dsh-audiogen 0.4.4 → 0.4.6

package/lib/index.js

@@ -24,6 +24,8 @@ const SETTINGS_API = {
 const GENERATE_API = "/api/dsh-audiogen/generate";
 /** Loopback-only task cancellation route (aborts the host-side upstream call). */
 const TASK_API = { cancel: "/api/dsh-audiogen/task/cancel" };
+/** Loopback-only prompt enhancement route (uses the agent's default model). */
+const ENHANCE_API = "/api/dsh-audiogen/prompt/enhance";
 /** Host-mediated built-in provider catalog (channels the user can instantiate). */
 const PRESETS_API = "/api/dsh-audiogen/presets";
 /** Host-mediated model/voice discovery endpoint. */
@@ -864,6 +866,66 @@ async function generateAudio(channel, request, signal) {
 	return genericAudio(channel, request, signal);
 }
 //#endregion
+//#region src/prompt-enhance.ts
+/** 按生成模式给出增强指令（系统提示）。 */
+function instructionsFor(mode) {
+	return [
+		"你是一个音频提示词增强助手。用户给出一个粗略的音频生成需求，",
+		"请将其扩写为一段可直接提交给音频生成模型的中文或英文描述。",
+		"只输出增强后的描述本身，不要输出任何解释、前后缀、引号或代码块。",
+		"保持用户原始意图，不要改变其核心内容；为最终生成的音频服务。",
+		"描述控制在 200-600 字左右。"
+	].join("") + {
+		tts: "这是文本转语音（TTS）任务：让文本更适合朗读——口语化、自然、带合适的情感标签（如 (laughs)、(whisper)），避免生僻多音字和超长句，可适当补足上下文使语句完整，但不要改写原意。",
+		music: "这是音乐生成任务：扩写音乐风格/情绪/乐器/结构/节奏变化/氛围，使用音频模型熟悉的描述词汇（如 cinematic orchestral、lo-fi、bpm、弦乐进出、旋律动机、前中后段结构），如用户未指定可补充风格建议，但保持原方向。",
+		sfx: "这是音效生成任务：扩写声音材质、动作过程、空间感、节奏（先轻后重、清脆短促等）、环境氛围，用具体拟声与材质词，避免抽象概括。",
+		voice_design: "这是音色设计任务：扩写人声/音色特征——性别年龄、音域、音质（低沉/清亮/沙哑）、语速、情绪性格、适用场景，用可感知的描述，方便语音模型合成。"
+	}[mode];
+}
+/** 读取 Agent 默认模型并调用 LLM 增强，返回增强后的文本。 */
+async function enhancePromptText(deps, prompt, mode) {
+	const text = prompt.trim();
+	if (text === "") throw new AudioGenError("提示词为空，无法增强", "enhance-empty-prompt");
+	const value = (deps.settings.describe({ redactSecrets: true }) ?? []).find((candidate) => String(candidate.ns) === "agent-default-model")?.value ?? {};
+	const provider = typeof value.provider === "string" && value.provider.trim() !== "" ? value.provider.trim() : "";
+	const model = typeof value.model === "string" && value.model.trim() !== "" ? value.model.trim() : "";
+	if (provider === "" || model === "") throw new AudioGenError("未找到 Agent 默认模型（agent-default-model）：请先在「设置 → 模型」中配置默认模型", "no-default-model");
+	const runtime = deps.llm?.();
+	if (runtime === void 0 || runtime.stream === void 0) throw new AudioGenError("宿主 LLM 服务不可用（ctx.llm 未注册）", "llm-unavailable");
+	const controller = new AbortController();
+	const timer = setTimeout(() => controller.abort(new DOMException("The operation timed out.", "TimeoutError")), 3e4);
+	timer.unref?.();
+	let output = "";
+	try {
+		for await (const chunk of runtime.stream({
+			provider,
+			model,
+			messages: [{
+				role: "user",
+				content: text
+			}],
+			system: instructionsFor(mode),
+			temperature: .7,
+			maxTokens: 1200,
+			signal: controller.signal
+		})) {
+			const record = chunk;
+			if (record.type === "text-delta" && typeof record.text === "string") output += record.text;
+			else if (record.type === "block-end" && record.block !== void 0 && record.block.type === "text" && typeof record.block.text === "string") output += record.block.text;
+		}
+	} finally {
+		clearTimeout(timer);
+	}
+	const result = stripFences(output.trim());
+	if (result === "") throw new AudioGenError("模型未返回增强内容：请检查「设置 → 模型」的默认模型是否可用（或稍后重试）", "enhance-empty-result");
+	return result;
+}
+/** 去掉模型可能包裹的 ``` 代码围栏。 */
+function stripFences(value) {
+	if (value === "") return value;
+	return value.replace(/^```[a-zA-Z]*\s*\n?/, "").replace(/\n?```\s*$/, "").trim();
+}
+//#endregion
 //#region src/audio-presets.ts
 const AUDIO_PRESETS = [
 	{
@@ -2046,6 +2108,36 @@ function makeRoutes(deps) {
 				});
 			}
 		},
+		{
+			kind: "exact",
+			path: ENHANCE_API,
+			handler: async (req, res) => {
+				if (!guard(req, res, "POST")) return;
+				const body = await readJsonBody(req);
+				const prompt = typeof body?.prompt === "string" ? body.prompt.trim() : "";
+				if (prompt === "") {
+					writeJson(res, 200, {
+						ok: false,
+						code: "bad-request",
+						message: "prompt is required"
+					});
+					return;
+				}
+				const mode = body?.mode === "music" ? "music" : body?.mode === "sfx" ? "sfx" : body?.mode === "voice_design" ? "voice_design" : "tts";
+				try {
+					writeJson(res, 200, {
+						ok: true,
+						enhanced: await deps.enhance(prompt, mode)
+					});
+				} catch (error) {
+					writeJson(res, 200, {
+						ok: false,
+						code: "enhance-failed",
+						message: messageOf(error)
+					});
+				}
+			}
+		},
 		{
 			kind: "prefix",
 			path: AUDIO_API.file,
@@ -2483,6 +2575,10 @@ function registerAgentAudioTools(ctx, resolve) {
 				type: "string",
 				description: "Optional preview text for voice_design."
 			},
+			enhance_prompt: {
+				type: "boolean",
+				description: "Enhance the prompt with the agent default model before generating (uses the configured model settings, no extra key). Best-effort: on failure the original prompt is used."
+			},
 			speed: {
 				type: "number",
 				description: "Optional speaking rate / speed multiplier where supported. MiniMax range 0.5-2.0 (default 1)."
@@ -2725,6 +2821,9 @@ function registerAgentAudioTools(ctx, resolve) {
 			/** 单模型执行：生成 + 保存文件 + 历史 + 可选资源库；错误收敛为分组结果。 */
 			const runOne = async (picked) => {
 				const request = buildRequest(picked);
+				if (args.enhance_prompt === true && config.enhance !== void 0) try {
+					request.prompt = await config.enhance(request.prompt, request.mode);
+				} catch {}
 				try {
 					const release = await (config.budget?.acquire(exec.signal) ?? Promise.resolve(() => {}));
 					let outputs;
@@ -3137,6 +3236,14 @@ function apply(ctx, config) {
 		};
 	};
 	const budget = createGenerationBudget(() => resolve().maxConcurrentGenerations);
+	const enhance = async (prompt, mode) => {
+		const seam = ctx.get("settings");
+		if (seam?.describe === void 0) throw new AudioGenError("设置服务不可用，无法增强提示词", "settings-unavailable");
+		return enhancePromptText({
+			settings: seam,
+			llm: () => ctx.get("llm")
+		}, prompt, mode);
+	};
 	const channelsView = () => {
 		const value = resolve();
 		return {
@@ -3151,7 +3258,8 @@ function apply(ctx, config) {
 				settings: seam,
 				resolveChannels: channelsView,
 				autoSave: () => resolve().autoSaveToLibrary,
-				budget
+				budget,
+				enhance
 			}).map((route) => ctx.webServer.register(route));
 			return () => {
 				for (const dispose of disposers) dispose();
@@ -3167,7 +3275,8 @@ function apply(ctx, config) {
 				channels: value.channels,
 				defaultChannelId: value.defaultChannelId,
 				autoSaveToLibrary: value.autoSaveToLibrary,
-				budget
+				budget,
+				enhance
 			};
 		}), "dsh-audiogen: agent audio tools");
 	});

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "dsh-audiogen",
   "description": "AI audio generation plugin for the dsh web GUI: multi-vendor TTS/music/sound-effect channels (OpenAI-compatible, ElevenLabs, MiniMax, Stability AI and custom), per-channel model/voice catalogs, Agent tool and a sidebar AI 音频 panel.",
-  "version": "0.4.4",
+  "version": "0.4.6",
   "type": "module",
   "main": "lib/index.js",
   "exports": {

package/src/agent-audio-tools.ts

@@ -21,6 +21,8 @@ export interface AgentAudioToolConfig {
   autoSaveToLibrary: boolean
   /** 全局并发闸门（与面板路由共享「最大并发生成数」）。 */
   budget?: GenerationBudget
+  /** 提示词增强（复用 Agent 默认模型）；enhance_prompt=true 时在生成前调用。 */
+  enhance?: (prompt: string, mode: AudioMode) => Promise<string>
 }
 
 interface AgentAudioRef {
@@ -157,6 +159,7 @@ export function registerAgentAudioTools(ctx: Context, resolve: () => AgentAudioT
       },
       voice: { type: 'string', description: 'Optional voice id/name for TTS providers. Required for MiniMax TTS (e.g. male-qn-qingse, female-shaonv); fetch the account voices in Settings > Plugins > AI Audio.' },
       preview_text: { type: 'string', description: 'Optional preview text for voice_design.' },
+      enhance_prompt: { type: 'boolean', description: 'Enhance the prompt with the agent default model before generating (uses the configured model settings, no extra key). Best-effort: on failure the original prompt is used.' },
       speed: { type: 'number', description: 'Optional speaking rate / speed multiplier where supported. MiniMax range 0.5-2.0 (default 1).' },
       duration: { type: 'number', description: 'Requested duration in seconds for music/sfx.' },
       lyrics: { type: 'string', description: 'Lyrics for music generation (MiniMax music-3.0/music-cover). Required unless is_instrumental is true. Split verses with an empty line.' },
@@ -302,6 +305,14 @@ export function registerAgentAudioTools(ctx: Context, resolve: () => AgentAudioT
       /** 单模型执行：生成 + 保存文件 + 历史 + 可选资源库；错误收敛为分组结果。 */
       const runOne = async (picked: { channel: AudioChannel; alias: string; upstream: string }): Promise<AgentAudioGroup> => {
         const request = buildRequest(picked)
+        // 可选：生成前用 Agent 默认模型增强 prompt（失败则沿用原文）
+        if (args.enhance_prompt === true && config.enhance !== undefined) {
+          try {
+            request.prompt = await config.enhance(request.prompt, request.mode)
+          } catch {
+            // 增强失败不阻断生成
+          }
+        }
         try {
           // 与面板路由共享全局并发闸门（限流时排队；取消时立即出队）。
           const release = await (config.budget?.acquire(exec.signal) ?? Promise.resolve(() => { /* 默认不限制 */ }))

package/src/client/api.ts

@@ -4,7 +4,7 @@
  */
 
 import {
-  GENERATE_API, HISTORY_API, LIBRARY_API, TASK_API,
+  ENHANCE_API, GENERATE_API, HISTORY_API, LIBRARY_API, TASK_API,
   type GenerateAudioRequest, type GeneratedAudio, type HistoryEntry,
   type LibraryEntry, type LibrarySaveRequest, type LibraryUpdateRequest,
 } from '../protocol.ts'
@@ -42,6 +42,13 @@ export class AudiogenApi {
     await postJson(TASK_API.cancel, { taskId }).catch(() => { /* best-effort */ })
   }
 
+  /** 提示词增强（复用 Agent 默认模型）。 */
+  async enhancePrompt(prompt: string, mode: string): Promise<{ ok: boolean; enhanced?: string; code?: string; message?: string }> {
+    const response = await postJson(ENHANCE_API, { prompt, mode })
+    const body = await response.json() as { ok?: boolean; enhanced?: string; code?: string; message?: string }
+    return { ok: body.ok === true, ...(body.enhanced === undefined ? {} : { enhanced: body.enhanced }), ...(body.code === undefined ? {} : { code: body.code }), ...(body.message === undefined ? {} : { message: body.message }) }
+  }
+
   async history(): Promise<HistoryEntry[]> {
     const response = await postJson(HISTORY_API.list, {})
     const body = await response.json() as { ok?: boolean; history?: HistoryEntry[] }

package/src/client/audio-panel.module.css

@@ -97,9 +97,9 @@
   flex-direction: column;
   gap: 11px;
   flex: none;
-  width: 300px;
-  min-width: 260px;
-  max-width: 340px;
+  width: 380px;
+  min-width: 320px;
+  max-width: 440px;
   min-height: 0;
   overflow-y: auto;
   padding: 2px;
@@ -111,8 +111,8 @@
 .formCol::-webkit-scrollbar-thumb { background: var(--dsw-alias-border-l2); border-radius: 999px; }
 
 .modeRow {
-  display: grid;
-  grid-template-columns: repeat(4, minmax(0, 1fr));
+  display: flex;
+  flex-wrap: wrap;
   gap: 6px;
 }
 
@@ -120,8 +120,10 @@
   display: flex;
   align-items: center;
   justify-content: center;
+  flex: 1 1 auto;
+  min-width: 0;
   min-height: 34px;
-  padding: 6px 4px;
+  padding: 6px 10px;
   border: 1px solid var(--dsw-alias-border-l2, #d1d5db);
   border-radius: 9px;
   background: var(--dsw-alias-bg-layer-1, #fff);
@@ -258,6 +260,7 @@
 .resultCol {
   flex: 1;
   min-width: 0;
+  max-width: 760px;
   display: flex;
   flex-direction: column;
   gap: 10px;
@@ -960,7 +963,7 @@
 
 .resultGroupError {
   font-size: 12px;
-  color: #dc2626;
+  color: var(--dsw-alias-state-error-primary, #dc2626);
 }
 
 .resultGroupCount {
@@ -1021,7 +1024,7 @@
 }
 
 .taskCard[data-state='failed'] {
-  border-color: #dc2626;
+  border-color: var(--dsw-alias-state-error-primary, #dc2626);
 }
 
 .taskCard[data-state='cancelled'] {
@@ -1051,11 +1054,11 @@
 }
 
 .taskStatus[data-state='done'] {
-  color: #16a34a;
+  color: var(--dsw-alias-state-success-primary, #16a34a);
 }
 
 .taskStatus[data-state='failed'] {
-  color: #dc2626;
+  color: var(--dsw-alias-state-error-primary, #dc2626);
 }
 
 .taskActions {
@@ -1107,11 +1110,11 @@
 .historyCompareBadge {
   font-size: 11px;
   font-weight: 600;
-  color: #7c3aed;
-  border: 1px solid #ddd6fe;
+  color: var(--dsw-alias-label-secondary, #6b7280);
+  border: 1px solid var(--dsw-alias-border-l2, #d1d5db);
   border-radius: 999px;
   padding: 1px 8px;
-  background: #f5f3ff;
+  background: var(--dsw-alias-bg-layer-3, #fff);
   white-space: nowrap;
 }
 
@@ -1221,7 +1224,7 @@
 }
 
 .historyIcon[title^='删除']:hover {
-  color: #dc2626;
+  color: var(--dsw-alias-state-error-primary, #dc2626);
 }
 
 /* 历史 prompt 两行截断 */
@@ -1231,3 +1234,90 @@
   -webkit-box-orient: vertical;
   overflow: hidden;
 }
+
+/* ------------------------------------------------------- P3 打磨：响应式与统一质感 */
+
+/* 焦点态统一（键盘可达性） */
+.input:focus-visible,
+.select:focus-visible,
+.textarea:focus-visible,
+.modeButton:focus-visible,
+.ghostButton:focus-visible,
+.historyAction:focus-visible,
+.historyIcon:focus-visible {
+  outline: 2px solid var(--dsw-alias-interactive-bg-hover-accent, rgba(38, 49, 72, 0.35));
+  outline-offset: 1px;
+}
+
+.generate:focus-visible {
+  outline: 2px solid var(--dsw-alias-interactive-bg-hover-accent, rgba(38, 49, 72, 0.35));
+  outline-offset: 2px;
+}
+
+/* 中窄屏：历史移到下方整行 */
+@media (max-width: 1100px) {
+  .studio {
+    flex-wrap: wrap;
+  }
+
+  .historyCol {
+    width: 100%;
+    min-width: 0;
+    max-width: none;
+  }
+}
+
+/* 窄屏：单列堆叠 */
+@media (max-width: 720px) {
+  .formCol {
+    width: 100%;
+    min-width: 0;
+    max-width: none;
+  }
+
+  .resultCol {
+    min-height: 320px;
+  }
+}
+
+/* 提示词增强 */
+.formSectionRow {
+  display: flex;
+  align-items: flex-end;
+  justify-content: space-between;
+  gap: 8px;
+}
+
+.formSectionRow .formSection {
+  flex: 1;
+}
+
+.enhanceCard {
+  border: 1px solid var(--dsw-alias-border-l1, #e5e7eb);
+  border-radius: 10px;
+  padding: 10px;
+  background: var(--dsw-alias-bg-layer-2, #fafafa);
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+
+.enhanceCardHead {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  gap: 8px;
+  font-size: 12px;
+  color: var(--dsw-alias-label-secondary, #6b7280);
+}
+
+.enhanceActions {
+  display: flex;
+  gap: 6px;
+}
+
+.enhanceActionsRow {
+  display: flex;
+  justify-content: flex-end;
+  margin-top: -4px;
+}

package/src/client/studio-view.tsx

@@ -230,6 +230,9 @@ export function StudioView(props: {
   const [bitrate, setBitrate] = useState('')
   const [audioChannel, setAudioChannel] = useState('')
   const [subtitle, setSubtitle] = useState(false)
+  // 提示词增强
+  const [enhancing, setEnhancing] = useState(false)
+  const [enhancePreview, setEnhancePreview] = useState<string | null>(null)
   // Stable Audio 参数（仅 Stability 渠道显示）
   const [seed, setSeed] = useState('')
   const [steps, setSteps] = useState('')
@@ -524,6 +527,8 @@ export function StudioView(props: {
     }
     const bool = (key: string): boolean | undefined => typeof params[key] === 'boolean' ? params[key] as boolean : undefined
     setMode(modeValue)
+    const promptValue = str('prompt')
+    if (promptValue !== undefined) setPrompt(promptValue)
     const modelValue = str('model') ?? singleModel
     if (modelValue !== '') setModel(modelValue)
     if (compareModelsRestore !== undefined && compareModelsRestore.length > 0) {
@@ -578,6 +583,28 @@ export function StudioView(props: {
     props.showToast('已恢复该次生成的配置，可直接再次生成')
   }
 
+  /** 调用宿主增强（Agent 默认模型），结果先预览再应用。 */
+  const runEnhance = async (): Promise<void> => {
+    if (prompt.trim() === '') {
+      setError('请先输入文本/提示词，再点击增强')
+      return
+    }
+    setEnhancing(true)
+    setError(null)
+    try {
+      const result = await api.enhancePrompt(prompt.trim(), mode)
+      if (result.ok !== true || result.enhanced === undefined || result.enhanced.trim() === '') {
+        setError(result.message ?? '增强失败，请稍后重试')
+        return
+      }
+      setEnhancePreview(result.enhanced.trim())
+    } catch (err) {
+      setError(err instanceof Error ? err.message : String(err))
+    } finally {
+      setEnhancing(false)
+    }
+  }
+
   /** 删除历史记录（对比任务卡删除该任务的全部模型条目）。 */
   const deleteHistoryEntries = async (ids: string[]): Promise<void> => {
     try {
@@ -875,6 +902,24 @@ export function StudioView(props: {
           <span>{mode === 'voice_design' ? '音色描述' : mode === 'tts' ? '文本' : '提示词'}</span>
           <textarea className={css.textarea} value={prompt} onChange={event => setPrompt(event.target.value)} placeholder={tt('prompt.placeholder')} />
         </label>
+        <div className={css.enhanceActionsRow}>
+          <button type="button" className={css.ghostButton} disabled={enhancing} onClick={() => void runEnhance()}>
+            {enhancing ? '增强中…' : '✨ 增强提示词'}
+          </button>
+        </div>
+        {enhancePreview !== null ? (
+          <div className={css.enhanceCard}>
+            <div className={css.enhanceCardHead}>
+              <strong>增强结果（{modeLabelOf(mode)}）</strong>
+              <span className={css.enhanceActions}>
+                <button type="button" className={css.ghostButton} onClick={() => { setPrompt(enhancePreview); setEnhancePreview(null); props.showToast('已应用增强结果') }}>应用</button>
+                <button type="button" className={css.ghostButton} disabled={enhancing} onClick={() => void runEnhance()}>重新生成</button>
+                <button type="button" className={css.ghostButton} onClick={() => setEnhancePreview(null)}>放弃</button>
+              </span>
+            </div>
+            <textarea className={css.textarea} value={enhancePreview} readOnly />
+          </div>
+        ) : null}
 
         {mode === 'voice_design' ? (
           <>

package/src/index.ts

@@ -16,8 +16,10 @@ import z from 'schemastery'
 import type {} from '@deepseek-ai/dsh-host-webserver'
 import type {} from '@deepseek-ai/dsh-system-prompt'
 import type {} from '@deepseek-ai/dsh-tools'
-import { AUDIOGEN_SETTINGS_NAMESPACE, type ChannelConfig, type ModelMapping } from './protocol.ts'
+import { AUDIOGEN_SETTINGS_NAMESPACE, type AudioMode, type ChannelConfig, type ModelMapping } from './protocol.ts'
 import { createGenerationBudget } from './audio-scheduler.ts'
+import { enhancePromptText } from './prompt-enhance.ts'
+import { AudioGenError } from './audio-engine.ts'
 import { makeRoutes, type ChannelsView, type SettingsSeam } from './routes.ts'
 import type { AudioChannel } from './audio-engine.ts'
 import { registerAgentAudioTools, type AgentAudioToolConfig } from './agent-audio-tools.ts'
@@ -196,6 +198,13 @@ export function apply(ctx: Context, config?: Config): void {
   // 全局并发闸门：所有上游调用（面板路由 + Agent 工具）共享「最大并发生成数」。
   const budget = createGenerationBudget(() => resolve().maxConcurrentGenerations)
 
+  // 提示词增强：复用 Agent 默认模型（agent-default-model 设置），面板与 Agent 工具共用。
+  const enhance = async (prompt: string, mode: AudioMode): Promise<string> => {
+    const seam = ctx.get('settings') as unknown as SettingsSeam
+    if (seam?.describe === undefined) throw new AudioGenError('设置服务不可用，无法增强提示词', 'settings-unavailable')
+    return enhancePromptText({ settings: seam, llm: () => ctx.get('llm') }, prompt, mode)
+  }
+
   const channelsView = (): ChannelsView => {
     const value = resolve()
     return { channels: value.channels, defaultChannelId: value.defaultChannelId }
@@ -209,6 +218,7 @@ export function apply(ctx: Context, config?: Config): void {
         resolveChannels: channelsView,
         autoSave: () => resolve().autoSaveToLibrary,
         budget,
+        enhance,
       })
       const disposers = routes.map(route => ctx.webServer.register(route))
       return () => { for (const dispose of disposers) dispose() }
@@ -225,6 +235,7 @@ export function apply(ctx: Context, config?: Config): void {
         defaultChannelId: value.defaultChannelId,
         autoSaveToLibrary: value.autoSaveToLibrary,
         budget,
+        enhance,
       }
     }), 'dsh-audiogen: agent audio tools')
   })

package/src/prompt-enhance.ts

@@ -0,0 +1,89 @@
+/**
+ * 提示词增强：复用 Agent 当前默认模型（设置「模型」里的 provider/model，
+ * 即 agent-default-model 命名空间），宿主端发起一次 LLM 调用把用户 prompt
+ * 扩展成更适合生成的任务描述。不新增 API key 配置。
+ *
+ * 面板「✨ 增强提示词」与 generate_audio 工具的 enhance_prompt 都走这里。
+ */
+
+import type { AudioMode } from './protocol.ts'
+import { AudioGenError } from './audio-engine.ts'
+
+/** 按生成模式给出增强指令（系统提示）。 */
+function instructionsFor(mode: AudioMode): string {
+  const common = [
+    '你是一个音频提示词增强助手。用户给出一个粗略的音频生成需求，',
+    '请将其扩写为一段可直接提交给音频生成模型的中文或英文描述。',
+    '只输出增强后的描述本身，不要输出任何解释、前后缀、引号或代码块。',
+    '保持用户原始意图，不要改变其核心内容；为最终生成的音频服务。',
+    '描述控制在 200-600 字左右。',
+  ].join('')
+  const perMode: Record<AudioMode, string> = {
+    tts: '这是文本转语音（TTS）任务：让文本更适合朗读——口语化、自然、带合适的情感标签（如 (laughs)、(whisper)），避免生僻多音字和超长句，可适当补足上下文使语句完整，但不要改写原意。',
+    music: '这是音乐生成任务：扩写音乐风格/情绪/乐器/结构/节奏变化/氛围，使用音频模型熟悉的描述词汇（如 cinematic orchestral、lo-fi、bpm、弦乐进出、旋律动机、前中后段结构），如用户未指定可补充风格建议，但保持原方向。',
+    sfx: '这是音效生成任务：扩写声音材质、动作过程、空间感、节奏（先轻后重、清脆短促等）、环境氛围，用具体拟声与材质词，避免抽象概括。',
+    voice_design: '这是音色设计任务：扩写人声/音色特征——性别年龄、音域、音质（低沉/清亮/沙哑）、语速、情绪性格、适用场景，用可感知的描述，方便语音模型合成。',
+  }
+  return common + perMode[mode]
+}
+
+export interface PromptEnhanceDeps {
+  /** DSH 设置 seam（读 agent-default-model）。 */
+  settings: { describe(options?: { redactSecrets?: boolean }): Array<{ ns: unknown; value?: unknown }> }
+  /** 宿主 LLM 运行时访问器（延迟读取，调用时才获取）。 */
+  llm?: () => unknown
+}
+
+/** 读取 Agent 默认模型并调用 LLM 增强，返回增强后的文本。 */
+export async function enhancePromptText(deps: PromptEnhanceDeps, prompt: string, mode: AudioMode): Promise<string> {
+  const text = prompt.trim()
+  if (text === '') throw new AudioGenError('提示词为空，无法增强', 'enhance-empty-prompt')
+  const descriptor = (deps.settings.describe({ redactSecrets: true }) ?? []).find(candidate => String(candidate.ns) === 'agent-default-model')
+  const value = (descriptor?.value ?? {}) as { provider?: unknown; model?: unknown }
+  const provider = typeof value.provider === 'string' && value.provider.trim() !== '' ? value.provider.trim() : ''
+  const model = typeof value.model === 'string' && value.model.trim() !== '' ? value.model.trim() : ''
+  if (provider === '' || model === '') {
+    throw new AudioGenError('未找到 Agent 默认模型（agent-default-model）：请先在「设置 → 模型」中配置默认模型', 'no-default-model')
+  }
+  const runtime = deps.llm?.() as { stream?: (options: unknown) => AsyncIterable<unknown> } | undefined
+  if (runtime === undefined || runtime.stream === undefined) {
+    throw new AudioGenError('宿主 LLM 服务不可用（ctx.llm 未注册）', 'llm-unavailable')
+  }
+  const controller = new AbortController()
+  const timer = setTimeout(() => controller.abort(new DOMException('The operation timed out.', 'TimeoutError')), 30_000)
+  timer.unref?.()
+  let output = ''
+  try {
+    for await (const chunk of runtime.stream({
+      provider,
+      model,
+      messages: [{ role: 'user', content: text }],
+      system: instructionsFor(mode),
+      temperature: 0.7,
+      maxTokens: 1200,
+      signal: controller.signal,
+    })) {
+      const record = chunk as { type?: string; text?: string; block?: { type?: string; text?: string } }
+      if (record.type === 'text-delta' && typeof record.text === 'string') {
+        output += record.text
+      } else if (record.type === 'block-end' && record.block !== undefined
+        && record.block.type === 'text' && typeof record.block.text === 'string') {
+        output += record.block.text
+      }
+    }
+  } finally {
+    clearTimeout(timer)
+  }
+  const result = stripFences(output.trim())
+  if (result === '') {
+    throw new AudioGenError('模型未返回增强内容：请检查「设置 → 模型」的默认模型是否可用（或稍后重试）', 'enhance-empty-result')
+  }
+  return result
+}
+
+/** 去掉模型可能包裹的 ``` 代码围栏。 */
+function stripFences(value: string): string {
+  if (value === '') return value
+  const withoutFence = value.replace(/^```[a-zA-Z]*\s*\n?/, '').replace(/\n?```\s*$/, '')
+  return withoutFence.trim()
+}

package/src/protocol.ts

@@ -8,7 +8,7 @@
 export const AUDIOGEN_SETTINGS_NAMESPACE = 'dsh-audiogen'
 
 /** Published package version shared by the host updater and the client UI. */
-export const PLUGIN_VERSION = '0.4.4'
+export const PLUGIN_VERSION = '0.4.6'
 
 /** Same-origin route family (loopback-only, mirroring dsh-imagegen). */
 export const SETTINGS_API = {
@@ -24,6 +24,9 @@ export const TASK_API = {
   cancel: '/api/dsh-audiogen/task/cancel',
 } as const
 
+/** Loopback-only prompt enhancement route (uses the agent's default model). */
+export const ENHANCE_API = '/api/dsh-audiogen/prompt/enhance' as const
+
 /** Host-mediated built-in provider catalog (channels the user can instantiate). */
 export const PRESETS_API = '/api/dsh-audiogen/presets' as const