app update carries a newer god-tts-server.py → swap it into an installed voice engine
// Feedback capture: a fail (or fn↓) during a guide raises the notch note field + arms the fn-drag grab.
CursorGuide.shared.onFeedbackBegin = { [weak self] _ in Task { @MainActor in self?.showFeedbackNote() } }
CursorGuide.shared.onFeedbackEnd = { [weak self] in Task { @MainActor in self?.hideFeedbackNote() } }
Expand Down Expand Up @@ -8990,6 +9010,10 @@ struct ActionConsentDrop: View {
// ambient notch is already there the instant the phase drop clears (no flash of nothing).
orb?.orderFrontRegardless()
godStatusPanel?.orderOut(nil)
// Drop the content too: showGodStatus() builds a fresh hosting view on every show, and an ordered-out
// NSHostingView keeps its NotchFieldLED / DotMatrix TimelineView running (30 fps forever after the
// first God turn or dictation).
godStatusPanel?.contentView = nil
}

// ── Ambient mode ─────────────────────────────────────────────────────────────────────────────
Expand Down Expand Up @@ -9078,6 +9102,7 @@ struct ActionConsentDrop: View {
ambientSuppressUntil = Date().addingTimeInterval(120) // a manual dismiss hushes ambient for 2 min
}

private var glyphKey = ""
private func poll() {
checkReachable { ok in
self.checkWorking { busy in
Expand All @@ -9086,15 +9111,24 @@ struct ActionConsentDrop: View {
// Rung 4: only meaningful once the daemon is up. When up-but-signed-out, the glyph
// goes RED and the tooltip names the one fix — the cliff caught before the first call.
let signedIn = ok ? readSignedIn() : true
self.model.running = ok
self.model.working = ok && busy && signedIn
self.model.signedIn = signedIn
self.model.updateAvailable = updateReady
// Write @Published state ONLY on change: an unconditional assignment fires objectWillChange
// every 1.6 s, re-rendering every view that observes Model (the orb, the hidden panel…).
let working = ok && busy && signedIn
if self.model.running != ok { self.model.running = ok }
if self.model.working != working { self.model.working = working }
if self.model.signedIn != signedIn { self.model.signedIn = signedIn }
if self.model.updateAvailable != updateReady { self.model.updateAvailable = updateReady }
self.phase += 1
self.statusItem.button?.image = glyphImage(running: ok, working: self.model.working, signedIn: signedIn, phase: self.phase)
self.statusItem.button?.toolTip = !ok ? "Switchboard — sidekick offline"
// The glyph only changes with state (+ the blink phase while working) — skip identical redraws.
let glyphKey = "\(ok)|\(working)|\(signedIn)|\(working ? self.phase % 2 : 0)"
if glyphKey != self.glyphKey {
self.glyphKey = glyphKey
self.statusItem.button?.image = glyphImage(running: ok, working: working, signedIn: signedIn, phase: self.phase)
}
let tip = !ok ? "Switchboard — sidekick offline"
: !signedIn ? "Switchboard — \(SIGN_IN_HINT)"
: (ok && busy) ? "Switchboard — your model is working…" : "Switchboard — connected"
if self.statusItem.button?.toolTip != tip { self.statusItem.button?.toolTip = tip }
if self.panel.isVisible { self.model.refreshFiles(); self.ollama.refresh() }
}
}
Expand Down Expand Up @@ -9241,6 +9275,29 @@ struct ActionConsentDrop: View {
DispatchQueue.main.asyncAfter(deadline: .now() + 2.5) { [weak self] in self?.model.toast = nil }
}

/// The voice engine is opt-in: install-voice-engine.sh copies the bundled god-tts-server.py into
/// ~/.relay/tts ONCE, so a fix to the server (e.g. 0.4.2's MLX cache cap: 5.3 GB → ~0.6 GB resident)
/// would never reach a machine that already has the engine. On launch, if the engine is installed and
/// its server differs from this build's copy, replace it and restart the LaunchAgent. Never installs
/// the engine where it isn't (no venv/model download) — only refreshes the one script.
private func refreshInstalledVoiceServer() {
DispatchQueue.global(qos: .utility).async {
let fm = FileManager.default
let bundled = ((Bundle.main.resourcePath ?? "") as NSString).appendingPathComponent("god/tts/god-tts-server.py")
let installed = (RELAY_DIR as NSString).appendingPathComponent("tts/god-tts-server.py")
guard fm.fileExists(atPath: installed), let new = fm.contents(atPath: bundled),
let old = fm.contents(atPath: installed), new != old else { return }
do {
try new.write(to: URL(fileURLWithPath: installed), options: .atomic)
} catch { godLog("voice server refresh failed: \(error.localizedDescription)"); return }
godLog("voice server updated from the app bundle — restarting com.relay.godtts")
let p = Process()
p.executableURL = URL(fileURLWithPath: "/bin/launchctl")
p.arguments = ["kickstart", "-k", "gui/\(getuid())/com.relay.godtts"]
try? p.run(); p.waitUntilExit()
}
}

private func launchctl(_ args: [String]) {
let p = Process()
p.executableURL = URL(fileURLWithPath: "/bin/launchctl")
Expand Down
Loading