@oh-my-pi/pi-coding-agent 6.8.5 → 6.9.0

package/src/core/voice-supervisor.ts

@@ -1,976 +0,0 @@
-import { logger, ptree } from "@oh-my-pi/pi-utils";
-import {
-	RealtimeAgent,
-	RealtimeSession,
-	type RealtimeSessionConfig,
-	type TransportEvent,
-	type TransportLayerAudio,
-	tool,
-} from "@openai/agents/realtime";
-import type { ReadableStreamDefaultReader as WebReadableStreamDefaultReader } from "stream/web";
-import { z } from "zod";
-import type { ModelRegistry } from "./model-registry";
-
-const DEFAULT_REALTIME_MODEL = process.env.OMP_VOICE_REALTIME_MODEL ?? "gpt-realtime";
-const DEFAULT_REALTIME_VOICE = process.env.OMP_VOICE_REALTIME_VOICE ?? "marin";
-const DEFAULT_SAMPLE_RATE = 24000;
-const DEFAULT_CHANNELS = 1;
-const DEFAULT_BITS = 16;
-const INTERRUPT_DEBOUNCE_MS = 200;
-const MAX_RESULT_CHARS = 6000;
-const MAX_PROGRESS_CHARS = 1400;
-const PLAYBACK_ACTIVE_WINDOW_MS = 350;
-// Echo cancellation: only suppress mic when playback is active and mic is much quieter
-const ECHO_SUPPRESSION_RATIO = 2.5;
-// Minimum RMS to ever send (absolute noise floor)
-const MIC_NOISE_FLOOR = 0.005;
-const PLAYBACK_ERROR_COOLDOWN_MS = 2000;
-
-const SUPERVISOR_INSTRUCTIONS = [
-	"You are the realtime voice supervisor for a terminal coding agent.",
-	"Manage conversation flow, turn-taking, and what gets spoken aloud.",
-	"For user speech: if unclear, ask exactly one short question.",
-	"If clear, call send_to_agent with a concise instruction for the coding agent.",
-	"If the user is greeting/smalltalk or gives no actionable request, respond briefly and do not call send_to_agent.",
-	"Keep spoken responses to 1-2 short sentences (<=40 words).",
-	"You will receive system updates prefixed with SYSTEM_EVENT, PROGRESS_UPDATE, or AGENT_OUTPUT.",
-	"For AGENT_OUTPUT, always respond with a brief spoken summary and any single question needed.",
-	"For PROGRESS_UPDATE, speak a short update only if it helps the user stay oriented.",
-	"Do not call send_to_agent for system updates.",
-	"If the user asks to stop or cancel work, call interrupt_agent.",
-].join(" ");
-
-type VoiceSupervisorCallbacks = {
-	onSendToAgent: (text: string) => Promise<void> | void;
-	onInterruptAgent: (reason?: string) => Promise<void> | void;
-	onStatus: (status?: string) => void;
-	onError: (error: Error) => void;
-	onWarning?: (message: string) => void;
-};
-
-function normalizeText(text: string): string {
-	return text.replace(/\s+/g, " ").trim();
-}
-
-function truncateText(text: string, maxChars: number): string {
-	if (text.length <= maxChars) return text;
-	return `${text.slice(0, maxChars)}...`;
-}
-
-function toArrayBuffer(chunk: Uint8Array): ArrayBuffer {
-	const buffer = chunk.buffer;
-	if (buffer instanceof ArrayBuffer) {
-		if (chunk.byteOffset === 0 && chunk.byteLength === buffer.byteLength) {
-			return buffer;
-		}
-		return buffer.slice(chunk.byteOffset, chunk.byteOffset + chunk.byteLength);
-	}
-	const copy = new Uint8Array(chunk.byteLength);
-	copy.set(chunk);
-	return copy.buffer;
-}
-
-function describeError(error: unknown): string {
-	if (error instanceof Error) return error.message;
-	if (typeof error === "string") return error;
-	if (error && typeof error === "object") {
-		const maybeMessage = (error as { message?: unknown }).message;
-		if (typeof maybeMessage === "string") return maybeMessage;
-		const nested = (error as { error?: unknown }).error;
-		if (nested) return describeError(nested);
-		try {
-			return JSON.stringify(error);
-		} catch {
-			return String(error);
-		}
-	}
-	return String(error);
-}
-
-type AudioToolStatus = {
-	capture: { available: boolean; tool?: string; command?: string[] };
-	playback: { available: boolean; tool?: string; command?: string[] };
-};
-
-function checkAudioTools(sampleRate: number, channels: number): AudioToolStatus {
-	const captureResult = buildCaptureCommand(sampleRate, channels);
-	const playbackCmd = buildPlaybackCommand(sampleRate, channels);
-
-	return {
-		capture: {
-			available: captureResult !== null,
-			tool: captureResult?.command[0],
-			command: captureResult?.command,
-		},
-		playback: {
-			available: playbackCmd !== null,
-			tool: playbackCmd?.[0],
-			command: playbackCmd ?? undefined,
-		},
-	};
-}
-
-function getMissingToolsMessage(): string {
-	const platform = process.platform;
-	const lines: string[] = ["Voice mode requires audio tools. Install one of the following:"];
-
-	if (platform === "linux") {
-		lines.push("");
-		lines.push("  For capture (microphone):");
-		lines.push("    • sox (recommended): sudo dnf install sox");
-		lines.push("    • pulseaudio-utils: sudo dnf install pulseaudio-utils");
-		lines.push("    • alsa-utils: sudo dnf install alsa-utils");
-		lines.push("    • ffmpeg: sudo dnf install ffmpeg");
-		lines.push("");
-		lines.push("  For playback (speaker):");
-		lines.push("    • sox (recommended): sudo dnf install sox");
-		lines.push("    • ffmpeg: sudo dnf install ffmpeg");
-		lines.push("");
-		lines.push("  Set OMP_VOICE_CAPTURE_DEVICE to override the default capture device.");
-		lines.push("  (Applies to all tools; for sox, this sets AUDIODEV internally.)");
-	} else if (platform === "darwin") {
-		lines.push("");
-		lines.push("  • sox (recommended): brew install sox");
-		lines.push("  • ffmpeg: brew install ffmpeg");
-	} else if (platform === "win32") {
-		lines.push("");
-		lines.push("  • sox: choco install sox");
-		lines.push("  • ffmpeg: choco install ffmpeg");
-	}
-
-	return lines.join("\n");
-}
-
-type CaptureCommand = { command: string[]; env?: Record<string, string> };
-
-function buildCaptureCommand(sampleRate: number, channels: number): CaptureCommand | null {
-	const platform = process.platform;
-	// Allow user to override capture device via environment
-	const captureDevice = process.env.OMP_VOICE_CAPTURE_DEVICE;
-
-	// Prefer sox/rec as they work well across platforms
-	const soxPath = Bun.which("sox") ?? Bun.which("rec");
-	if (soxPath) {
-		const command = [
-			soxPath,
-			"-q",
-			"-d",
-			"-t",
-			"raw",
-			"-r",
-			String(sampleRate),
-			"-e",
-			"signed-integer",
-			"-b",
-			String(DEFAULT_BITS),
-			"-c",
-			String(channels),
-			"-",
-		];
-		// sox uses AUDIODEV env var to override the default device
-		const env = captureDevice ? { AUDIODEV: captureDevice } : undefined;
-		return { command, env };
-	}
-
-	// On Linux, try PulseAudio first (parecord)
-	if (platform === "linux") {
-		const parecordPath = Bun.which("parecord");
-		if (parecordPath) {
-			const command = [parecordPath, "--raw", "--format=s16le", `--rate=${sampleRate}`, `--channels=${channels}`];
-			if (captureDevice) {
-				command.push(`--device=${captureDevice}`);
-			}
-			return { command };
-		}
-	}
-
-	// ALSA arecord as fallback on Linux
-	const arecordPath = Bun.which("arecord");
-	if (arecordPath) {
-		const device = captureDevice ?? "default";
-		return {
-			command: [
-				arecordPath,
-				"-q",
-				"-D",
-				device,
-				"-f",
-				"S16_LE",
-				"-r",
-				String(sampleRate),
-				"-c",
-				String(channels),
-				"-t",
-				"raw",
-			],
-		};
-	}
-
-	// ffmpeg fallback with platform-specific input
-	const ffmpegPath = Bun.which("ffmpeg");
-	if (ffmpegPath) {
-		if (platform === "darwin") {
-			const device = captureDevice ?? ":0";
-			return {
-				command: [
-					ffmpegPath,
-					"-hide_banner",
-					"-loglevel",
-					"error",
-					"-f",
-					"avfoundation",
-					"-i",
-					device,
-					"-ac",
-					String(channels),
-					"-ar",
-					String(sampleRate),
-					"-f",
-					"s16le",
-					"-",
-				],
-			};
-		}
-		if (platform === "linux") {
-			// Try PulseAudio format first, fall back to ALSA
-			const hasPulse = Bun.which("pulseaudio") || Bun.which("pipewire-pulse") || process.env.PULSE_SERVER;
-			const format = hasPulse ? "pulse" : "alsa";
-			const device = captureDevice ?? "default";
-			return {
-				command: [
-					ffmpegPath,
-					"-hide_banner",
-					"-loglevel",
-					"error",
-					"-f",
-					format,
-					"-i",
-					device,
-					"-ac",
-					String(channels),
-					"-ar",
-					String(sampleRate),
-					"-f",
-					"s16le",
-					"-",
-				],
-			};
-		}
-		if (platform === "win32") {
-			const device = captureDevice ?? "audio=default";
-			return {
-				command: [
-					ffmpegPath,
-					"-hide_banner",
-					"-loglevel",
-					"error",
-					"-f",
-					"dshow",
-					"-i",
-					device,
-					"-ac",
-					String(channels),
-					"-ar",
-					String(sampleRate),
-					"-f",
-					"s16le",
-					"-",
-				],
-			};
-		}
-	}
-
-	return null;
-}
-
-function buildPlaybackCommand(sampleRate: number, channels: number): string[] | null {
-	const preferred = process.env.OMP_VOICE_PLAYBACK?.toLowerCase();
-	const ffplayPath = Bun.which("ffplay");
-	const playPath = Bun.which("play");
-	const soxPath = Bun.which("sox");
-
-	const playCommand = playPath
-		? [
-				playPath,
-				"-q",
-				"-t",
-				"raw",
-				"-r",
-				String(sampleRate),
-				"-e",
-				"signed-integer",
-				"-b",
-				String(DEFAULT_BITS),
-				"-c",
-				String(channels),
-				"-",
-			]
-		: null;
-
-	const soxCommand = soxPath
-		? [
-				soxPath,
-				"-q",
-				"-t",
-				"raw",
-				"-r",
-				String(sampleRate),
-				"-e",
-				"signed-integer",
-				"-b",
-				String(DEFAULT_BITS),
-				"-c",
-				String(channels),
-				"-",
-				"-d",
-			]
-		: null;
-
-	const ffplayCommand = ffplayPath
-		? [
-				ffplayPath,
-				"-nodisp",
-				"-autoexit",
-				"-hide_banner",
-				"-loglevel",
-				"error",
-				"-fflags",
-				"nobuffer",
-				"-flags",
-				"low_delay",
-				"-f",
-				"s16le",
-				"-ar",
-				String(sampleRate),
-				"-ac",
-				String(channels),
-				"-",
-			]
-		: null;
-
-	if (preferred === "ffplay") return ffplayCommand;
-	if (preferred === "play") return playCommand ?? soxCommand;
-	if (preferred === "sox") return soxCommand ?? playCommand;
-
-	return playCommand ?? soxCommand ?? ffplayCommand;
-}
-
-function rms16le(buffer: Uint8Array): number {
-	if (buffer.byteLength < 2) return 0;
-	const view = new DataView(buffer.buffer, buffer.byteOffset, buffer.byteLength);
-	let sum = 0;
-	let count = 0;
-	for (let i = 0; i + 1 < buffer.byteLength; i += 2) {
-		const sample = view.getInt16(i, true) / 32768;
-		sum += sample * sample;
-		count += 1;
-	}
-	if (count === 0) return 0;
-	return Math.sqrt(sum / count);
-}
-
-export class VoiceSupervisor {
-	private session: RealtimeSession | undefined = undefined;
-	private captureProcess: ptree.ChildProcess | undefined = undefined;
-	private captureReader: WebReadableStreamDefaultReader<Uint8Array> | undefined = undefined;
-	private playbackProcess: ptree.ChildProcess | undefined = undefined;
-	private playbackWriter:
-		| {
-				write: (chunk: Uint8Array) => Promise<void>;
-				close: () => Promise<void>;
-		  }
-		| undefined = undefined;
-	private active = false;
-	private connected = false;
-	private sessionReady = false;
-	private lastInterruptAt = 0;
-	private lastPlaybackAt = 0;
-	private lastPlaybackRms = 0;
-	private lastPlaybackErrorAt = 0;
-	// Fallback transcript handling: track user speech when no tool call is made
-	private pendingTranscript = "";
-	private pendingResponseHasToolCall = false;
-	private pendingResponseHasAudioOutput = false;
-
-	constructor(
-		private registry: ModelRegistry,
-		private callbacks: VoiceSupervisorCallbacks,
-	) {}
-
-	/**
-	 * Check if audio tools are available for voice mode.
-	 * Returns null if all tools are available, or an error message if not.
-	 */
-	static checkAvailability(): { available: boolean; error?: string; tools?: AudioToolStatus } {
-		const status = checkAudioTools(DEFAULT_SAMPLE_RATE, DEFAULT_CHANNELS);
-		if (status.capture.available && status.playback.available) {
-			return { available: true, tools: status };
-		}
-
-		const missing: string[] = [];
-		if (!status.capture.available) missing.push("capture");
-		if (!status.playback.available) missing.push("playback");
-
-		return {
-			available: false,
-			error: `Missing audio ${missing.join(" and ")} tools.\n\n${getMissingToolsMessage()}`,
-			tools: status,
-		};
-	}
-
-	get isActive(): boolean {
-		return this.active;
-	}
-
-	async start(): Promise<void> {
-		if (this.active) return;
-
-		const apiKey = await this.registry.getApiKeyForProvider("openai");
-		if (!apiKey) {
-			throw new Error("OpenAI API key not found (set OPENAI_API_KEY or login).");
-		}
-
-		this.active = true;
-		this.lastInterruptAt = 0;
-		this.sessionReady = false;
-		this.lastPlaybackErrorAt = 0;
-		this.pendingTranscript = "";
-		this.pendingResponseHasToolCall = false;
-		this.pendingResponseHasAudioOutput = false;
-		this.callbacks.onStatus("Connecting realtime voice...");
-
-		try {
-			const agent = this.createSupervisorAgent();
-			const session = new RealtimeSession(agent, {
-				transport: "websocket",
-				model: DEFAULT_REALTIME_MODEL,
-				config: this.buildSessionConfig(),
-			});
-
-			this.session = session;
-			this.bindSessionEvents(session);
-			await session.connect({ apiKey });
-			this.connected = session.transport.status === "connected";
-			this.sessionReady = this.connected;
-			if (!this.connected) {
-				await this.waitForConnection(session, 5000);
-			}
-			await this.waitForSessionReady(session, 5000);
-			await this.startCapture();
-			await this.ensurePlayback();
-			this.callbacks.onStatus("Listening... (auto-send on silence, Ctrl+Y to stop)");
-		} catch (error) {
-			await this.stop();
-			throw new Error(describeError(error));
-		}
-	}
-
-	async stop(): Promise<void> {
-		if (!this.active) return;
-		this.active = false;
-		this.connected = false;
-		this.sessionReady = false;
-		await this.stopCapture();
-		await this.resetPlayback();
-		if (this.session) {
-			this.session.close();
-			this.session = undefined;
-		}
-		this.callbacks.onStatus(undefined);
-	}
-
-	notifyProgress(text: string): void {
-		this.sendSystemMessage("PROGRESS_UPDATE", text, MAX_PROGRESS_CHARS);
-	}
-
-	notifyResult(text: string): void {
-		this.sendSystemMessage("AGENT_OUTPUT", text, MAX_RESULT_CHARS);
-	}
-
-	private sendSystemMessage(prefix: string, text: string, maxChars: number): void {
-		if (!this.session || !this.active) return;
-		if (!this.connected || !this.sessionReady || this.session.transport.status !== "connected") return;
-		const trimmed = normalizeText(text);
-		if (!trimmed) return;
-		const payload = `${prefix}: ${truncateText(trimmed, maxChars)}`;
-		try {
-			this.session.transport.sendEvent({
-				type: "conversation.item.create",
-				item: {
-					type: "message",
-					role: "system",
-					content: [{ type: "input_text", text: payload }],
-				},
-			});
-			this.session.transport.sendEvent({ type: "response.create" });
-		} catch (error) {
-			const message = describeError(error);
-			if (message.includes("WebSocket is not connected")) return;
-			this.callbacks.onError(error instanceof Error ? error : new Error(message));
-		}
-	}
-
-	private createSupervisorAgent(): RealtimeAgent {
-		const sendToAgentTool = tool({
-			name: "send_to_agent",
-			description: "Send a concise instruction to the coding agent.",
-			parameters: z.object({
-				text: z.string().min(1),
-			}),
-			execute: async ({ text }) => {
-				const cleaned = normalizeText(text);
-				if (cleaned) {
-					await this.callbacks.onSendToAgent(cleaned);
-				}
-				return "sent";
-			},
-		});
-
-		const interruptAgentTool = tool({
-			name: "interrupt_agent",
-			description: "Interrupt the coding agent immediately.",
-			parameters: z.object({
-				reason: z.string().optional(),
-			}),
-			execute: async ({ reason }) => {
-				await this.callbacks.onInterruptAgent(reason);
-				return "interrupted";
-			},
-		});
-
-		return new RealtimeAgent({
-			name: "Voice Supervisor",
-			instructions: SUPERVISOR_INSTRUCTIONS,
-			tools: [sendToAgentTool, interruptAgentTool],
-			voice: DEFAULT_REALTIME_VOICE,
-		});
-	}
-
-	private buildSessionConfig(): Partial<RealtimeSessionConfig> {
-		return {
-			outputModalities: ["audio"],
-			audio: {
-				input: {
-					format: { type: "audio/pcm", rate: DEFAULT_SAMPLE_RATE },
-					noiseReduction: { type: "near_field" },
-					turnDetection: {
-						type: "semantic_vad",
-						createResponse: true,
-						interruptResponse: true,
-					},
-				},
-				output: {
-					format: { type: "audio/pcm", rate: DEFAULT_SAMPLE_RATE },
-					...(DEFAULT_REALTIME_VOICE ? { voice: DEFAULT_REALTIME_VOICE } : {}),
-				},
-			},
-		};
-	}
-
-	private bindSessionEvents(session: RealtimeSession): void {
-		session.transport.on("connection_change", (status) => {
-			this.connected = status === "connected";
-			if (this.connected) {
-				this.sessionReady = true;
-			} else {
-				this.sessionReady = false;
-			}
-			if (!this.active) return;
-			if (this.connected) {
-				this.callbacks.onStatus("Listening... (auto-send on silence, Ctrl+Y to stop)");
-			} else {
-				this.callbacks.onStatus("Reconnecting realtime voice...");
-			}
-		});
-
-		session.on("audio", (event: TransportLayerAudio) => {
-			void this.handleAudio(event);
-		});
-
-		session.on("audio_start", () => {
-			if (!this.active) return;
-			this.pendingResponseHasAudioOutput = true;
-			this.callbacks.onStatus("Speaking...");
-		});
-
-		session.on("audio_stopped", () => {
-			if (!this.active) return;
-			this.callbacks.onStatus("Listening... (auto-send on silence, Ctrl+Y to stop)");
-		});
-
-		session.on("audio_interrupted", () => {
-			void this.resetPlayback();
-			if (!this.active) return;
-			this.callbacks.onStatus("Listening... (auto-send on silence, Ctrl+Y to stop)");
-		});
-
-		session.on("transport_event", (event: TransportEvent) => {
-			this.handleTransportEvent(event);
-		});
-
-		session.on("error", (error) => {
-			const message = describeError(error);
-			logger.debug("voice-supervisor: realtime error", { error: message });
-			if (message.includes("WebSocket is not connected")) {
-				if (this.active) {
-					this.callbacks.onStatus("Reconnecting realtime voice...");
-				}
-				return;
-			}
-			this.callbacks.onError(new Error(message));
-		});
-	}
-
-	private handleTransportEvent(event: TransportEvent): void {
-		if (!this.active) return;
-
-		// Session ready
-		if (event.type === "session.created") {
-			this.sessionReady = true;
-			return;
-		}
-
-		// User speech started - interrupt agent and reset tracking
-		if (event.type === "input_audio_buffer.speech_started") {
-			const now = Date.now();
-			if (now - this.lastInterruptAt < INTERRUPT_DEBOUNCE_MS) return;
-			this.lastInterruptAt = now;
-			this.pendingTranscript = "";
-			this.pendingResponseHasToolCall = false;
-			this.pendingResponseHasAudioOutput = false;
-			void this.callbacks.onInterruptAgent();
-			return;
-		}
-
-		// User speech transcript completed - store for fallback
-		if (event.type === "conversation.item.input_audio_transcription.completed") {
-			const transcript = (event as { transcript?: string }).transcript;
-			if (transcript && typeof transcript === "string") {
-				this.pendingTranscript = normalizeText(transcript);
-				logger.debug("voice-supervisor: transcript captured", { transcript: this.pendingTranscript });
-			}
-			return;
-		}
-
-		// Response started - begin tracking
-		if (event.type === "response.created") {
-			this.pendingResponseHasToolCall = false;
-			this.pendingResponseHasAudioOutput = false;
-			return;
-		}
-
-		// Tool call detected - mark so we know not to use fallback
-		// Check multiple event types for robustness against API changes
-		if (
-			event.type === "function_call" ||
-			event.type === "response.function_call_arguments.done" ||
-			event.type === "response.function_call_arguments.delta" ||
-			event.type === "response.output_item.added"
-		) {
-			// For output_item.added, only mark if it's a function_call type
-			if (event.type === "response.output_item.added") {
-				const item = (event as { item?: { type?: string } }).item;
-				if (item?.type === "function_call") {
-					this.pendingResponseHasToolCall = true;
-				}
-			} else {
-				this.pendingResponseHasToolCall = true;
-			}
-			return;
-		}
-
-		// Audio output detected - mark so we don't fallback
-		if (
-			event.type === "response.output_audio.delta" ||
-			event.type === "response.output_audio.done" ||
-			event.type === "response.output_audio_transcript.delta" ||
-			event.type === "response.output_audio_transcript.done" ||
-			event.type === "response.content_part.added" ||
-			event.type === "response.content_part.done"
-		) {
-			this.pendingResponseHasAudioOutput = true;
-			return;
-		}
-
-		// Response completed - check if we need fallback
-		if (event.type === "response.done") {
-			// Only use fallback if we have a transcript AND there was no tool call AND no audio output
-			// This prevents duplicate responses when the realtime assistant already spoke
-			if (this.pendingTranscript && !this.pendingResponseHasToolCall && !this.pendingResponseHasAudioOutput) {
-				logger.debug("voice-supervisor: using fallback transcript path", {
-					transcript: this.pendingTranscript,
-				});
-				const transcript = this.pendingTranscript;
-				this.pendingTranscript = "";
-				// Queue the fallback asynchronously to avoid blocking
-				setImmediate(() => {
-					if (this.active) {
-						void this.callbacks.onSendToAgent(transcript);
-					}
-				});
-			}
-			return;
-		}
-	}
-
-	private async handleAudio(event: TransportLayerAudio): Promise<void> {
-		if (!this.active) return;
-		const now = Date.now();
-		try {
-			await this.ensurePlayback();
-		} catch (error) {
-			this.callbacks.onError(new Error(describeError(error)));
-			return;
-		}
-		if (!this.playbackWriter) return;
-		try {
-			await this.playbackWriter.write(new Uint8Array(event.data));
-			this.lastPlaybackAt = now;
-			this.lastPlaybackRms = rms16le(new Uint8Array(event.data));
-		} catch (error) {
-			logger.debug("voice-supervisor: playback write failed", {
-				error: describeError(error),
-			});
-			void this.resetPlayback();
-		}
-	}
-
-	private async startCapture(): Promise<void> {
-		const captureResult = buildCaptureCommand(DEFAULT_SAMPLE_RATE, DEFAULT_CHANNELS);
-		if (!captureResult) {
-			throw new Error(`No audio capture tool found.\n\n${getMissingToolsMessage()}`);
-		}
-
-		const { command, env: captureEnv } = captureResult;
-		logger.debug("voice-supervisor: starting mic capture", { command, env: captureEnv });
-		const proc = ptree.cspawn(command, {
-			env: captureEnv ? { ...process.env, ...captureEnv } : undefined,
-		});
-		this.captureProcess = proc;
-		const reader = proc.stdout.getReader();
-		this.captureReader = reader;
-
-		(async () => {
-			while (this.active) {
-				const { value, done } = await reader.read();
-				if (done || !this.active) break;
-				if (!value || !this.session) continue;
-				if (!this.connected || !this.sessionReady || this.session.transport.status !== "connected") {
-					continue;
-				}
-
-				const micRms = rms16le(value);
-				const now = Date.now();
-				const playbackActive = now - this.lastPlaybackAt < PLAYBACK_ACTIVE_WINDOW_MS;
-
-				// Echo suppression: only skip if playback is active AND mic is very quiet relative to playback
-				// This prevents feedback loops while allowing user to speak over the assistant
-				if (playbackActive && micRms < MIC_NOISE_FLOOR && micRms < this.lastPlaybackRms / ECHO_SUPPRESSION_RATIO) {
-					continue;
-				}
-
-				// Send all audio to realtime API - let semantic_vad handle turn detection
-				const buffer = toArrayBuffer(value);
-				if (buffer.byteLength === 0) continue;
-				try {
-					this.session.sendAudio(buffer);
-				} catch (error) {
-					const message = describeError(error);
-					logger.debug("voice-supervisor: sendAudio failed", { error: message });
-					if (message.includes("WebSocket is not connected")) {
-						continue;
-					}
-					this.callbacks.onError(error instanceof Error ? error : new Error(message));
-					return;
-				}
-			}
-			if (this.active) {
-				this.callbacks.onError(new Error("Voice capture stopped unexpectedly."));
-			}
-		})().catch((error) => {
-			if (!this.active) return;
-			logger.debug("voice-supervisor: capture loop error", {
-				error: describeError(error),
-			});
-			this.callbacks.onError(new Error(describeError(error)));
-		});
-	}
-
-	private async stopCapture(): Promise<void> {
-		if (this.captureReader) {
-			try {
-				await this.captureReader.cancel();
-			} catch {
-				// ignore
-			}
-			this.captureReader = undefined;
-		}
-		if (this.captureProcess) {
-			try {
-				this.captureProcess.kill("SIGINT");
-			} catch {
-				// ignore
-			}
-			await this.captureProcess.exited;
-			this.captureProcess = undefined;
-		}
-	}
-
-	private async ensurePlayback(): Promise<void> {
-		if (this.playbackProcess && this.playbackWriter) return;
-		const command = buildPlaybackCommand(DEFAULT_SAMPLE_RATE, DEFAULT_CHANNELS);
-		if (!command) {
-			throw new Error(`No audio playback tool found.\n\n${getMissingToolsMessage()}`);
-		}
-
-		logger.debug("voice-supervisor: starting audio playback", { command });
-		const proc = ptree.cspawn(command, {
-			stdin: "pipe",
-		});
-		const startedAt = Date.now();
-
-		this.playbackProcess = proc;
-		const stdin = proc.stdin;
-		if (!stdin) {
-			throw new Error("Audio playback stdin unavailable.");
-		}
-		if ("getWriter" in stdin && typeof stdin.getWriter === "function") {
-			const writer = (stdin as unknown as WritableStream<Uint8Array>).getWriter();
-			this.playbackWriter = {
-				write: async (chunk) => {
-					await writer.write(chunk);
-				},
-				close: async () => {
-					await writer.close();
-				},
-			};
-		} else if ("write" in stdin && typeof (stdin as { write?: unknown }).write === "function") {
-			const sink = stdin as unknown as {
-				write: (chunk: Uint8Array) => undefined | number | Promise<undefined | number>;
-				end?: () => undefined | number | Promise<undefined | number>;
-				close?: () => undefined | number | Promise<undefined | number>;
-			};
-			this.playbackWriter = {
-				write: async (chunk) => {
-					await sink.write(chunk);
-				},
-				close: async () => {
-					if (sink.end) {
-						await sink.end();
-					} else if (sink.close) {
-						await sink.close();
-					}
-				},
-			};
-		} else {
-			throw new Error("Audio playback stdin is not writable.");
-		}
-
-		proc.exited
-			.then(() => {
-				const code = proc.exitCode;
-				if (this.playbackProcess === proc) {
-					this.playbackProcess = undefined;
-					this.playbackWriter = undefined;
-				}
-				const trimmed = proc.peekStderr().trim();
-				if (trimmed) {
-					logger.debug("voice-supervisor: playback stderr", { stderr: trimmed });
-				}
-				const elapsed = Date.now() - startedAt;
-				if (code !== 0 && elapsed < 2000 && this.active && code !== null) {
-					this.maybeWarnPlaybackFailure(trimmed || `exit code ${code}`);
-				}
-			})
-			.catch(() => {
-				// ignore
-			});
-	}
-
-	private async resetPlayback(): Promise<void> {
-		if (this.playbackWriter) {
-			try {
-				await this.playbackWriter.close();
-			} catch {
-				// ignore
-			}
-		}
-		if (this.playbackProcess) {
-			try {
-				this.playbackProcess.kill();
-			} catch {
-				// ignore
-			}
-			await this.playbackProcess.exited;
-		}
-		this.playbackProcess = undefined;
-		this.playbackWriter = undefined;
-	}
-
-	private maybeWarnPlaybackFailure(message: string): void {
-		if (!this.callbacks.onWarning) return;
-		const now = Date.now();
-		if (now - this.lastPlaybackErrorAt < PLAYBACK_ERROR_COOLDOWN_MS) return;
-		this.lastPlaybackErrorAt = now;
-		this.callbacks.onWarning(`Audio playback failed: ${message}`);
-	}
-
-	private async waitForConnection(session: RealtimeSession, timeoutMs: number): Promise<void> {
-		if (session.transport.status === "connected") {
-			this.connected = true;
-			return;
-		}
-		await new Promise<void>((resolve, reject) => {
-			const timeout = setTimeout(() => {
-				cleanup();
-				reject(new Error("Realtime voice connection timed out."));
-			}, timeoutMs);
-
-			const cleanup = () => {
-				clearTimeout(timeout);
-				session.transport.off("connection_change", onChange);
-			};
-
-			const onChange = (status: string) => {
-				if (status === "connected") {
-					this.connected = true;
-					cleanup();
-					resolve();
-				}
-			};
-
-			session.transport.on("connection_change", onChange);
-		});
-	}
-
-	private async waitForSessionReady(session: RealtimeSession, timeoutMs: number): Promise<void> {
-		if (this.sessionReady) return;
-		await new Promise<void>((resolve, reject) => {
-			let resolved = false;
-
-			const cleanup = () => {
-				clearTimeout(timeout);
-				session.off("transport_event", onEvent);
-			};
-
-			const timeout = setTimeout(() => {
-				if (resolved) return;
-				resolved = true;
-				cleanup();
-				reject(new Error("Realtime voice session not ready."));
-			}, timeoutMs);
-
-			const onEvent = (event: TransportEvent) => {
-				if (resolved) return;
-				if (event.type === "session.created") {
-					this.sessionReady = true;
-					resolved = true;
-					cleanup();
-					resolve();
-				}
-			};
-
-			session.on("transport_event", onEvent);
-		});
-	}
-}