dsh-audiogen 0.4.6 → 0.4.8

package/src/index.ts

@@ -16,7 +16,7 @@ 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 AudioMode, type ChannelConfig, type ModelMapping } from './protocol.ts'
+import { AUDIOGEN_SETTINGS_NAMESPACE, type AudioMode, type ChannelConfig, type LlmModelOption, type ModelMapping } from './protocol.ts'
 import { createGenerationBudget } from './audio-scheduler.ts'
 import { enhancePromptText } from './prompt-enhance.ts'
 import { AudioGenError } from './audio-engine.ts'
@@ -45,6 +45,11 @@ export interface Config {
   autoSaveToLibrary?: boolean
   /** 设置卡按文本编辑，保存值可能是数字或数字字符串。 */
   maxConcurrentGenerations?: number | string
+  /**
+   * 提示词增强模型，格式 "provider|model"（如 "deepseek-official|deepseek-v4-flash-vision-exp"）；
+   * 空串表示跟随 Agent 默认模型（「设置 → 模型」）。
+   */
+  enhanceModel?: string
 }
 
 const DEFAULT_MAX_CONCURRENT = 5
@@ -68,6 +73,7 @@ export const Config: z<Config> = z.object({
   defaultModel: z.string().default(''),
   autoSaveToLibrary: z.boolean().default(false),
   maxConcurrentGenerations: z.union([z.number(), z.string()]).default(DEFAULT_MAX_CONCURRENT),
+  enhanceModel: z.string().default(''),
 })
 
 const DEFAULT_ENABLED = true
@@ -158,6 +164,8 @@ export interface EffectiveConfig {
   defaultChannelId: string
   defaultModel: string
   autoSaveToLibrary: boolean
+  /** 提示词增强模型（"provider|model"；空串 = 跟随 Agent 默认模型）。 */
+  enhanceModel: string
   maxConcurrentGenerations: number
 }
 
@@ -186,6 +194,7 @@ export function apply(ctx: Context, config?: Config): void {
       })),
       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
@@ -198,11 +207,54 @@ 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。
+  // 提示词增强：优先用设置的增强模型（enhanceModel），否则复用 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)
+    return enhancePromptText({ settings: seam, llm: () => ctx.get('llm') }, 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: EffectiveConfig): { provider: string; model: string } | undefined => {
+    const raw = config.enhanceModel ?? ''
+    const sep = raw.indexOf('|')
+    if (sep <= 0 || sep >= raw.length - 1) return undefined
+    const provider = raw.slice(0, sep).trim()
+    const model = raw.slice(sep + 1).trim()
+    return provider !== '' && model !== '' ? { provider, model } : undefined
+  }
+
+  /** 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 (): Promise<LlmModelOption[]> => {
+    const llm = ctx.get('llm') as {
+      listProviders?: () => Array<{ id?: string; name?: string }>
+      listConfigurableProviders?: () => Array<{ provider?: string; displayName?: string }>
+      listModels?: (provider: string) => Promise<Array<{ id?: string; name?: string }>>
+    } | undefined
+    if (llm === undefined || llm.listProviders === undefined || llm.listModels === undefined) return []
+    const options: LlmModelOption[] = []
+    const directory = new Map<string, string>()
+    if (llm.listConfigurableProviders !== undefined) {
+      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: Array<{ id?: string; name?: string }> = []
+      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 = (): ChannelsView => {
@@ -219,6 +271,7 @@ export function apply(ctx: Context, config?: Config): void {
         autoSave: () => resolve().autoSaveToLibrary,
         budget,
         enhance,
+        llmModelOptions,
       })
       const disposers = routes.map(route => ctx.webServer.register(route))
       return () => { for (const dispose of disposers) dispose() }

package/src/prompt-enhance.ts

@@ -34,12 +34,29 @@ export interface PromptEnhanceDeps {
   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> {
+/** 设置中显式选择的增强模型（provider + model）。 */
+export interface EnhanceModelSelection {
+  provider: string
+  model: string
+}
+
+/** 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。 */
+export async function enhancePromptText(
+  deps: PromptEnhanceDeps,
+  prompt: string,
+  mode: AudioMode,
+  override?: EnhanceModelSelection,
+): 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 selected = override !== undefined && override.provider.trim() !== '' && override.model.trim() !== ''
+    ? { provider: override.provider.trim(), model: override.model.trim() }
+    : undefined
+  const descriptor = selected === undefined
+    ? (deps.settings.describe({ redactSecrets: true }) ?? []).find(candidate => String(candidate.ns) === 'agent-default-model')
+    : undefined
+  const value = selected ?? ((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 === '') {
@@ -53,22 +70,32 @@ export async function enhancePromptText(deps: PromptEnhanceDeps, prompt: string,
   const timer = setTimeout(() => controller.abort(new DOMException('The operation timed out.', 'TimeoutError')), 30_000)
   timer.unref?.()
   let output = ''
+  let terminalFailure = ''
   try {
     for await (const chunk of runtime.stream({
       provider,
       model,
-      messages: [{ role: 'user', content: text }],
+      // dsh-llm 的 Message.content 是 ContentBlock[]（如 [{ type: 'text', text }]），不是纯字符串；
+      // 传字符串会在适配器 contentHasImage() 里抛 "content.some is not a function"，
+      // 主机把该异常转成终止的 finish/error 分片，导致下面永远收集不到文本。
+      messages: [{ role: 'user', content: [{ type: 'text', 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 } }
+      const record = chunk as { type?: string; text?: string; block?: { type?: string; text?: string }; reason?: { kind?: string; failure?: { message?: string; code?: 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
+      } else if (record.type === 'finish' && record.reason !== undefined
+        && record.reason.kind !== 'stop' && record.reason.kind !== undefined && terminalFailure === '') {
+        const failure = record.reason.failure
+        terminalFailure = typeof failure?.message === 'string' && failure.message.trim() !== ''
+          ? `${failure.message}${typeof failure.code === 'string' ? `（${failure.code}）` : ''}`
+          : `stream ${record.reason.kind}`
       }
     }
   } finally {
@@ -76,6 +103,9 @@ export async function enhancePromptText(deps: PromptEnhanceDeps, prompt: string,
   }
   const result = stripFences(output.trim())
   if (result === '') {
+    if (terminalFailure !== '') {
+      throw new AudioGenError(`增强失败：LLM 调用出错（${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 result

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 = '0.4.6'
+export const PLUGIN_VERSION = '0.4.8'
 
 /** Same-origin route family (loopback-only, mirroring dsh-imagegen). */
 export const SETTINGS_API = {
@@ -30,6 +30,9 @@ 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
 
+/** 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）。 */
+export const LLM_MODELS_API = '/api/dsh-audiogen/llm/models' as const
+
 /** Host-mediated model/voice discovery endpoint. */
 export const MODEL_API = {
   discover: '/api/dsh-audiogen/models/discover',
@@ -91,6 +94,18 @@ export interface ModelMapping {
   category?: AudioModelCategory
 }
 
+/** 一个「设置 → 模型」中的 LLM 候选模型（提示词增强模型下拉用）。 */
+export interface LlmModelOption {
+  /** 提供方路由（如 deepseek-official / google）。 */
+  provider: string
+  /** 提供方展示名（如 DeepSeek / google）。 */
+  providerName: string
+  /** 模型 id（上游型号，如 deepseek-v4-flash-vision-exp）。 */
+  id: string
+  /** 模型展示名。 */
+  name: string
+}
+
 /** A model/voice discovered from a vendor endpoint (not yet persisted). */
 export interface DiscoveredAudioModel extends ModelMapping {
   /** Optional human-readable description from the vendor. */

package/src/routes.ts

@@ -16,9 +16,9 @@ import { discoverAudioModels } from './audio-models.ts'
 import { AUDIO_PRESETS } from './audio-presets.ts'
 import { appendHistory, clearHistory, listHistory, readAudioFile, removeHistory, saveAudioFile, listLibrary, saveToLibrary, updateLibraryEntry, removeLibraryEntries, readLibraryFile } from './audio-store.ts'
 import {
-  AUDIO_API, AUDIOGEN_SETTINGS_NAMESPACE, ENHANCE_API, GENERATE_API, HISTORY_API, LIBRARY_API, MODEL_API, PRESETS_API, SETTINGS_API, TASK_API,
+  AUDIO_API, AUDIOGEN_SETTINGS_NAMESPACE, ENHANCE_API, GENERATE_API, HISTORY_API, LIBRARY_API, LLM_MODELS_API, MODEL_API, PRESETS_API, SETTINGS_API, TASK_API,
   LIBRARY_TYPES,
-  type GenerateAudioRequest, type GeneratedAudio, type HistoryEntryInput, type LibraryAudioInput, type LibraryProvenance, type LibraryType,
+  type GenerateAudioRequest, type GeneratedAudio, type HistoryEntryInput, type LibraryAudioInput, type LibraryProvenance, type LibraryType, type LlmModelOption,
 } from './protocol.ts'
 
 const MAX_JSON_BODY_BYTES = 16 * 1024 * 1024
@@ -49,6 +49,8 @@ export interface AudiogenRoutesDeps {
   budget: GenerationBudget
   /** 提示词增强：调用 Agent 默认模型，返回增强后的文本。 */
   enhance: (prompt: string, mode: GenerateAudioRequest['mode']) => Promise<string>
+  /** 「设置 → 模型」提供方列表 + 各自可广播模型（增强模型下拉候选）。 */
+  llmModelOptions: () => Promise<LlmModelOption[]>
 }
 
 function isLoopbackRequest(request: IncomingMessage): boolean {
@@ -142,6 +144,8 @@ function parseGenerateRequest(body: Record<string, unknown>): GenerateAudioReque
     ...(num(body.cfgScale) !== undefined ? { cfgScale: num(body.cfgScale)! } : {}),
     ...(typeof body.format === 'string' && body.format.trim() !== '' ? { format: body.format.trim() } : {}),
     ...(typeof body.channelId === 'string' && body.channelId !== '' ? { channelId: body.channelId } : {}),
+    // 对比任务的任务 id：随 params 一起入历史，供客户端按 taskId 聚合出对比卡片。
+    ...(typeof body.taskId === 'string' && body.taskId.trim() !== '' ? { taskId: body.taskId.trim() } : {}),
     // ---- MiniMax TTS 专属字段（其他厂商忽略） ----
     ...(str(body.emotion) !== undefined ? { emotion: str(body.emotion)! } : {}),
     ...(num(body.vol) !== undefined ? { vol: num(body.vol)! } : {}),
@@ -327,6 +331,20 @@ export function makeRoutes(deps: AudiogenRoutesDeps): WebRoute[] {
         writeJson(res, 200, { ok: true, presets: AUDIO_PRESETS })
       },
     },
+    // --------------------------------------------- LLM models (enhance)
+    {
+      kind: 'exact',
+      path: LLM_MODELS_API,
+      handler: async (req, res) => {
+        if (!guard(req, res, 'POST')) return
+        try {
+          const providers = await deps.llmModelOptions()
+          writeJson(res, 200, { ok: true, providers })
+        } catch (error) {
+          writeJson(res, 200, { ok: false, code: 'llm-models-failed', message: messageOf(error) })
+        }
+      },
+    },
     // ---------------------------------------------- model/voice discovery
     {
       kind: 'exact',