dsh-audiogen 0.4.5 → 0.4.7

package/lib/index.js

@@ -28,6 +28,8 @@ const TASK_API = { cancel: "/api/dsh-audiogen/task/cancel" };
 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";
+/** LLM 模型目录：提示词增强模型的候选（来自「设置 → 模型」各提供方）。 */
+const LLM_MODELS_API = "/api/dsh-audiogen/llm/models";
 /** Host-mediated model/voice discovery endpoint. */
 const MODEL_API = { discover: "/api/dsh-audiogen/models/discover" };
 /** Loopback-only audio file reader for panel/tool-result previews. */
@@ -882,32 +884,60 @@ function instructionsFor(mode) {
 		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) {
+/** 读取 Agent 默认模型并调用 LLM 增强，返回增强后的文本。
+*  @param override - 用户设置的增强模型；缺省/为空时回退到 agent-default-model。 */
+async function enhancePromptText(deps, prompt, mode, override) {
 	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 selected = override !== void 0 && override.provider.trim() !== "" && override.model.trim() !== "" ? {
+		provider: override.provider.trim(),
+		model: override.model.trim()
+	} : void 0;
+	const descriptor = selected === void 0 ? (deps.settings.describe({ redactSecrets: true }) ?? []).find((candidate) => String(candidate.ns) === "agent-default-model") : void 0;
+	const value = selected ?? descriptor?.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 = "";
-	for await (const chunk of runtime.stream({
-		provider,
-		model,
-		messages: [{
-			role: "user",
-			content: text
-		}],
-		system: instructionsFor(mode),
-		temperature: .7,
-		maxTokens: 1200
-	})) {
-		const record = chunk;
-		if (record.type === "text-delta" && typeof record.text === "string") output += record.text;
+	let terminalFailure = "";
+	try {
+		for await (const chunk of runtime.stream({
+			provider,
+			model,
+			messages: [{
+				role: "user",
+				content: [{
+					type: "text",
+					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;
+			else if (record.type === "finish" && record.reason !== void 0 && record.reason.kind !== "stop" && record.reason.kind !== void 0 && terminalFailure === "") {
+				const failure = record.reason.failure;
+				terminalFailure = typeof failure?.message === "string" && failure.message.trim() !== "" ? `${failure.message}${typeof failure.code === "string" ? `（${failure.code}）` : ""}` : `stream ${record.reason.kind}`;
+			}
+		}
+	} finally {
+		clearTimeout(timer);
+	}
+	const result = stripFences(output.trim());
+	if (result === "") {
+		if (terminalFailure !== "") throw new AudioGenError(`增强失败：LLM 调用出错（${terminalFailure}）。请检查「设置 → 模型」的默认模型是否可用`, "enhance-llm-error");
+		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 stripFences(output.trim());
+	return result;
 }
 /** 去掉模型可能包裹的 ``` 代码围栏。 */
 function stripFences(value) {
@@ -1860,6 +1890,25 @@ function makeRoutes(deps) {
 				});
 			}
 		},
+		{
+			kind: "exact",
+			path: LLM_MODELS_API,
+			handler: async (req, res) => {
+				if (!guard(req, res, "POST")) return;
+				try {
+					writeJson(res, 200, {
+						ok: true,
+						providers: await deps.llmModelOptions()
+					});
+				} catch (error) {
+					writeJson(res, 200, {
+						ok: false,
+						code: "llm-models-failed",
+						message: messageOf(error)
+					});
+				}
+			}
+		},
 		{
 			kind: "exact",
 			path: MODEL_API.discover,
@@ -3123,7 +3172,8 @@ const Config = z.object({
 	defaultChannelId: z.string().default(""),
 	defaultModel: z.string().default(""),
 	autoSaveToLibrary: z.boolean().default(false),
-	maxConcurrentGenerations: z.union([z.number(), z.string()]).default(DEFAULT_MAX_CONCURRENT)
+	maxConcurrentGenerations: z.union([z.number(), z.string()]).default(DEFAULT_MAX_CONCURRENT),
+	enhanceModel: z.string().default("")
 });
 const DEFAULT_ENABLED = true;
 const DEFAULT_ANNOUNCE = true;
@@ -3216,6 +3266,7 @@ function apply(ctx, config) {
 			})),
 			defaultChannelId,
 			defaultModel: typeof value.defaultModel === "string" ? value.defaultModel.trim() : "",
+			enhanceModel: typeof value.enhanceModel === "string" ? value.enhanceModel.trim() : "",
 			autoSaveToLibrary: value.autoSaveToLibrary === true,
 			maxConcurrentGenerations: (() => {
 				const rawMax = value.maxConcurrentGenerations;
@@ -3231,7 +3282,51 @@ function apply(ctx, config) {
 		return enhancePromptText({
 			settings: seam,
 			llm: () => ctx.get("llm")
-		}, prompt, mode);
+		}, prompt, mode, enhanceSelectionOf(resolve()));
+	};
+	/** 解析设置的增强模型（"provider|model"）；空/非法值返回 undefined = 跟随默认。 */
+	const enhanceSelectionOf = (config) => {
+		const raw = config.enhanceModel ?? "";
+		const sep = raw.indexOf("|");
+		if (sep <= 0 || sep >= raw.length - 1) return void 0;
+		const provider = raw.slice(0, sep).trim();
+		const model = raw.slice(sep + 1).trim();
+		return provider !== "" && model !== "" ? {
+			provider,
+			model
+		} : void 0;
+	};
+	/** 读取「设置 → 模型」目录：各提供方 + 可广播的模型列表（增强模型下拉候选）。 */
+	const llmModelOptions = async () => {
+		const llm = ctx.get("llm");
+		if (llm === void 0 || llm.listProviders === void 0 || llm.listModels === void 0) return [];
+		const options = [];
+		const directory = /* @__PURE__ */ new Map();
+		if (llm.listConfigurableProviders !== void 0) {
+			for (const entry of llm.listConfigurableProviders() ?? []) if (typeof entry.provider === "string" && entry.provider !== "" && typeof entry.displayName === "string") directory.set(entry.provider, entry.displayName);
+		}
+		for (const info of llm.listProviders() ?? []) {
+			const provider = typeof info.id === "string" ? info.id : "";
+			if (provider === "") continue;
+			let models = [];
+			try {
+				models = await llm.listModels(provider) ?? [];
+			} catch {
+				models = [];
+			}
+			for (const model of models) {
+				const id = typeof model.id === "string" ? model.id.trim() : "";
+				if (id === "") continue;
+				const name = typeof model.name === "string" && model.name.trim() !== "" ? model.name.trim() : id;
+				options.push({
+					provider,
+					providerName: directory.get(provider) ?? provider,
+					id,
+					name
+				});
+			}
+		}
+		return options;
 	};
 	const channelsView = () => {
 		const value = resolve();
@@ -3248,7 +3343,8 @@ function apply(ctx, config) {
 				resolveChannels: channelsView,
 				autoSave: () => resolve().autoSaveToLibrary,
 				budget,
-				enhance
+				enhance,
+				llmModelOptions
 			}).map((route) => ctx.webServer.register(route));
 			return () => {
 				for (const dispose of disposers) dispose();

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": "0.4.5",
+  "version": "0.4.7",
   "type": "module",
   "main": "lib/index.js",
   "exports": {

package/src/client/SettingsCard.tsx

@@ -18,7 +18,7 @@ import { createSnapshotStore, type SnapshotStore } from '@deepseek-ai/dsh-client
 import { CardForm, booleanField, textField, type CardActions, type CardShell, type FieldState as CardFieldState } from './settings-form.ts'
 import { ChannelsForm, type ChannelDraft, type ChannelsFormActions, type ChannelsFormState } from './channels-form.ts'
 import type { AudiogenScope } from './settings-scope.ts'
-import { MODEL_API, PRESETS_API, type AudioModelCategory, type DiscoveredAudioModel, type ModelMapping, type PresetProviderView } from '../protocol.ts'
+import { MODEL_API, PRESETS_API, LLM_MODELS_API, type AudioModelCategory, type DiscoveredAudioModel, type LlmModelOption, type ModelMapping, type PresetProviderView } from '../protocol.ts'
 import type { AudioGenKey } from './locales.ts'
 import css from './settings-card.module.css'
 
@@ -29,6 +29,7 @@ export interface AudioGenSettings {
   defaultModel?: string
   autoSaveToLibrary?: boolean
   maxConcurrentGenerations?: number
+  enhanceModel?: string
 }
 
 export interface AudioGenSettingsCardState extends CardShell {
@@ -39,6 +40,7 @@ export interface AudioGenSettingsCardState extends CardShell {
   defaultModel: CardFieldState
   autoSaveToLibrary: CardFieldState
   maxConcurrentGenerations: CardFieldState
+  enhanceModel: CardFieldState
 }
 
 export interface AudioGenSettingsCardFace extends CardActions {
@@ -60,6 +62,7 @@ export class AudioGenSettingsCardController {
       textField('defaultModel'),
       booleanField('autoSaveToLibrary'),
       textField('maxConcurrentGenerations'),
+      textField('enhanceModel'),
     ])
     this.channelsForm = new ChannelsForm(scope)
   }
@@ -76,6 +79,7 @@ export class AudioGenSettingsCardController {
       defaultModel: this.form.field('defaultModel'),
       autoSaveToLibrary: this.form.field('autoSaveToLibrary'),
       maxConcurrentGenerations: this.form.field('maxConcurrentGenerations'),
+      enhanceModel: this.form.field('enhanceModel'),
     }
   }
 
@@ -387,6 +391,9 @@ function ChannelEditor(props: ChannelEditorProps): React.JSX.Element {
         onAdopt={() => adoptCandidates()}
         onCloseCandidates={closeCandidates}
       />
+      {/stability/i.test(`${props.channel?.preset ?? invited?.id ?? presetId}|${url}`) ? (
+        <p className={css.sectionHint}>{t('channel.stabilityHint')}</p>
+      ) : null}
       <label className={css.field}>
         <span className={css.label}>
           <input type="checkbox" checked={isDefault} disabled={!writable} onChange={event => setIsDefault(event.target.checked)} /> {t('channel.default')}
@@ -469,20 +476,32 @@ function ModelCatalog(props: ModelCatalogProps): React.JSX.Element {
                 disabled={!writable}
                 onChange={event => props.onPatchModel(index, { id: event.target.value })}
               />
-              <select
-                className={`${css.select} ${css.modelCategorySelect}`}
-                value={model.category ?? ''}
-                aria-label={`${t('channel.modelCategory')} ${index + 1}`}
-                disabled={!writable}
-                onChange={event => {
-                  const value = event.target.value as AudioModelCategory | ''
-                  props.onPatchModel(index, value === '' ? { category: undefined } : { category: value })
-                }}
-              >
-                {MODEL_CATEGORIES.map(category => (
-                  <option key={category ?? 'auto'} value={category ?? ''}>{category === undefined ? t('channel.category.auto') : t(`channel.category.${category}`)}</option>
-                ))}
-              </select>
+              {/^stable-audio-/i.test(model.id) ? (
+                <select
+                  className={`${css.select} ${css.modelCategorySelect}`}
+                  value="__stable-dual__"
+                  aria-label={`${t('channel.modelCategory')} ${index + 1}`}
+                  disabled
+                  title={t('channel.category.stableDualHint')}
+                >
+                  <option value="__stable-dual__">{t('channel.category.stableDual')}</option>
+                </select>
+              ) : (
+                <select
+                  className={`${css.select} ${css.modelCategorySelect}`}
+                  value={model.category ?? ''}
+                  aria-label={`${t('channel.modelCategory')} ${index + 1}`}
+                  disabled={!writable}
+                  onChange={event => {
+                    const value = event.target.value as AudioModelCategory | ''
+                    props.onPatchModel(index, value === '' ? { category: undefined } : { category: value })
+                  }}
+                >
+                  {MODEL_CATEGORIES.map(category => (
+                    <option key={category ?? 'auto'} value={category ?? ''}>{category === undefined ? t('channel.category.auto') : t(`channel.category.${category}`)}</option>
+                  ))}
+                </select>
+              )}
               <button
                 type="button"
                 className={css.modelRowRemove}
@@ -554,6 +573,8 @@ export function AudioGenSettingsCard(props: AudioGenSettingsCardProps) {
   const [presetLoading, setPresetLoading] = useState(false)
   const [presetError, setPresetError] = useState<string | null>(null)
   const [confirmDeleteId, setConfirmDeleteId] = useState<string | null>(null)
+  const [llmModels, setLlmModels] = useState<LlmModelOption[] | null>(null)
+  const [llmModelsError, setLlmModelsError] = useState<string | null>(null)
 
   const channels = state.channels.channels
   const editingChannel = editor !== null && editor.kind === 'edit'
@@ -575,6 +596,18 @@ export function AudioGenSettingsCard(props: AudioGenSettingsCardProps) {
     }
   }, [editor, presets.length, presetError, presetLoading])
 
+  // 展开卡片时加载「设置 → 模型」的 LLM 提供方/模型列表（提示词增强模型下拉）。
+  useEffect(() => {
+    if (!open || llmModels !== null || llmModelsError !== null) return
+    void fetch(LLM_MODELS_API, { method: 'POST' })
+      .then(async response => {
+        const body = await response.json() as { ok?: boolean; providers?: LlmModelOption[]; message?: string }
+        if (!response.ok || body.ok !== true || body.providers === undefined) throw new Error(body.message ?? `HTTP ${response.status}`)
+        setLlmModels(body.providers)
+      })
+      .catch(error => setLlmModelsError(error instanceof Error ? error.message : String(error)))
+  }, [open, llmModels, llmModelsError])
+
   if (!state.available) return null
 
   const blocked = !state.dirty || state.invalid || state.saving || state.channels.saving
@@ -707,6 +740,31 @@ export function AudioGenSettingsCard(props: AudioGenSettingsCardProps) {
             ) : null}
           </section>
 
+          <div className={css.field}>
+            <label className={css.label}>
+              <span>{t('settings.enhanceModel')}</span>
+              <select
+                className={css.select}
+                value={state.enhanceModel.text}
+                disabled={!state.writable}
+                onChange={event => props.edit('enhanceModel', event.target.value)}
+              >
+                <option value="">{t('settings.enhanceModelDefault')}</option>
+                {llmModels !== null ? llmModels.map(option => (
+                  <option key={`${option.provider}|${option.id}`} value={`${option.provider}|${option.id}`}>
+                    {`${option.providerName} — ${option.name}${option.name !== option.id ? `（${option.id}）` : ''}`}
+                  </option>
+                )) : null}
+                {llmModels !== null && state.enhanceModel.text !== ''
+                  && !llmModels.some(option => `${option.provider}|${option.id}` === state.enhanceModel.text) ? (
+                    <option value={state.enhanceModel.text}>{state.enhanceModel.text}</option>
+                  ) : null}
+              </select>
+              <span className={css.sectionHint}>{t('settings.enhanceModelHint')}</span>
+              {llmModelsError !== null ? <span className={css.failed}>{t('settings.enhanceModelFailed')}：{llmModelsError}</span> : null}
+            </label>
+          </div>
+
           <div className={css.field}>
             <label className={css.label}>
               <input type="checkbox" checked={state.enabled.text === 'true' || state.enabled.text === ''} disabled={!state.writable} onChange={event => props.edit('enabled', String(event.target.checked))} /> {t('settings.enabled')}

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;
@@ -260,6 +260,7 @@
 .resultCol {
   flex: 1;
   min-width: 0;
+  max-width: 760px;
   display: flex;
   flex-direction: column;
   gap: 10px;
@@ -1314,3 +1315,9 @@
   display: flex;
   gap: 6px;
 }
+
+.enhanceActionsRow {
+  display: flex;
+  justify-content: flex-end;
+  margin-top: -4px;
+}

package/src/client/locales.ts

@@ -37,6 +37,10 @@ export const zh = {
   'settings.allowAgentAudio': '允许 Agent 调用音频生成',
   'settings.autoSaveLibrary': '生成后自动保存到资源库',
   'settings.maxConcurrent': '最大并发生成数（同时打到上游的请求数，默认 5）',
+  'settings.enhanceModel': '提示词增强模型（LLM）',
+  'settings.enhanceModelDefault': '跟随 Agent 默认模型（设置 → 模型）',
+  'settings.enhanceModelHint': '用于「✨ 增强提示词」与 generate_audio 的 enhance_prompt 参数；留空则使用「设置 → 模型」中的默认模型。',
+  'settings.enhanceModelFailed': '模型列表加载失败',
   'settings.save': '保存',
   'settings.saving': '保存中…',
   'settings.discard': '放弃修改',
@@ -89,6 +93,8 @@ export const zh = {
   'channel.category.sfx': '音效',
   'channel.category.voice_design': '音色设计',
   'channel.category.voice_clone': '音色克隆',
+  'channel.category.stableDual': '音乐 + 音效（自动）',
+  'channel.category.stableDualHint': 'Stable Audio 模型按模型名自动识别为音乐+音效，无需设置分类。',
   'channel.addModel': '添加模型',
   'channel.removeModel': '移除',
   'channel.fetchModels': '获取可用模型',
@@ -97,6 +103,7 @@ export const zh = {
   'channel.fetchEmpty': '未发现音频相关模型。',
   'channel.discoverSource': '来源：{source}',
   'channel.candidates': '发现 {n} 个可用模型/音色',
+  'channel.stabilityHint': 'Stable Audio 模型（stable-audio-*）同时适用于「音乐生成」与「音效生成」，面板按模型名自动识别；分类仅影响默认展示。官方不支持 TTS。',
   'channel.selectAll': '全选',
   'channel.clearSelection': '清空',
   'channel.adoptSelected': '添加所选（{n}）',
@@ -142,6 +149,10 @@ export const en: Record<AudioGenKey, string> = {
   'settings.allowAgentAudio': 'Allow agents to generate audio',
   'settings.autoSaveLibrary': 'Auto-save generated audio to the library',
   'settings.maxConcurrent': 'Max concurrent generations (in-flight upstream calls, default 5)',
+  'settings.enhanceModel': 'Prompt enhance model (LLM)',
+  'settings.enhanceModelDefault': 'Follow the agent default model (Settings → Models)',
+  'settings.enhanceModelHint': 'Used by "✨ Enhance prompt" and the generate_audio enhance_prompt parameter; empty follows the default model in Settings → Models.',
+  'settings.enhanceModelFailed': 'Failed to load the model list',
   'settings.save': 'Save',
   'settings.saving': 'Saving…',
   'settings.discard': 'Discard',
@@ -194,6 +205,8 @@ export const en: Record<AudioGenKey, string> = {
   'channel.category.sfx': 'SFX',
   'channel.category.voice_design': 'Voice design',
   'channel.category.voice_clone': 'Voice clone',
+  'channel.category.stableDual': 'Music + SFX (auto)',
+  'channel.category.stableDualHint': 'Stable Audio models are auto-detected as Music + SFX by model name; no category needed.',
   'channel.addModel': 'Add model',
   'channel.removeModel': 'Remove',
   'channel.fetchModels': 'Fetch available models',
@@ -202,6 +215,7 @@ export const en: Record<AudioGenKey, string> = {
   'channel.fetchEmpty': 'No audio-related models found.',
   'channel.discoverSource': 'Source: {source}',
   'channel.candidates': '{n} available model(s)/voice(s) found',
+  'channel.stabilityHint': 'Stable Audio models (stable-audio-*) apply to both Music and SFX; the panel detects them by model name. The category only affects the default display. Official API does not support TTS.',
   'channel.selectAll': 'Select all',
   'channel.clearSelection': 'Clear',
   'channel.adoptSelected': 'Add selected ({n})',

package/src/client/settings-scope.ts

@@ -31,6 +31,8 @@ export interface AudiogenConfig {
   defaultModel?: string
   /** 生成完成后自动保存到资源库（面板与 Agent 生成均生效）。 */
   autoSaveToLibrary?: boolean
+  /** 提示词增强模型（"provider|model"；空串 = 跟随 Agent 默认模型）。 */
+  enhanceModel?: string
 }
 
 /** One settings path-op as the bridge consumes it. */

package/src/client/studio-view.tsx

@@ -16,11 +16,26 @@ import {
   HISTORY_API,
   type AudioMode, type GeneratedAudio, type GenerateAudioRequest, type HistoryEntry, type LibraryEntry,
 } from '../protocol.ts'
+import type { ModelOption } from './settings-scope.ts'
 import { AudioPlayer } from './audio-player.tsx'
 import { LibrarySaveDialog, type SaveDialogContext } from './library-save-dialog.tsx'
 import { CheckIcon, DownloadIcon, StarIcon } from './icons.tsx'
 import css from './audio-panel.module.css'
 
+/**
+ * 模型是否适用于当前模式。
+ * - stable-audio-*（Stable Audio 系列）：官方 text-to-audio 协议对音乐与音效是同一接口，
+ *   因此同时适用于「音乐生成」与「音效生成」；官方不支持 TTS（语音合成），故不出现在 TTS。
+ * - 其余模型：按设置中的分类（category）匹配，auto（未分类）适用于全部模式。
+ */
+function modelSuitableForMode(entry: ModelOption, mode: AudioMode): boolean {
+  if (mode === 'voice_design') return false
+  if (/^stable-audio-/i.test(entry.alias)) return mode === 'music' || mode === 'sfx'
+  if (entry.category === undefined) return true
+  if (entry.category === mode) return true
+  return entry.category === 'tts' && mode === 'tts'
+}
+
 export interface StudioReuse {
   nonce: number
   mode: AudioMode
@@ -209,7 +224,13 @@ export function StudioView(props: {
   })
 
   const [mode, setMode] = useState<AudioMode>('tts')
-  const [prompt, setPrompt] = useState('')
+  /** 每个模式独立的输入内容（TTS 文本 / 音乐·音效提示词 / 音色描述），切模式互不干扰。 */
+  const [promptByMode, setPromptByMode] = useState<Record<AudioMode, string>>({ tts: '', music: '', sfx: '', voice_design: '' })
+  const prompt = promptByMode[mode] ?? ''
+  const setPrompt = (next: string): void => {
+    const target = mode
+    setPromptByMode(current => (current[target] === next ? current : { ...current, [target]: next }))
+  }
   const [previewText, setPreviewText] = useState('')
   const [model, setModel] = useState('')
   const [voice, setVoice] = useState('')
@@ -233,6 +254,8 @@ export function StudioView(props: {
   // 提示词增强
   const [enhancing, setEnhancing] = useState(false)
   const [enhancePreview, setEnhancePreview] = useState<string | null>(null)
+  // 切换模式后增强预览属于旧模式，清除以免误用（prompt 本身按模式独立保留）。
+  useEffect(() => { setEnhancePreview(null) }, [mode])
   // Stable Audio 参数（仅 Stability 渠道显示）
   const [seed, setSeed] = useState('')
   const [steps, setSteps] = useState('')
@@ -273,7 +296,7 @@ export function StudioView(props: {
   const visibleModels = useMemo(() => {
     if (mode === 'voice_design') return []
     return modelOptions.models
-      .filter(entry => entry.category === undefined || entry.category === 'tts' && mode === 'tts' || entry.category === mode)
+      .filter(entry => modelSuitableForMode(entry, mode))
       .map(entry => entry.alias)
   }, [modelOptions.models, mode])
 
@@ -527,6 +550,11 @@ 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) {
+      // 直接写入被恢复模式自己的槽位（此刻闭包 mode 仍是旧值）。
+      setPromptByMode(current => (current[modeValue] === promptValue ? current : { ...current, [modeValue]: promptValue }))
+    }
     const modelValue = str('model') ?? singleModel
     if (modelValue !== '') setModel(modelValue)
     if (compareModelsRestore !== undefined && compareModelsRestore.length > 0) {
@@ -591,11 +619,11 @@ export function StudioView(props: {
     setError(null)
     try {
       const result = await api.enhancePrompt(prompt.trim(), mode)
-      if (result.ok !== true || result.enhanced === undefined) {
+      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)
+      setEnhancePreview(result.enhanced.trim())
     } catch (err) {
       setError(err instanceof Error ? err.message : String(err))
     } finally {
@@ -876,6 +904,8 @@ export function StudioView(props: {
 
   const needModel = mode !== 'voice_design'
   const runningCount = tasks.filter(task => task.status === 'running').length
+  /** 结果列只显示当前模式的任务（历史面板已按模式分组）。 */
+  const visibleTasks = useMemo(() => tasks.filter(task => task.mode === mode), [tasks, mode])
 
   return (
     <div className={css.studio}>
@@ -895,16 +925,16 @@ export function StudioView(props: {
           ))}
         </div>
 
-        <div className={css.formSectionRow}>
-          <p className={css.formSection}>输入</p>
-          <button type="button" className={css.ghostButton} disabled={enhancing} onClick={() => void runEnhance()}>
-            {enhancing ? '增强中…' : '✨ 增强提示词'}
-          </button>
-        </div>
+        <p className={css.formSection}>输入</p>
         <label className={css.label}>
           <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}>
@@ -1029,7 +1059,11 @@ export function StudioView(props: {
                 {visibleModels.length === 0 ? <option value="">（当前模式暂无可用模型）</option> : null}
                 {groupedModels.map(group => (
                   <optgroup key={group.channelId} label={group.channelName}>
-                    {group.models.map(item => <option key={item.alias} value={item.alias}>{item.alias}</option>)}
+                    {group.models.map(item => (
+                      <option key={item.alias} value={item.alias}>
+                        {/^stable-audio-/i.test(item.alias) ? `${item.alias} · 音乐/音效` : item.alias}
+                      </option>
+                    ))}
                   </optgroup>
                 ))}
               </select>
@@ -1072,7 +1106,7 @@ export function StudioView(props: {
 
       <div className={css.resultCol}>
         {error !== null ? <p className={css.error}>{error}</p> : null}
-        {tasks.length === 0 ? (
+        {visibleTasks.length === 0 ? (
           <div className={css.resultEmpty}>
             <span className={css.resultEmptyIcon}>🎵</span>
             <p>{tt('result.empty')}</p>
@@ -1089,7 +1123,7 @@ export function StudioView(props: {
           </div>
         ) : (
           <div className={css.taskList}>
-            {tasks.map(task => {
+            {visibleTasks.map(task => {
               const elapsed = task.finishedAt !== undefined
                 ? Math.round((task.finishedAt - task.startedAt) / 1000)
                 : Math.round((Date.now() - task.startedAt) / 1000)