feat: selectable audio source

Add an audio source picker to the GUI so users can capture a specific
PipeWire device (sink monitor or microphone) instead of only the system
default output.

- vendor go2tv.app/screencast and patch the audio stream to accept a
  target PipeWire node serial (PW_KEY_TARGET_OBJECT); the upstream lib
  only ever auto-connected to the default
- capture: ListAudioSources enumerates PipeWire sinks/sources via pw-dump;
  OpenPipeWire takes the selected node serial
- flinger/config/gui: AudioSource config field, persisted and exposed as
  an Audio source dropdown (default output + enumerated devices)

Also fixes two pre-existing bugs surfaced by stop/start testing:
- engine Stop now waits for the video/audio/stats goroutines before
  destroying the encoder (was a use-after-free SIGSEGV)
- Start/Stop now pause/resume the engine instead of tearing down and
  re-opening the portal session, which the portal cannot reliably do
  in-process (2nd CreateSession returned Ended/cancelled). Capture
  session stays open across stop/start.

go.mod

@@ -44,3 +44,5 @@ require (
 	golang.org/x/text v0.42.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )
+
+replace go2tv.app/screencast => ./third_party/screencast

internal/capture/audio.go

@@ -0,0 +1,97 @@
+// Audio source enumeration for the settings UI.
+
+package capture
+
+import (
+	"encoding/json"
+	"os/exec"
+)
+
+// AudioDevice describes a capturable PipeWire audio node.
+type AudioDevice struct {
+	Serial   uint64 // PipeWire object.serial, passed to the capture backend
+	ID       uint32 // PipeWire node id (informational)
+	Name     string // node name, e.g. "alsa_output...analog-stereo"
+	Desc     string // human-readable description
+	IsOutput bool   // true = a sink (system output); false = a source (mic)
+}
+
+// pwDumpNode is the subset of `pw-dump` output we parse.
+type pwDumpNode struct {
+	ID   uint32 `json:"id"`
+	Info struct {
+		Props map[string]any `json:"props"`
+	} `json:"info"`
+}
+
+// ListAudioSources enumerates PipeWire audio sinks and sources via `pw-dump`.
+// To capture system audio we attach to a sink's monitor; to capture a
+// microphone we attach to an Audio/Source node. This lists the capturable
+// audio nodes so the GUI can present a picker instead of always using the
+// system default output.
+func ListAudioSources() ([]AudioDevice, error) {
+	out, err := exec.Command("pw-dump").Output()
+	if err != nil {
+		return nil, err
+	}
+
+	var nodes []pwDumpNode
+	if err := json.Unmarshal(out, &nodes); err != nil {
+		return nil, err
+	}
+
+	var sources []AudioDevice
+	for _, n := range nodes {
+		props := n.Info.Props
+		mediaClass, _ := props["media.class"].(string)
+		if !audioNodeClass(mediaClass) {
+			continue
+		}
+		desc, _ := props["node.description"].(string)
+		name, _ := props["node.name"].(string)
+		serial := toUint64(props["object.serial"])
+		if serial == 0 {
+			continue
+		}
+		sources = append(sources, AudioDevice{
+			Serial:   serial,
+			ID:       n.ID,
+			Name:     name,
+			Desc:     desc,
+			IsOutput: mediaClass == "Audio/Sink",
+		})
+	}
+	return sources, nil
+}
+
+// audioNodeClass reports whether a media.class is a capturable audio node.
+func audioNodeClass(mediaClass string) bool {
+	switch mediaClass {
+	case "Audio/Sink", "Audio/Source":
+		return true
+	default:
+		return false
+	}
+}
+
+// toUint64 best-effort converts a pw-dump property value to uint64.
+// PipeWire serials are small, so the conversions cannot overflow in practice.
+//
+//nolint:gosec // safe: JSON numbers from pw-dump are small node serials
+func toUint64(v any) uint64 {
+	switch t := v.(type) {
+	case float64:
+		return uint64(t)
+	case int:
+		return uint64(t)
+	case int64:
+		return uint64(t)
+	case uint64:
+		return t
+	case json.Number:
+		if n, err := t.Int64(); err == nil {
+			return uint64(n)
+		}
+	}
+	return 0
+}

internal/capture/audio_test.go

@@ -0,0 +1,64 @@
+package capture
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+// TestPWParseAudioNodes parses a representative pw-dump payload into the
+// node shape used by ListAudioSources.
+func TestPWParseAudioNodes(t *testing.T) {
+	payload := `[
+  {"id": 51, "info": {"props": {"media.class": "Audio/Sink", "node.name": "alsa_out_speaker", "node.description": "Speaker", "object.serial": 1179}}},
+  {"id": 56, "info": {"props": {"media.class": "Audio/Source", "node.name": "alsa_in_mic", "node.description": "Stereo Mic", "object.serial": 1183}}},
+  {"id": 64, "info": {"props": {"media.class": "Audio/Device", "node.description": "Not capturable", "object.serial": 1172}}}
+]`
+
+	var nodes []pwDumpNode
+	if err := json.Unmarshal([]byte(payload), &nodes); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+
+	var devs []AudioDevice
+	for _, n := range nodes {
+		props := n.Info.Props
+		mc, _ := props["media.class"].(string)
+		if !audioNodeClass(mc) {
+			continue
+		}
+		devs = append(devs, AudioDevice{
+			Serial:   toUint64(props["object.serial"]),
+			ID:       n.ID,
+			Name:     props["node.name"].(string),
+			Desc:     props["node.description"].(string),
+			IsOutput: mc == "Audio/Sink",
+		})
+	}
+
+	if len(devs) != 2 {
+		t.Fatalf("got %d capturable devices, want 2", len(devs))
+	}
+	// Sink serial parsed and flagged as output.
+	if devs[0].Serial != 1179 || !devs[0].IsOutput {
+		t.Errorf("sink wrong: %+v", devs[0])
+	}
+	// Source serial parsed and flagged as input.
+	if devs[1].Serial != 1183 || devs[1].IsOutput {
+		t.Errorf("source wrong: %+v", devs[1])
+	}
+}
+
+// TestAudioNodeClass verifies which media classes are capturable.
+func TestAudioNodeClass(t *testing.T) {
+	cases := map[string]bool{
+		"Audio/Sink":   true,
+		"Audio/Source": true,
+		"Audio/Device": false,
+		"Video/Source": false,
+	}
+	for cls, want := range cases {
+		if got := audioNodeClass(cls); got != want {
+			t.Errorf("audioNodeClass(%q) = %v, want %v", cls, got, want)
+		}
+	}
+}

internal/capture/pipewire.go

@@ -24,12 +24,14 @@ type PipeWire struct {
 }
 
 // OpenPipeWire opens a PipeWire capture session. streamIndex selects which
-// monitor to capture when multiple are present. Triggering the portal
-// consent dialog is expected; the compositor decides whether to show it.
-func OpenPipeWire(streamIndex int, audio bool) (*PipeWire, error) {
+// monitor to capture when multiple are present; audioSourceSerial optionally
+// selects a specific PipeWire audio node (0 = system default). Triggering the
+// portal consent dialog is expected; the compositor decides whether to show it.
+func OpenPipeWire(streamIndex int, audio bool, audioSourceSerial uint64) (*PipeWire, error) {
 	s, err := capture.Open(&capture.Options{
 		StreamIndex:       streamIndex,
 		IncludeAudio:      audio,
+		AudioSourceSerial: audioSourceSerial,
 	})
 	if err != nil {
 		return nil, err

internal/config/config.go

@@ -22,6 +22,7 @@ type Config struct {
 	FPS         int     `json:"fps"`
 	Source      string  `json:"source"`
 	Audio       bool    `json:"audio"`
+	AudioSource uint64  `json:"audio_source"`
 	StreamIndex int     `json:"stream_index"`
 	Scale       float64 `json:"scale"`
 	// Announce is a *bool so an absent JSON key (older config files) keeps
@@ -40,6 +41,7 @@ func Default() Config {
 		FPS:         c.FPS,
 		Source:      c.Source,
 		Audio:       c.Audio,
+		AudioSource: c.AudioSource,
 		StreamIndex: c.StreamIndex,
 		Scale:       c.Scale,
 		Announce:    &announce,
@@ -59,6 +61,7 @@ func (c Config) ToFlinger() flinger.Config {
 		FPS:         c.FPS,
 		Source:      c.Source,
 		Audio:       c.Audio,
+		AudioSource: c.AudioSource,
 		StreamIndex: c.StreamIndex,
 		Scale:       c.Scale,
 		Announce:    announce,

internal/flinger/flinger.go

@@ -48,6 +48,9 @@ type Config struct {
 	Source string
 	// Audio enables system audio capture and streaming.
 	Audio bool
+	// AudioSource is the PipeWire node serial to capture audio from (0 = system
+	// default output). Only applies when Audio is true and Source is "screen".
+	AudioSource uint64
 	// StreamIndex selects which monitor to capture (screen source only).
 	StreamIndex int
 	// Scale downscales the captured frame before encoding. 1.0 = native
@@ -122,8 +125,18 @@ type Engine struct {
 	start time.Time
 	stop  chan struct{}
 
+	// running gates packet sending: when false the loops consume capture
+	// frames but do not transmit. This lets the GUI "stop" without tearing
+	// down the portal/PipeWire session (which the portal cannot reliably
+	// recreate in-process), so a subsequent Start just flips running on.
+	running atomic.Bool
+
 	frames atomic.Int64
 
+	// wg tracks the audio/video/stats goroutines so Stop can wait for them to
+	// finish before destroying shared resources (encoder, sender, capture).
+	wg sync.WaitGroup
+
 	// scaleBuf is the cached downscaled frame, reallocated only when the
 	// configured scale or source dimensions change.
 	scaleBuf *capture.VideoFrame
@@ -162,6 +175,7 @@ func (e *Engine) SetConfig(cfg Config) error {
 	cfg.Source = e.cfg.Source
 	cfg.Port = e.cfg.Port
 	cfg.StreamIndex = e.cfg.StreamIndex
+	cfg.AudioSource = e.cfg.AudioSource
 
 	e.cfg = cfg
 	return nil
@@ -199,7 +213,7 @@ func New(cfg Config) (*Engine, error) {
 
 	switch cfg.Source {
 	case "screen":
-		cam, err := capture.OpenPipeWire(cfg.StreamIndex, cfg.Audio)
+		cam, err := capture.OpenPipeWire(cfg.StreamIndex, cfg.Audio, cfg.AudioSource)
 		if err != nil {
 			e.encoder.Close()
 			sender.Close()
@@ -222,11 +236,13 @@ func New(cfg Config) (*Engine, error) {
 // config) starts announcing the stream. It is idempotent.
 func (e *Engine) Start() {
 	if e.stop != nil {
+		e.Resume()
 		return
 	}
 
 	e.start = time.Now()
 	e.stop = make(chan struct{})
+	e.running.Store(true)
 
 	if e.cfg.Announce {
 		e.announcer = discovery.Start(e.cfg.Name, e.sender.Port())
@@ -243,32 +259,68 @@ func (e *Engine) Start() {
 		src = capture.NewSilenceSource()
 	}
 
-	go e.audioLoop(src)
-	go e.videoLoop()
-	go e.statsLoop()
+	e.wg.Add(3)
+	go func() {
+		defer e.wg.Done()
+		e.audioLoop(src)
+	}()
+	go func() {
+		defer e.wg.Done()
+		e.videoLoop()
+	}()
+	go func() {
+		defer e.wg.Done()
+		e.statsLoop()
+	}()
 }
 
-// Stop halts the loops, stops announcing and closes all resources. It is
-// idempotent. After Stop the engine must not be restarted.
+// Stop pauses the stream: loops keep consuming capture frames but stop
+// transmitting, and the portal/PipeWire session stays open so the engine can
+// be resumed with Start without re-opening the portal (which the portal
+// cannot reliably do in-process). Idempotent.
 func (e *Engine) Stop() {
+	e.running.Store(false)
+}
+
+// Pause is an alias for Stop; kept for clarity at call sites.
+func (e *Engine) Pause() { e.Stop() }
+
+// Resume restarts transmission on a paused engine. It is a no-op if the
+// engine was never started.
+func (e *Engine) Resume() {
+	if e.stop == nil {
+		return
+	}
+	e.running.Store(true)
+}
+
+// Close fully tears down the engine: stops the loops, closes the portal
+// session and frees the encoder/sender. After Close the engine must not be
+// reused.
+func (e *Engine) Close() {
 	if e.stop == nil {
 		return
 	}
 
 	close(e.stop)
-	// Give the loops a moment to observe the stop signal.
-	time.Sleep(50 * time.Millisecond)
+
+	// Closing the capture source unblocks the audio and video loops that may
+	// be stuck in a read. Do this before waiting so they can observe stop.
+	if e.cam != nil {
+		if err := e.cam.Close(); err != nil {
+			log.Printf("flinger: capture close: %v", err)
+		}
+	}
+
+	// Wait for the audio/video/stats goroutines to finish so they no longer
+	// touch the encoder or sender before we destroy them.
+	e.wg.Wait()
 
 	if e.announcer != nil {
 		e.announcer.Stop()
 	}
 	e.sender.Close()
 	e.encoder.Close()
-	if e.cam != nil {
-		if err := e.cam.Close(); err != nil {
-			log.Printf("flinger: capture close: %v", err)
-		}
-	}
 }
 
 // Status returns a snapshot of the running engine.
@@ -278,7 +330,7 @@ func (e *Engine) Status() Status {
 	e.errMu.RUnlock()
 
 	return Status{
-		Running: e.stop != nil,
+		Running: e.running.Load() && e.stop != nil,
 		Frames:  e.frames.Load(),
 		Dropped: e.sender.Dropped(),
 		Conns:   e.sender.NumConns(),
@@ -333,8 +385,17 @@ func (e *Engine) audioLoop(src io.ReadCloser) {
 	buf := make([]byte, chunkBytes)
 
 	for {
+		select {
+		case <-e.stop:
+			return
+		default:
+		}
+
 		n, err := io.ReadFull(src, buf)
 		if n > 0 {
+			if !e.running.Load() {
+				continue
+			}
 			frames := n / (speakers * 2)
 			ts := uint64(time.Since(e.start))
 			packet, perr := protocol.BuildWavePacket(ts, protocol.AudioFormatS16, sampleRate, speakers, int32(frames), buf[:n])
@@ -433,6 +494,12 @@ func (e *Engine) videoLoop() {
 	next := e.start
 
 	for {
+		select {
+		case <-e.stop:
+			return
+		default:
+		}
+
 		frame, err := e.loop.Next()
 		if err != nil {
 			select {
@@ -464,6 +531,13 @@ func (e *Engine) videoLoop() {
 		}
 		next = now.Add(frameInterval)
 
+		// When paused (Stop), keep consuming frames to stay alive but do not
+		// encode or transmit. The portal session is left open so a subsequent
+		// Start can resume without re-opening the portal.
+		if !e.running.Load() {
+			continue
+		}
+
 		ts := uint64(now.Sub(e.start))
 		enc := e.scaleFrame(frame, cfg.Scale)
 		buf, err := e.encoder.EncodeBGRA(enc.Pix, enc.Width, enc.Height, cfg.Quality)

internal/flinger/flinger_test.go

@@ -22,7 +22,7 @@ func TestEnginePatternStartStop(t *testing.T) {
 	}
 
 	eng.Start()
-	defer eng.Stop()
+	defer eng.Close()
 
 	// Connect a raw receiver and read until the engine reports frames sent.
 	conn, err := net.Dial("tcp", "127.0.0.1:"+itoa(eng.sender.Port()))
@@ -143,7 +143,7 @@ func TestBitrateMeasurement(t *testing.T) {
 		t.Fatalf("New: %v", err)
 	}
 	eng.Start()
-	defer eng.Stop()
+	defer eng.Close()
 
 	// Connect a receiver so packets actually flow, and wait for the bitrate
 	// window to produce a measurement.

internal/gui/gui.go

@@ -93,11 +93,17 @@ type App struct {
 	scaleSel    *widget.Select
 	nameEnt     *widget.Entry
 	audioChk    *widget.Check
+	audioSel    *widget.Select
+	audioDevs   []capture.AudioDevice
 	announceChk *widget.Check
 	srcSel      *widget.Select
 	monSel      *widget.Select
 	monitors    []capture.Monitor
 	statsDone   chan struct{}
+
+	// lastStart records the capture-affecting config the current engine was
+	// created with, so Start can resume instead of reopening the portal.
+	lastStart config.Config
 }
 
 // Run starts the GUI and blocks until the app exits. configPath selects a
@@ -188,8 +194,9 @@ func (g *App) buildUI() {
 	g.presetSel.SetSelected("High")
 
 	// Audio.
-	g.audioChk = widget.NewCheck("Capture system audio", nil)
+	g.audioChk = widget.NewCheck("Capture system audio", func(bool) { g.applyLiveSettings("") })
 	g.audioChk.SetChecked(g.cfg.Audio)
+	g.setupAudioPicker()
 
 	// Announce over multicast.
 	g.announceChk = widget.NewCheck("Announce on LAN", nil)
@@ -217,7 +224,8 @@ func (g *App) buildUI() {
 			{Text: "Quality", Widget: g.qualitySel},
 			{Text: "Frame rate", Widget: g.fpsSel},
 			{Text: "Scale", Widget: g.scaleSel},
-			{Text: "", Widget: g.audioChk},
+			{Text: "Audio", Widget: g.audioChk},
+			{Text: "Audio source", Widget: g.audioSel},
 			{Text: "", Widget: g.announceChk},
 		},
 	}
@@ -283,6 +291,54 @@ func (g *App) selectedMonitorIndex() int {
 	return g.cfg.StreamIndex
 }
 
+// setupAudioPicker populates the audio source dropdown from PipeWire. The
+// first option is "Default output"; the rest are the enumerated sinks and
+// microphones. A no-op if enumeration is unavailable.
+func (g *App) setupAudioPicker() {
+	g.audioDevs, _ = capture.ListAudioSources()
+
+	names := []string{"Default output"}
+	for _, d := range g.audioDevs {
+		label := d.Desc
+		if label == "" {
+			label = d.Name
+		}
+		if d.IsOutput {
+			label = "Output: " + label
+		} else {
+			label = "Input: " + label
+		}
+		names = append(names, label)
+	}
+
+	g.audioSel = widget.NewSelect(names, func(string) { g.applyLiveSettings("") })
+
+	// Preselect the configured serial if it matches an enumerated device.
+	if g.cfg.AudioSource > 0 {
+		for i, d := range g.audioDevs {
+			if d.Serial == g.cfg.AudioSource {
+				g.audioSel.SetSelectedIndex(i + 1)
+				break
+			}
+		}
+	} else {
+		g.audioSel.SetSelectedIndex(0)
+	}
+}
+
+// selectedAudioSerial returns the PipeWire serial chosen in the picker, or 0
+// for the default output.
+func (g *App) selectedAudioSerial() uint64 {
+	if g.audioSel == nil || g.audioSel.SelectedIndex() <= 0 {
+		return 0
+	}
+	idx := g.audioSel.SelectedIndex() - 1
+	if idx >= 0 && idx < len(g.audioDevs) {
+		return g.audioDevs[idx].Serial
+	}
+	return 0
+}
+
 // applyPreset applies a named quality/fps preset to the Quality and Frame
 // rate selectors, then pushes it live if the engine is running.
 func (g *App) applyPreset(string) {
@@ -320,6 +376,7 @@ func (g *App) applyLiveSettings(string) {
 		FPS:         fps,
 		Source:      g.cfg.Source,
 		Audio:       g.audioChk.Checked,
+		AudioSource: g.selectedAudioSerial(),
 		StreamIndex: g.cfg.StreamIndex,
 		Scale:       parseScale(g.scaleSel.Selected),
 		Announce:    announce,
@@ -332,6 +389,7 @@ func (g *App) applyLiveSettings(string) {
 	g.cfg.Quality = quality
 	g.cfg.FPS = fps
 	g.cfg.Audio = cfg.Audio
+	g.cfg.AudioSource = cfg.AudioSource
 	g.cfg.Name = cfg.Name
 	g.cfg.Scale = cfg.Scale
 	ann := announce
@@ -343,31 +401,37 @@ func (g *App) toggleStream() {
 	if g.lock {
 		return
 	}
-	if g.eng != nil {
+	// Engine exists and is currently running → pause it.
+	if g.eng != nil && g.eng.Status().Running {
 		g.stop()
 		return
 	}
 	g.start()
 }
 
-// start reads the form into cfg, saves it, and boots the engine.
+// start reads the form into cfg, saves it, and boots the engine. If an engine
+// already exists and the capture-affecting settings are unchanged, it resumes
+// the paused engine instead of reopening the portal session (which cannot be
+// reliably re-created in-process).
 func (g *App) start() {
 	port, _ := atoi(g.portEnt.Text)
 	quality, _ := atoi(g.qualitySel.Selected)
 	fps, _ := atoi(g.fpsSel.Selected)
 
 	announce := g.announceChk.Checked
-	g.cfg = config.Config{
+	newCfg := config.Config{
 		Name:        g.nameEnt.Text,
 		Port:        port,
 		Quality:     quality,
 		FPS:         fps,
 		Source:      g.srcSel.Selected,
 		Audio:       g.audioChk.Checked,
+		AudioSource: g.selectedAudioSerial(),
 		StreamIndex: g.selectedMonitorIndex(),
 		Scale:       parseScale(g.scaleSel.Selected),
 		Announce:    &announce,
 	}
+	g.cfg = newCfg
 	if g.configPath != "" {
 		if err := config.SaveTo(g.configPath, g.cfg); err != nil {
 			log.Printf("gui: config save: %v", err)
@@ -376,17 +440,47 @@ func (g *App) start() {
 		log.Printf("gui: config save: %v", err)
 	}
 
+	// Resume the existing engine if capture-affecting fields are unchanged.
+	if g.eng != nil && captureConfigEqual(g.lastStart, newCfg) {
+		// Push live-applicable changes, then resume.
+		g.applyLiveSettings("")
+		g.eng.Resume()
+		g.startedUI(newCfg)
+		return
+	}
+
+	// Otherwise close any existing engine and create a fresh one.
+	if g.eng != nil {
+		g.eng.Close()
+		g.eng = nil
+	}
+
 	eng, err := flinger.New(g.cfg.ToFlinger())
 	if err != nil {
 		dialog.ShowError(err, g.win)
 		return
 	}
 	g.eng = eng
+	g.lastStart = newCfg
 	g.eng.Start()
 
+	g.startedUI(newCfg)
+}
+
+// captureConfigEqual reports whether two configs agree on the fields that
+// require reopening the capture session (source, port, monitor, audio
+// source). Live-applicable fields (quality, fps, scale, name, announce,
+// audio-on) are ignored.
+func captureConfigEqual(a, b config.Config) bool {
+	return a.Source == b.Source && a.Port == b.Port &&
+		a.StreamIndex == b.StreamIndex && a.AudioSource == b.AudioSource
+}
+
+// startedUI updates the UI to the streaming state after a start/resume.
+func (g *App) startedUI(cfg config.Config) {
 	g.startBtn.SetText("Stop")
 	g.startBtn.Importance = widget.DangerImportance
-	g.statusLab.SetText("Streaming (port " + itoa(g.cfg.Port) + ")")
+	g.statusLab.SetText("Streaming (port " + itoa(cfg.Port) + ")")
 	g.statusLab.Importance = widget.SuccessImportance
 	g.setTrayState(true, "TeleportFling · Streaming")
 	g.refresh()
@@ -460,7 +554,8 @@ func formatBitrate(bps int64) string {
 	}
 }
 
-// stop halts the engine and returns the UI to the stopped state.
+// stop pauses the engine (keeping the portal session open) and returns the
+// UI to the stopped state.
 func (g *App) stop() {
 	if g.statsDone != nil {
 		close(g.statsDone)
@@ -468,7 +563,6 @@ func (g *App) stop() {
 	}
 	if g.eng != nil {
 		g.eng.Stop()
-		g.eng = nil
 	}
 	g.startBtn.SetText("Start")
 	g.startBtn.Importance = widget.HighImportance