refactor: make transcription and capture context-owned per run

docs/features/tui/index.md

@@ -127,7 +127,9 @@ export OPENAI_API_KEY=your_openai_api_key
 docker agent run agent.yaml
 ```
 
-Run `/speak` to start voice input. Press <kbd>Enter</kbd> to stop and send the transcribed message, or <kbd>Escape</kbd> to stop without sending it.
+Run `/speak` to start voice input. Press <kbd>Enter</kbd> to finish transcription and send the message once the final text arrives. Press <kbd>Escape</kbd> to cancel without sending; text already received stays in the draft. Errors or a finalization timeout keep the draft instead of sending a partial message.
+
+Dictation stays with the tab where it started. Switching tabs while transcription finishes keeps the text in the original draft without sending it.
 
 ### Plans Sidebar
 

pkg/audio/capture/capture.go

@@ -0,0 +1,26 @@
+// Package capture provides microphone capture using CoreAudio on macOS.
+package capture
+
+import "errors"
+
+// Common sample rates.
+const (
+	SampleRate44100 = 44100 // CD quality, used for WAV recording
+	SampleRate24000 = 24000 // Required by OpenAI Realtime API
+)
+
+// ErrNotSupported is returned when audio capture is unavailable on the current platform.
+var ErrNotSupported = errors.New("audio capture is only supported on macOS")
+
+// AudioHandler receives mono, 16-bit signed PCM (little-endian) audio chunks.
+type AudioHandler func(data []byte)
+
+// Capturer configures independent microphone capture runs.
+type Capturer struct {
+	sampleRate int
+}
+
+// NewCapturer creates a capturer with the specified sample rate.
+func NewCapturer(sampleRate int) *Capturer {
+	return &Capturer{sampleRate: sampleRate}
+}