@flowingspring/dsh-voco 0.2.0 → 0.2.2

package/lib/plugins/voice-assistant.js

@@ -13,25 +13,21 @@ const VOICE_MESSAGE_SECTION_ORDER = 118;
 * Install `send_voice_message` and its guidance into one delegated task Agent scope.
 * @param agentCtx - task-Agent context receiving the tool and guidance.
 * @param send - binding-owned message receiver.
+* @param resolveDelegationId - resolves the active delegation without exposing its id to the model or transcript.
 * @returns disposer that revokes both registrations.
 */
-function installVoiceMessageTool(agentCtx, send) {
+function installVoiceMessageTool(agentCtx, send, resolveDelegationId) {
 	const disposeSection = agentCtx.systemPrompt.section({
 		name: "tool:send-voice-message",
 		order: VOICE_MESSAGE_SECTION_ORDER,
-		text: "When a user message contains a <realtime_delegation> request or <realtime_delegation_update>, use send_voice_message with its delegation_id to keep the realtime voice assistant informed. Report structured events with type and detail. Use progress for meaningful progress, warning or error for important conditions, question when user input is required, and result exactly once before a successful final turn ends. Put the complete factual content in detail; the Voice layer independently rewrites every reported event against the user request. Keep detail complete enough to support an accurate rewrite. The full report remains in the task UI. The voice assistant does not automatically see your transcript, tool output, or reasoning. A result is held until the turn actually succeeds. A question leaves the task waiting for a user reply. Reporting never ends your turn. Do not use this tool for ordinary requests without a realtime_delegation envelope."
+		text: "This Agent handles tasks delegated by the realtime voice assistant. While a realtime voice delegation is active, use send_voice_message to keep the voice assistant informed. The tool is already bound to the exact active delegation; never ask for, invent, or expose an internal delegation id. Report structured events with type and detail. Use progress for meaningful progress, warning or error for important conditions, question when user input is required, and result exactly once before a successful final turn ends. Put the complete factual content in detail; the Voice layer independently rewrites every reported event against the user request. Keep detail complete enough to support an accurate rewrite. The full report remains in the task UI. The voice assistant does not automatically see your transcript, tool output, or reasoning. A result is held until the turn actually succeeds. A question leaves the task waiting for a user reply. Reporting never ends your turn. Do not use this tool when no realtime voice delegation is active."
 	});
 	let disposeTool;
 	try {
 		disposeTool = agentCtx.tools.register(defineTool({
 			name: "send_voice_message",
-			description: "Send a user-facing status or final result to the realtime voice assistant for the exact active delegation using type and complete factual detail. Progress, warnings, errors, and questions may repeat when meaningful. result may be called once. The Voice layer independently rewrites detail into conversational speech, so this tool accepts no Agent-authored speech field. COMPLETE is held until the Agent turn succeeds; it does not finish the turn. The voice assistant does not otherwise see this Agent transcript or tool output.",
+			description: "Send a user-facing status or final result to the realtime voice assistant for the exact active delegation using type and complete factual detail. The delegation is bound automatically; no delegation id is accepted. Progress, warnings, errors, and questions may repeat when meaningful. result may be called once. The Voice layer independently rewrites detail into conversational speech, so this tool accepts no Agent-authored speech field. COMPLETE is held until the Agent turn succeeds; it does not finish the turn. The voice assistant does not otherwise see this Agent transcript or tool output.",
 			parameters: {
-				delegation_id: {
-					type: "string",
-					required: true,
-					description: "Exact delegation_id from the realtime_delegation request envelope."
-				},
 				type: {
 					type: "string",
 					enum: [
@@ -85,8 +81,10 @@ function installVoiceMessageTool(agentCtx, send) {
 				const type = args.type ?? (args.channel === "COMPLETE" ? "result" : "progress");
 				const detail = args.detail.trim();
 				if (detail === "") throw new Error("send_voice_message detail must be non-empty");
+				const delegationId = resolveDelegationId();
+				if (delegationId === void 0) throw new Error("send_voice_message has no active voice delegation");
 				return Promise.resolve(send({
-					delegationId: VoiceTaskId(args.delegation_id),
+					delegationId,
 					channel: type === "result" ? "COMPLETE" : "STATUS",
 					type,
 					detail
@@ -154,6 +152,9 @@ try {
 		if (hostSession.KNOWN_SESSION_EVENT_TYPES !== void 0) registerVoiceSessionEventTypes(hostSession.KNOWN_SESSION_EVENT_TYPES);
 	}
 } catch {}
+function sameModelSelection(left, right) {
+	return left.provider === right.provider && left.model === right.model && left.reasoningEffort === right.reasoningEffort;
+}
 const REWRITE_SYSTEM_PROMPT = `你是语音模式下的自然回复编辑器。你的任务是把后台处理结果改写成准确、自然、适合直接朗读的中文回复。
 
 必须遵守：
@@ -249,7 +250,13 @@ function apply(ctx, config = {}) {
 		const voiceId = binding.voiceSessionId;
 		if (voiceId === void 0 || !binding.voiceAttached || text.trim() === "") return;
 		if (ctx.voice.appendSpeechText(voiceId, text)) {
-			if (flush) ctx.voice.requestResponse(voiceId, { kind: "automatic" });
+			if (flush) {
+				debugVoiceLatency("tts-request", {
+					taskId,
+					textLength: text.length
+				});
+				ctx.voice.requestResponse(voiceId, { kind: "automatic" });
+			}
 			return;
 		}
 		append(binding, {
@@ -279,6 +286,14 @@ function apply(ctx, config = {}) {
 		const abort = new AbortController();
 		binding.rewriteAbort = abort;
 		const selection = ctx.agentDefaultModel.currentSelection();
+		debugVoiceLatency("rewrite-start", {
+			taskId,
+			eventType,
+			provider: selection.provider,
+			model: selection.model,
+			selectedReasoningEffort: selection.reasoningEffort,
+			originalLength: original.length
+		});
 		const prompt = [
 			"根据用户原话和处理结果，生成一条可以直接显示并朗读的最终回复。",
 			REWRITE_EVENT_INSTRUCTIONS[eventType],
@@ -309,13 +324,13 @@ function apply(ctx, config = {}) {
 			for await (const chunk of llm.stream({
 				provider: selection.provider,
 				model: selection.model,
-				...selection.reasoningEffort === void 0 ? {} : { reasoningEffort: selection.reasoningEffort },
 				messages: [message],
 				system: REWRITE_SYSTEM_PROMPT,
 				signal: abort.signal
 			})) {
 				if (generation !== binding.rewriteGeneration || abort.signal.aborted) return;
 				if (chunk.type !== "text-delta") continue;
+				if (rewritten === "") debugVoiceLatency("rewrite-first-text", { taskId });
 				rewritten += chunk.text;
 				pending += chunk.text;
 				const split = speechFragments(pending, false);
@@ -329,6 +344,10 @@ function apply(ctx, config = {}) {
 			if (rewritten.trim() === "" && pending.trim() === "") speakFragment(binding, taskId, fallbackEventSpeech(eventType, original));
 		} catch (error) {
 			if (!abort.signal.aborted && generation === binding.rewriteGeneration) {
+				debugVoiceLatency("rewrite-error", {
+					taskId,
+					error: String(error)
+				});
 				ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
 				speakFragment(binding, taskId, fallbackEventSpeech(eventType, original));
 			}
@@ -429,12 +448,40 @@ function apply(ctx, config = {}) {
 		binding.agent = handle.agent;
 		return handle.agent;
 	};
-	const createTaskAgent = async (binding, taskSessionId) => {
+	const loggedSourceSelection = (binding) => {
+		const event = ctx.sessions.get(binding.sessionId)?.events.findLast((candidate) => candidate.type === "request/header");
+		if (event?.type !== "request/header") return void 0;
+		const config = event.data.header.config;
+		return {
+			selection: {
+				provider: config.provider,
+				model: config.model,
+				...config.reasoningEffort === void 0 ? {} : { reasoningEffort: config.reasoningEffort }
+			},
+			seq: event.seq
+		};
+	};
+	const taskModelSelection = (binding) => {
+		const defaults = ctx.agentDefaultModel.currentSelection();
+		const logged = loggedSourceSelection(binding);
+		const current = binding.continuousTaskAgent?.selection;
+		if (current === void 0) return {
+			selection: logged?.selection ?? defaults,
+			sourceHeaderSeq: logged?.seq
+		};
+		if (logged !== void 0 && logged.seq !== binding.continuousTaskAgent?.sourceHeaderSeq) return {
+			selection: logged.selection,
+			sourceHeaderSeq: logged.seq
+		};
+		return {
+			selection: sameModelSelection(defaults, current) ? current : defaults,
+			sourceHeaderSeq: logged?.seq
+		};
+	};
+	const createTaskAgent = async (binding, taskSessionId, selection) => {
 		const sourceAgent = await ensureAgent(binding);
 		const sourceSession = ctx.sessions.get(binding.sessionId);
 		if (sourceSession === void 0) throw new Error(`voice-assistant: source session "${binding.sessionId}" is not live`);
-		const defaults = ctx.agentDefaultModel.currentSelection();
-		const header = foldRequestHeader(sourceSession.events);
 		const presetId = resolveSessionPreset(sourceSession);
 		const presets = ctx.get("agentPresets");
 		let disposeVoiceMessage;
@@ -445,12 +492,12 @@ function apply(ctx, config = {}) {
 				...presetId === void 0 ? {} : { agentPreset: presetId }
 			},
 			agentOptions: {
-				provider: sourceAgent.options.provider ?? header?.config.provider ?? defaults.provider,
-				model: sourceAgent.options.model ?? header?.config.model ?? defaults.model
+				provider: selection.provider,
+				model: selection.model
 			},
 			setup: (agentCtx) => {
 				presets?.composeFrom(agentCtx, sourceAgent.ctx);
-				disposeVoiceMessage = installVoiceMessageTool(agentCtx, (input) => sendVoiceMessage(binding, input));
+				disposeVoiceMessage = installVoiceMessageTool(agentCtx, (input) => sendVoiceMessage(binding, input), () => binding.active?.id);
 			}
 		});
 		handles.set(taskSessionId, handle);
@@ -471,25 +518,24 @@ function apply(ctx, config = {}) {
 			disposeVoiceMessage
 		};
 	};
-	const resumeTaskAgent = async (binding, taskSessionId) => {
+	const resumeTaskAgent = async (binding, taskSessionId, selection) => {
 		const live = ctx.agents.get(taskSessionId);
 		if (live !== void 0) return {
 			agent: live,
-			disposeVoiceMessage: installVoiceMessageTool(live.ctx, (input) => sendVoiceMessage(binding, input))
+			disposeVoiceMessage: installVoiceMessageTool(live.ctx, (input) => sendVoiceMessage(binding, input), () => binding.active?.id)
 		};
 		const sourceAgent = await ensureAgent(binding);
-		const defaults = ctx.agentDefaultModel.currentSelection();
 		const presets = ctx.get("agentPresets");
 		let disposeVoiceMessage;
 		const handle = await ctx.agents.resume({
 			resumeSessionId: taskSessionId,
 			agentOptions: {
-				provider: sourceAgent.options.provider ?? defaults.provider,
-				model: sourceAgent.options.model ?? defaults.model
+				provider: selection.provider,
+				model: selection.model
 			},
 			setup: (agentCtx) => {
 				presets?.composeFrom(agentCtx, sourceAgent.ctx);
-				disposeVoiceMessage = installVoiceMessageTool(agentCtx, (input) => sendVoiceMessage(binding, input));
+				disposeVoiceMessage = installVoiceMessageTool(agentCtx, (input) => sendVoiceMessage(binding, input), () => binding.active?.id);
 			}
 		});
 		handles.set(taskSessionId, handle);
@@ -504,7 +550,40 @@ function apply(ctx, config = {}) {
 		};
 	};
 	const ensureContinuousTaskAgent = async (binding) => {
-		if (binding.continuousTaskAgent !== void 0) return binding.continuousTaskAgent;
+		const { selection, sourceHeaderSeq } = taskModelSelection(binding);
+		const current = binding.continuousTaskAgent;
+		if (current !== void 0 && sameModelSelection(current.selection, selection)) {
+			if (current.sourceHeaderSeq === sourceHeaderSeq) return current;
+			const refreshed = {
+				...current,
+				sourceHeaderSeq
+			};
+			binding.continuousTaskAgent = refreshed;
+			return refreshed;
+		}
+		if (current !== void 0) {
+			binding.continuousTaskAgent = void 0;
+			taskBindings.delete(current.taskSessionId);
+			try {
+				await current.agent.whenIdle();
+			} catch (error) {
+				ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+			}
+			try {
+				current.disposeVoiceMessage();
+			} catch (error) {
+				ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+			}
+			const handle = handles.get(current.taskSessionId);
+			if (handle !== void 0) {
+				handles.delete(current.taskSessionId);
+				try {
+					await handle.dispose();
+				} catch (error) {
+					ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+				}
+			}
+		}
 		const sourceSession = requireSourceSession(binding);
 		const previousState = sourceSession.events.findLast((event) => event.type === "voice/agent-binding-state");
 		const previousBound = sourceSession.events.findLast((event) => event.type === "voice/task-session-bound");
@@ -513,7 +592,9 @@ function apply(ctx, config = {}) {
 		if (previousTaskSessionId !== void 0) try {
 			const resource = {
 				taskSessionId: previousTaskSessionId,
-				...await resumeTaskAgent(binding, previousTaskSessionId)
+				selection,
+				sourceHeaderSeq,
+				...await resumeTaskAgent(binding, previousTaskSessionId, selection)
 			};
 			binding.continuousTaskAgent = resource;
 			taskBindings.set(resource.taskSessionId, binding);
@@ -524,7 +605,9 @@ function apply(ctx, config = {}) {
 		const taskSessionId = SessionId(`session-${randomUUID()}`);
 		const resource = {
 			taskSessionId,
-			...await createTaskAgent(binding, taskSessionId)
+			selection,
+			sourceHeaderSeq,
+			...await createTaskAgent(binding, taskSessionId, selection)
 		};
 		binding.continuousTaskAgent = resource;
 		taskBindings.set(taskSessionId, binding);
@@ -641,7 +724,7 @@ function apply(ctx, config = {}) {
 		}, false);
 		agent.steer(message);
 	};
-	const onTaskCommand = async (binding, voiceSessionId, call) => {
+	const onTaskCommand = async (binding, voiceSessionId, call, taskIdOverride) => {
 		const complete = (result) => {
 			ctx.voice.completeTaskCommand(voiceSessionId, call.id, result);
 		};
@@ -670,16 +753,21 @@ function apply(ctx, config = {}) {
 					route = await routeFrontendInput(ctx, requireSourceSession(binding).events, call.command.input);
 				} catch (error) {
 					ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
-					route = { action: "delegate" };
+					route = {
+						action: "delegate",
+						acknowledgement: DEFAULT_DELEGATION_ACKNOWLEDGEMENT
+					};
 				}
 				if (route.action === "delegate") {
+					const taskId = VoiceTaskId(randomUUID());
+					speakFragment(binding, taskId, route.acknowledgement);
 					await onTaskCommand(binding, voiceSessionId, {
 						id: call.id,
 						command: {
 							type: "realtime_delegation",
 							input: call.command.input
 						}
-					});
+					}, taskId);
 					return;
 				}
 				complete({ kind: "handled" });
@@ -696,16 +784,17 @@ function apply(ctx, config = {}) {
 					return;
 				}
 				cancelRewrite(binding);
-				const taskId = VoiceTaskId(randomUUID());
+				const taskId = taskIdOverride ?? VoiceTaskId(randomUUID());
 				const continuous = (config.taskSessionPolicy ?? "isolated") === "continuous";
 				let created;
 				try {
 					if (continuous) created = await ensureContinuousTaskAgent(binding);
 					else {
 						const taskSessionId = SessionId(`session-${randomUUID()}`);
+						const { selection } = taskModelSelection(binding);
 						created = {
 							taskSessionId,
-							...await createTaskAgent(binding, taskSessionId)
+							...await createTaskAgent(binding, taskSessionId, selection)
 						};
 					}
 				} catch (error) {
@@ -715,10 +804,21 @@ function apply(ctx, config = {}) {
 				const message = createUserMessage({
 					content: [{
 						type: "text",
-						text: renderRealtimeDelegation(taskId, call.command.input, call.command.transcriptDelta)
+						text: call.command.input
 					}],
 					source: { kind: "user" }
 				});
+				const transcriptContext = call.command.transcriptDelta?.trim();
+				const contextMessage = transcriptContext === void 0 || transcriptContext === "" ? void 0 : createUserMessage({
+					content: [{
+						type: "text",
+						text: `语音转写补充上下文：${transcriptContext}`
+					}],
+					source: {
+						kind: "plugin",
+						plugin: "voice-assistant"
+					}
+				});
 				const task = {
 					id: taskId,
 					interactionMode: "frontend-agent",
@@ -738,8 +838,12 @@ function apply(ctx, config = {}) {
 						taskSessionId: created.taskSessionId,
 						input: call.command.input
 					});
+					if (contextMessage !== void 0) created.agent.inject(contextMessage);
 					created.agent.followup(message);
 				} catch (error) {
+					if (contextMessage !== void 0) try {
+						created.agent.inbox.remove(contextMessage.id);
+					} catch {}
 					if (!continuous) taskBindings.delete(created.taskSessionId);
 					binding.active = void 0;
 					try {
@@ -780,7 +884,7 @@ function apply(ctx, config = {}) {
 				const message = createUserMessage({
 					content: [{
 						type: "text",
-						text: renderRealtimeDelegationUpdate(task.id, call.command.message)
+						text: call.command.message
 					}],
 					source: {
 						kind: "plugin",
@@ -953,6 +1057,8 @@ function apply(ctx, config = {}) {
 				if (session.interactionMode === "frontend-agent") enqueue(binding, () => binding.voiceSessionId === session.id ? onTaskCommand(binding, session.id, event.call) : void 0);
 				return;
 			case "output_audio.started":
+				debugVoiceLatency("audio-started");
+				return;
 			case "output_audio.delta":
 			case "output_audio.done":
 			case "task.observation":
@@ -1064,6 +1170,19 @@ function speechFragments(text, flush) {
 		rest
 	};
 }
+const DEFAULT_DELEGATION_ACKNOWLEDGEMENT = "好的，我先查看一下。";
+const MAX_DELEGATION_ACKNOWLEDGEMENT_LENGTH = 40;
+const VOICE_LATENCY_DEBUG_PREFIX = "[DEBUG-VOICE-LATENCY]";
+function debugVoiceLatency(...values) {
+	if (globalThis.process?.env?.DSH_VOICE_LATENCY_DEBUG !== "1") return;
+	console.info(VOICE_LATENCY_DEBUG_PREFIX, (/* @__PURE__ */ new Date()).toISOString(), ...values);
+}
+function delegationAcknowledgement(value) {
+	if (typeof value !== "string") return DEFAULT_DELEGATION_ACKNOWLEDGEMENT;
+	const normalized = value.replace(/\s+/gu, " ").trim();
+	if (normalized === "" || normalized.length > MAX_DELEGATION_ACKNOWLEDGEMENT_LENGTH) return DEFAULT_DELEGATION_ACKNOWLEDGEMENT;
+	return normalized;
+}
 async function routeFrontendInput(ctx, events, input) {
 	let llm;
 	try {
@@ -1071,7 +1190,10 @@ async function routeFrontendInput(ctx, events, input) {
 	} catch {
 		llm = void 0;
 	}
-	if (llm === void 0) return { action: "delegate" };
+	if (llm === void 0) return {
+		action: "delegate",
+		acknowledgement: DEFAULT_DELEGATION_ACKNOWLEDGEMENT
+	};
 	const selection = ctx.agentDefaultModel.currentSelection();
 	const recentConversation = events.flatMap((event) => {
 		if (event.type !== "voice/utterance-end" || event.data.state !== "completed") return [];
@@ -1085,10 +1207,12 @@ async function routeFrontendInput(ctx, events, input) {
 				"判断下面这句话应该由语音助手直接回答，还是委派给后台编码 Agent。",
 				"普通寒暄、日常对话、无需读取本地项目或调用工具即可回答的问题，选择 chat，并直接给出自然简洁的中文回复。",
 				"只有需要查看或修改工作区文件、运行命令、测试、安装依赖或执行其他工具操作时，才选择 delegate。",
+				"选择 delegate 时，同时给出一句简短自然的 acknowledgement，表示接下来要做什么。不能声称任务已经完成、已经找到结果，也不要提后台 Agent、工具或路由。",
+				"acknowledgement 只能有一句，最多 40 个字符，例如“好的，我先检查一下相关代码。”。",
 				"只输出一行 JSON，不要 Markdown。格式只能是：",
 				"{\"action\":\"chat\",\"reply\":\"...\"}",
 				"或：",
-				"{\"action\":\"delegate\"}",
+				"{\"action\":\"delegate\",\"acknowledgement\":\"...\"}",
 				"",
 				"最近对话：",
 				recentConversation || "（无）",
@@ -1115,7 +1239,10 @@ async function routeFrontendInput(ctx, events, input) {
 		action: "chat",
 		reply: parsed.reply.trim()
 	};
-	if (parsed.action === "delegate") return { action: "delegate" };
+	if (parsed.action === "delegate") return {
+		action: "delegate",
+		acknowledgement: delegationAcknowledgement(parsed.acknowledgement)
+	};
 	throw new Error("voice frontend router returned an invalid decision");
 }
 function fallbackSpeechText(text) {
@@ -1155,6 +1282,13 @@ function observeSessionEvent(binding, event, append, config, onDisposeError, rew
 	}
 	if (event.type !== "turn/end" || event.data.turn !== task.taskTurn) return false;
 	const status = terminalStatus(event.data.reason.kind);
+	const failureReason = status === "failed" ? event.data.reason.kind === "error" ? event.data.reason.error.message : event.data.reason.kind : void 0;
+	debugVoiceLatency("turn-end", {
+		taskId: task.id,
+		turn: task.taskTurn,
+		status,
+		...failureReason === void 0 ? {} : { failureReason }
+	});
 	if (status === "completed" && task.waitingUser && task.completionDetail === void 0) {
 		delete task.taskTurn;
 		delete task.lastAssistantMessage;
@@ -1180,7 +1314,7 @@ function observeSessionEvent(binding, event, append, config, onDisposeError, rew
 		},
 		...message === void 0 ? {} : { voiceMessage: message },
 		...announcement === void 0 ? {} : { announcement },
-		...status === "failed" ? { reason: event.data.reason.kind } : {}
+		...failureReason === void 0 ? {} : { reason: failureReason }
 	}, message !== void 0 || announcement !== void 0, true);
 	if (rewriteFinalResult) rewrite(task, resultText);
 	const disposeVoiceMessage = task.disposeVoiceMessage;
@@ -1216,26 +1350,6 @@ function taskRejection(binding, taskId) {
 		message: `task "${taskId}" is being cancelled`
 	};
 }
-function renderRealtimeDelegation(taskId, input, transcriptDelta) {
-	return [
-		"<realtime_delegation>",
-		`  <delegation_id>${escapeXmlText(taskId)}</delegation_id>`,
-		`  <input>${escapeXmlText(input)}</input>`,
-		...transcriptDelta === void 0 ? [] : [`  <transcript_delta>${escapeXmlText(transcriptDelta)}</transcript_delta>`],
-		"</realtime_delegation>"
-	].join("\n");
-}
-function renderRealtimeDelegationUpdate(taskId, message) {
-	return [
-		"<realtime_delegation_update>",
-		`  <delegation_id>${escapeXmlText(taskId)}</delegation_id>`,
-		`  <message>${escapeXmlText(message)}</message>`,
-		"</realtime_delegation_update>"
-	].join("\n");
-}
-function escapeXmlText(value) {
-	return value.replaceAll("&", "&amp;").replaceAll("<", "&lt;").replaceAll(">", "&gt;");
-}
 function terminalStatus(reason) {
 	if (reason === "completed") return "completed";
 	if (reason === "aborted") return "cancelled";

package/lib/plugins/voice-local.js

@@ -165,17 +165,17 @@ var LocalSession = class {
 					utteranceId,
 					text: event.text
 				});
-				if (this.interactionMode === "frontend-agent" && event.text.trim() !== "") {
-					const text = event.text.trim();
-					this.emitTaskCommand(this.activeTaskId === void 0 ? {
-						type: "route_transcription",
-						input: text
-					} : {
-						type: "send_task_message",
-						taskId: this.activeTaskId,
-						message: text
-					});
-				}
+				const text = event.text.trim();
+				if (text === "") return;
+				if (this.interactionMode === "frontend-agent") this.emitTaskCommand(this.activeTaskId === void 0 ? {
+					type: "route_transcription",
+					input: text
+				} : {
+					type: "send_task_message",
+					taskId: this.activeTaskId,
+					message: text
+				});
+				this.interruptResponse();
 				return;
 			}
 			case "transcription.failed":
@@ -460,7 +460,6 @@ var NodeSpeechBackend = class {
 		const utterance = this.active;
 		if (utterance === void 0 || utterance.confirmed) return;
 		utterance.confirmed = true;
-		this.interrupt();
 		this.emit?.({
 			type: "transcription.started",
 			utteranceId: utterance.id
@@ -557,9 +556,26 @@ const Config = z.object({
 	maxUtteranceMs: z.natural().min(1e3).default(6e4)
 });
 const PACKAGE_ROOT = dirname(fileURLToPath(import.meta.url));
-const PROJECT_ENV = resolve(PACKAGE_ROOT, "../../../.env");
+/**
+* Load development configuration from either the source workspace or the
+* bundled single-package layout. DSH-installed profiles still use the
+* process environment, so no package-local `.env` is required in production.
+*/
+function loadProjectEnv() {
+	const candidates = [
+		resolve(process.cwd(), ".env"),
+		resolve(PACKAGE_ROOT, "../../../.env"),
+		resolve(PACKAGE_ROOT, "../../../../.env")
+	];
+	const loaded = /* @__PURE__ */ new Set();
+	for (const filename of candidates) {
+		if (loaded.has(filename) || !existsSync(filename)) continue;
+		loaded.add(filename);
+		loadEnvFile(filename);
+	}
+}
 function apply(ctx, config = {}) {
-	if (existsSync(PROJECT_ENV)) loadEnvFile(PROJECT_ENV);
+	loadProjectEnv();
 	return ctx.voice.registerProvider({
 		id: "local",
 		available: () => true,

package/package.json

@@ -1,19 +1,32 @@
-{
+{
   "name": "@flowingspring/dsh-voco",
-  "description": "Voice interaction profile patch layered over dsh-web-app",
-  "version": "0.2.0",
-  "publishConfig": {
-    "access": "public"
-  },
-  "repository": {
-    "type": "git",
-    "url": "git+https://github.com/lgquan/dsh-voco.git",
-    "directory": "packages/voice-app"
+  "description": "Persistent voice conversations for DSH with cloud speech recognition, Edge TTS, and background Agent delegation",
+  "version": "0.2.2",
+  "keywords": [
+    "deepseek-harness",
+    "dsh",
+    "dsh-plugin",
+    "voice",
+    "speech-recognition",
+    "text-to-speech",
+    "agent"
+  ],
+  "homepage": "https://github.com/lgquan/dsh-voco#readme",
+  "bugs": {
+    "url": "https://github.com/lgquan/dsh-voco/issues"
   },
-  "type": "module",
-  "main": "lib/index.js",
-  "types": "lib/types/index.d.ts",
-  "exports": {
+  "publishConfig": {
+    "access": "public"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/lgquan/dsh-voco.git",
+    "directory": "packages/voice-app"
+  },
+  "type": "module",
+  "main": "lib/index.js",
+  "types": "lib/types/index.d.ts",
+  "exports": {
     ".": {
       "types": "./lib/types/index.d.ts",
       "default": "./lib/index.js"
@@ -32,17 +45,17 @@
     "./voice-web": "./lib/plugins/voice-web.js",
     "./cordis.patch.yml": "./cordis.patch.yml",
     "./package.json": "./package.json"
-  },
-  "files": [
+  },
+  "files": [
     "lib/index.js",
     "lib/invariant.js",
     "lib/client.js",
     "lib/client.js.map",
     "lib/plugins/*.js",
     "cordis.patch.yml",
-    "lib/types/**/*.d.ts"
-  ],
-  "license": "MIT",
+    "lib/types/**/*.d.ts"
+  ],
+  "license": "MIT",
   "dsh": {
     "bundle": {
       "patch": "./cordis.patch.yml"
@@ -105,15 +118,15 @@
     "@deepseek-ai/dsh-system-prompt": "^0.1.1-rc.1",
     "@deepseek-ai/dsh-tools": "^0.1.1-rc.1",
     "@deepseek-ai/dsh-workspace": "^0.1.1-rc.1",
+    "@flowingspring/dsh-client-ui-voice": "workspace:*",
+    "@flowingspring/dsh-llm-tool-call-compat": "workspace:*",
+    "@flowingspring/dsh-voice": "workspace:*",
+    "@flowingspring/dsh-voice-assistant": "workspace:*",
+    "@flowingspring/dsh-voice-local": "workspace:*",
+    "@flowingspring/dsh-voice-web": "workspace:*",
     "@types/react": "~18.3.1",
     "@types/ws": "^8.18.1",
     "react": "^18.2.0",
-    "react-dom": "^18.2.0",
-    "@flowingspring/dsh-client-ui-voice": "0.1.2",
-    "@flowingspring/dsh-voice-local": "0.1.0",
-    "@flowingspring/dsh-llm-tool-call-compat": "0.1.0",
-    "@flowingspring/dsh-voice-web": "0.1.2",
-    "@flowingspring/dsh-voice": "0.1.2",
-    "@flowingspring/dsh-voice-assistant": "0.1.2"
+    "react-dom": "^18.2.0"
   }
-}
+}