@flowingspring/dsh-voco 0.2.9 → 0.3.1

package/README.md

@@ -37,6 +37,7 @@ The plugin uses `XingChenAGI/XingChenASR-V3.2-Ultra` for cloud speech recognitio
 - Speaks a short contextual acknowledgement before starting delegated Agent work.
 - Keeps full task reports in the task UI while speaking a concise, purpose-written result.
 - Supports interruption, browser navigation, reconnects, and restored conversation history.
+- Pure voice sessions get a concise title from the first meaningful voice request, with the transcription as a fallback when the title model is unavailable.
 - Ships the server and browser surfaces as one public npm package.
 
 ## Configuration

package/README.zh.md

@@ -37,6 +37,7 @@ SILICONFLOW_API_KEY=你的密钥
 - 委派任务启动前立即播报一句贴合当前请求的简短确认语。
 - 完整任务报告保留在任务界面，语音只播报专门生成的简洁结果。
 - 支持语音打断、页面切换、断线重连以及历史对话恢复。
+- 纯语音会话会根据第一条有效语音请求生成简短标题；标题模型不可用时使用语音转写作为备用。
 - 服务端和浏览器界面统一通过一个公开 npm 包发行。
 
 ## 配置项

package/cordis.patch.yml

@@ -24,7 +24,8 @@
       name: '@flowingspring/dsh-voco/voice-assistant'
       config:
         maxPendingObservations: 64
-        restoreConversation: true
+        restoreConversation: true
+        memoryRecallTimeoutMs: 250
         maxRestoredUtterances: 24
         taskSessionPolicy: continuous
         completedAnnouncement: 任务已完成。

package/lib/client.js

@@ -18,11 +18,11 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var VoiceControl_module_css_default = {
+			"interruptButton": "TFZsua_interruptButton",
+			"active": "TFZsua_active",
 			"controls": "TFZsua_controls",
-			"voice-pulse": "TFZsua_voice-pulse",
 			"button": "TFZsua_button",
-			"active": "TFZsua_active",
-			"interruptButton": "TFZsua_interruptButton"
+			"voice-pulse": "TFZsua_voice-pulse"
 		};
 		//#endregion
 		//#region ../ui-voice/src/client/VoiceControl.tsx
@@ -141,23 +141,23 @@ window.__ModuleLoader__.load({
 			document.head.appendChild(tag);
 		}
 		var VoiceNodeViews_module_css_default = {
-			"taskInput": "Kt2smW_taskInput",
+			"taskDot": "Kt2smW_taskDot",
 			"miniWave": "Kt2smW_miniWave",
+			"taskInput": "Kt2smW_taskInput",
+			"taskUpdate": "Kt2smW_taskUpdate",
+			"taskCard": "Kt2smW_taskCard",
+			"taskTitle": "Kt2smW_taskTitle",
 			"taskSummary": "Kt2smW_taskSummary",
-			"meta": "Kt2smW_meta",
 			"bubble": "Kt2smW_bubble",
-			"taskStatus": "Kt2smW_taskStatus",
-			"taskActions": "Kt2smW_taskActions",
-			"taskLink": "Kt2smW_taskLink",
+			"meta": "Kt2smW_meta",
+			"taskChevron": "Kt2smW_taskChevron",
 			"taskDetails": "Kt2smW_taskDetails",
-			"taskDot": "Kt2smW_taskDot",
+			"taskLink": "Kt2smW_taskLink",
 			"utterance": "Kt2smW_utterance",
-			"voiceBadge": "Kt2smW_voiceBadge",
-			"taskCard": "Kt2smW_taskCard",
 			"taskCancel": "Kt2smW_taskCancel",
-			"taskChevron": "Kt2smW_taskChevron",
-			"taskUpdate": "Kt2smW_taskUpdate",
-			"taskTitle": "Kt2smW_taskTitle"
+			"taskActions": "Kt2smW_taskActions",
+			"voiceBadge": "Kt2smW_voiceBadge",
+			"taskStatus": "Kt2smW_taskStatus"
 		};
 		//#endregion
 		//#region ../ui-voice/src/client/VoiceNodeViews.tsx
@@ -286,9 +286,9 @@ window.__ModuleLoader__.load({
 		}
 		var VoiceOverlay_module_css_default = {
 			"wave": "GPPB2G_wave",
+			"stop": "GPPB2G_stop",
 			"root": "GPPB2G_root",
 			"voice-wave": "GPPB2G_voice-wave",
-			"stop": "GPPB2G_stop",
 			"status": "GPPB2G_status"
 		};
 		//#endregion

package/lib/plugins/voice-assistant.js

@@ -129,7 +129,8 @@ const Config = z.object({
 	completedAnnouncement: z.string().default("任务已完成。"),
 	failedAnnouncement: z.string().default("任务失败了，请查看屏幕上的错误信息。"),
 	cancelledAnnouncement: z.string().default("任务已取消。"),
-	interruptedAnnouncement: z.string().default("上次任务因服务关闭而中断，没有自动重放。你可以告诉我是否继续。")
+	interruptedAnnouncement: z.string().default("上次任务因服务关闭而中断，没有自动重放。你可以告诉我是否继续。"),
+	memoryRecallTimeoutMs: z.natural().min(1).default(250)
 });
 /** Plugin-owned durable session event types, registered with core at load. */
 const VOICE_SESSION_EVENT_TYPES = [
@@ -159,6 +160,41 @@ function optionalWorkspaceMemory(ctx) {
 		return;
 	}
 }
+function optionalSessionTitle(ctx) {
+	try {
+		return ctx.get("sessionTitle");
+	} catch {
+		return;
+	}
+}
+function recallWithSoftTimeout(operation, timeoutMs) {
+	const deadline = Math.max(1, timeoutMs);
+	return new Promise((resolve) => {
+		let settled = false;
+		const timer = setTimeout(() => {
+			if (settled) return;
+			settled = true;
+			resolve({ kind: "timeout" });
+		}, deadline);
+		operation().then((value) => {
+			if (settled) return;
+			settled = true;
+			clearTimeout(timer);
+			resolve({
+				kind: "resolved",
+				value
+			});
+		}, (error) => {
+			if (settled) return;
+			settled = true;
+			clearTimeout(timer);
+			resolve({
+				kind: "error",
+				error
+			});
+		});
+	});
+}
 function workspaceMemoryReference(memory) {
 	const sections = [];
 	if (memory.summary.trim() !== "") sections.push(`稳定摘要：\n${memory.summary.trim()}`);
@@ -193,13 +229,64 @@ const REWRITE_EVENT_INSTRUCTIONS = {
 	warning: "这是警告信息。简洁说明风险、影响以及用户需要知道的事项。",
 	error: "这是失败信息。明确说明没有完成、主要原因以及可行的下一步。"
 };
+const VOICE_TITLE_SYSTEM_PROMPT = `你是语音会话标题生成器。根据用户第一条有实际内容的语音请求，生成一个简短、准确、便于在会话列表中识别的中文标题。
+
+只输出标题本身，不要解释，不要引号，不要 Markdown，不要句末标点，不要使用“语音会话”“新会话”等泛化名称，不超过 12 个汉字。不要加入文件大小、行数、路径等执行细节。`;
+function voiceTitleCandidate(value) {
+	const normalized = value.replace(/```[\s\S]*?```/gu, "").replace(/^\s*(?:标题|title)\s*[:：]\s*/iu, "").replace(/["'「」『』]/gu, "").split(/\r?\n/u, 1)[0].replace(/[。！？!?；;，,：:]+$/u, "").replace(/\s+/gu, " ").trim();
+	return Array.from(normalized).slice(0, 12).join("");
+}
+function isMeaningfulVoiceTitleInput(value) {
+	const normalized = value.replace(/\s+/gu, "").trim();
+	if (Array.from(normalized).length < 4) return false;
+	return !/^(?:你好|您好|嗨|哈喽|喂|在吗|有人吗|测试|谢谢|好的)[，。！？!?、]*$/u.test(normalized);
+}
+async function generateVoiceSessionTitle(ctx, session, input, titleService) {
+	let candidate = "";
+	try {
+		const llm = ctx.get("llm");
+		if (llm !== void 0) {
+			const selection = ctx.agentDefaultModel.currentSelection();
+			const message = createUserMessage({
+				content: [{
+					type: "text",
+					text: `用户第一条有效语音请求：${input}`
+				}],
+				source: {
+					kind: "plugin",
+					plugin: "voice-assistant"
+				}
+			});
+			let output = "";
+			for await (const chunk of llm.stream({
+				provider: selection.provider,
+				model: selection.model,
+				...selection.reasoningEffort === void 0 ? {} : { reasoningEffort: selection.reasoningEffort },
+				messages: [message],
+				system: VOICE_TITLE_SYSTEM_PROMPT
+			})) if (chunk.type === "text-delta") output += chunk.text;
+			candidate = voiceTitleCandidate(output);
+		}
+	} catch (error) {
+		ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+	}
+	if (candidate === "") candidate = voiceTitleCandidate(input);
+	if (candidate === "" || titleService.get(session) !== void 0) return;
+	try {
+		titleService.rename(session, candidate);
+	} catch (error) {
+		ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+	}
+}
 /** Install the driver. @param ctx - composed Agent and voice context. @param config - driver copy and queue bounds. */
 function apply(ctx, config = {}) {
 	const bindings = /* @__PURE__ */ new Map();
 	const taskBindings = /* @__PURE__ */ new Map();
 	const handles = /* @__PURE__ */ new Map();
+	const audioResponsesSeen = /* @__PURE__ */ new Set();
 	const maxPending = config.maxPendingObservations ?? 64;
 	const maxRestoredUtterances = config.maxRestoredUtterances ?? 24;
+	const memoryRecallTimeoutMs = config.memoryRecallTimeoutMs ?? 250;
 	const loadConversationMemory = async (sessionId) => {
 		const events = ctx.sessions.get(sessionId)?.events ?? (await ctx.get("sessionPersistence")?.inspect(sessionId))?.events;
 		if (events === void 0) return void 0;
@@ -223,6 +310,7 @@ function apply(ctx, config = {}) {
 				interactionMode: void 0,
 				voiceAttached: false,
 				voiceTurnMarked: false,
+				voiceTitleRequested: false,
 				active: void 0,
 				continuousTaskAgent: void 0,
 				lastTerminalTaskId: void 0,
@@ -716,13 +804,22 @@ function apply(ctx, config = {}) {
 	};
 	const endUtterance = (binding, utteranceId, role, state, finalText, responseId) => {
 		const utterance = beginUtterance(binding, utteranceId, role, responseId);
-		requireSourceSession(binding).append("voice/utterance-end", {
+		const sourceSession = requireSourceSession(binding);
+		sourceSession.append("voice/utterance-end", {
 			utteranceId,
 			role,
 			text: finalText ?? utterance.text,
 			state,
 			...responseId === void 0 ? {} : { responseId }
 		});
+		const utteranceText = finalText ?? utterance.text;
+		if (role === "user" && state === "completed" && isMeaningfulVoiceTitleInput(utteranceText)) {
+			const titleService = optionalSessionTitle(ctx);
+			if (titleService !== void 0 && !binding.voiceTitleRequested && titleService.get(sourceSession) === void 0) {
+				binding.voiceTitleRequested = true;
+				generateVoiceSessionTitle(ctx, sourceSession, utteranceText, titleService);
+			}
+		}
 		binding.utterances.delete(utteranceId);
 	};
 	const interruptAssistantUtterances = (binding, responseId) => {
@@ -799,26 +896,69 @@ function apply(ctx, config = {}) {
 					});
 					return;
 				}
+				const routeStartedAt = Date.now();
+				const input = call.command.input;
+				debugVoiceLatency("route-command-received", {
+					callId: String(call.id),
+					inputLength: input.length
+				});
 				let route;
 				try {
 					let memoryReference = "";
 					const memory = optionalWorkspaceMemory(ctx);
-					if (memory !== void 0) try {
-						memoryReference = workspaceMemoryReference(await memory.recall({
+					if (memory !== void 0) {
+						const memoryStartedAt = Date.now();
+						debugVoiceLatency("memory-recall-start", { callId: String(call.id) });
+						const recallResult = await recallWithSoftTimeout(() => memory.recall({
 							sessionId: binding.sessionId,
-							query: call.command.input,
+							query: input,
 							maxBytes: 5e3
-						}));
-					} catch (error) {
-						ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
+						}), memoryRecallTimeoutMs);
+						if (recallResult.kind === "resolved") {
+							memoryReference = workspaceMemoryReference(recallResult.value);
+							debugVoiceLatency("memory-recall-end", {
+								callId: String(call.id),
+								durationMs: Date.now() - memoryStartedAt,
+								referenceLength: memoryReference.length
+							});
+						} else if (recallResult.kind === "timeout") debugVoiceLatency("memory-recall-timeout", {
+							callId: String(call.id),
+							durationMs: Date.now() - memoryStartedAt,
+							timeoutMs: memoryRecallTimeoutMs,
+							fallback: "empty-reference"
+						});
+						else {
+							debugVoiceLatency("memory-recall-error", {
+								callId: String(call.id),
+								durationMs: Date.now() - memoryStartedAt,
+								error: String(recallResult.error)
+							});
+							ctx.logger.warn(recallResult.error instanceof Error ? recallResult.error : new Error(String(recallResult.error)));
+						}
 					}
-					route = await routeFrontendInput(ctx, requireSourceSession(binding).events, call.command.input, memoryReference);
+					route = await routeFrontendInput(ctx, requireSourceSession(binding).events, input, memoryReference);
+					debugVoiceLatency("route-decision", {
+						callId: String(call.id),
+						action: route.action,
+						durationMs: Date.now() - routeStartedAt
+					});
 				} catch (error) {
+					debugVoiceLatency("route-error", {
+						callId: String(call.id),
+						durationMs: Date.now() - routeStartedAt,
+						error: String(error)
+					});
 					ctx.logger.warn(error instanceof Error ? error : new Error(String(error)));
-					route = fallbackDelegation(call.command.input, recentConversationText(requireSourceSession(binding).events));
+					route = fallbackDelegation(input, recentConversationText(requireSourceSession(binding).events));
 				}
 				if (route.action === "delegate") {
 					const taskId = VoiceTaskId(randomUUID());
+					debugVoiceLatency("ack-queued", {
+						callId: String(call.id),
+						taskId,
+						routeDurationMs: Date.now() - routeStartedAt,
+						textLength: route.acknowledgement.length
+					});
 					speakFragment(binding, taskId, route.acknowledgement);
 					await onTaskCommand(binding, voiceSessionId, {
 						id: call.id,
@@ -849,6 +989,12 @@ function apply(ctx, config = {}) {
 				const taskId = delegationOverride?.taskId ?? VoiceTaskId(randomUUID());
 				const requestText = delegationOverride?.requestText ?? call.command.input;
 				const continuous = (config.taskSessionPolicy ?? "isolated") === "continuous";
+				const delegationStartedAt = Date.now();
+				debugVoiceLatency("delegation-init-start", {
+					callId: String(call.id),
+					taskId,
+					policy: continuous ? "continuous" : "isolated"
+				});
 				let created;
 				try {
 					if (continuous) created = await ensureContinuousTaskAgent(binding);
@@ -860,7 +1006,19 @@ function apply(ctx, config = {}) {
 							...await createTaskAgent(binding, taskSessionId, selection)
 						};
 					}
+					debugVoiceLatency("delegation-init-end", {
+						callId: String(call.id),
+						taskId,
+						taskSessionId: created.taskSessionId,
+						durationMs: Date.now() - delegationStartedAt
+					});
 				} catch (error) {
+					debugVoiceLatency("delegation-init-error", {
+						callId: String(call.id),
+						taskId,
+						durationMs: Date.now() - delegationStartedAt,
+						error: String(error)
+					});
 					backendUnavailable(error);
 					return;
 				}
@@ -1087,6 +1245,11 @@ function apply(ctx, config = {}) {
 				});
 				return;
 			case "transcription.completed":
+				debugVoiceLatency("transcription-completed", {
+					voiceSessionId: String(session.id),
+					utteranceId: String(event.utteranceId),
+					textLength: event.text.length
+				});
 				enqueue(binding, async () => {
 					endUtterance(binding, event.utteranceId, "user", "completed", event.text);
 					if (session.interactionMode === "speech-shell" && binding.voiceSessionId === session.id) await onTranscription(binding, event.text);
@@ -1125,7 +1288,17 @@ function apply(ctx, config = {}) {
 				debugVoiceLatency("audio-started");
 				return;
 			case "output_audio.delta":
+				if (!audioResponsesSeen.has(String(event.responseId))) {
+					audioResponsesSeen.add(String(event.responseId));
+					debugVoiceLatency("audio-first-delta", {
+						responseId: String(event.responseId),
+						audioBytes: event.audio.byteLength
+					});
+				}
+				return;
 			case "output_audio.done":
+				audioResponsesSeen.delete(String(event.responseId));
+				return;
 			case "task.observation":
 			case "error":
 			case "closed": return;
@@ -1313,6 +1486,16 @@ async function routeFrontendInput(ctx, events, input, workspaceMemory = "") {
 	const recentConversation = recentConversationText(events);
 	if (llm === void 0) return fallbackDelegation(input, recentConversation);
 	const selection = ctx.agentDefaultModel.currentSelection();
+	const routeLlmStartedAt = Date.now();
+	let routeLlmFirstTextAt;
+	debugVoiceLatency("route-llm-start", {
+		provider: selection.provider,
+		model: selection.model,
+		reasoningEffort: selection.reasoningEffort,
+		inputLength: input.length,
+		recentConversationLength: recentConversation.length,
+		workspaceMemoryLength: workspaceMemory.length
+	});
 	const message = createUserMessage({
 		content: [{
 			type: "text",
@@ -1354,13 +1537,28 @@ async function routeFrontendInput(ctx, events, input, workspaceMemory = "") {
 		}
 	});
 	let output = "";
-	for await (const chunk of llm.stream({
-		provider: selection.provider,
-		model: selection.model,
-		...selection.reasoningEffort === void 0 ? {} : { reasoningEffort: selection.reasoningEffort },
-		messages: [message],
-		system: "你是语音前台路由器。严格按要求输出一个 JSON 对象。不要把普通对话委派给后台编码 Agent。"
-	})) if (chunk.type === "text-delta") output += chunk.text;
+	try {
+		for await (const chunk of llm.stream({
+			provider: selection.provider,
+			model: selection.model,
+			...selection.reasoningEffort === void 0 ? {} : { reasoningEffort: selection.reasoningEffort },
+			messages: [message],
+			system: "你是语音前台路由器。严格按要求输出一个 JSON 对象。不要把普通对话委派给后台编码 Agent。"
+		})) {
+			if (chunk.type !== "text-delta") continue;
+			if (routeLlmFirstTextAt === void 0) {
+				routeLlmFirstTextAt = Date.now();
+				debugVoiceLatency("route-llm-first-text", { delayMs: routeLlmFirstTextAt - routeLlmStartedAt });
+			}
+			output += chunk.text;
+		}
+	} finally {
+		debugVoiceLatency("route-llm-end", {
+			durationMs: Date.now() - routeLlmStartedAt,
+			firstTextDelayMs: routeLlmFirstTextAt === void 0 ? void 0 : routeLlmFirstTextAt - routeLlmStartedAt,
+			outputLength: output.length
+		});
+	}
 	const normalized = output.trim().replace(/^```(?:json)?\s*/iu, "").replace(/\s*```$/u, "");
 	const parsed = JSON.parse(normalized);
 	if (parsed.action === "chat" && typeof parsed.reply === "string" && parsed.reply.trim() !== "") return {

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@flowingspring/dsh-voco",
   "description": "Persistent voice conversations for DSH with cloud speech recognition, Edge TTS, and background Agent delegation",
-  "version": "0.2.9",
+  "version": "0.3.1",
   "keywords": [
     "deepseek-harness",
     "dsh",