@henjicc/ai-sdk 0.8.0 → 0.10.0

package/dist/index.d.ts

@@ -20,9 +20,17 @@ export * from './protocols/index.js';
 export * from './upload/index.js';
 export * from './providers/index.js';
 export * from './types/index.js';
+export type { SpeechTimings, GenerationOutputTimings } from './types/speechTimings.js';
 export * from './structured-output/index.js';
 export * from './catalog/index.js';
 export * from './pricing/index.js';
 export * from './llm/index.js';
 export * from './capabilities/index.js';
 export * from './discovery/index.js';
+export * from './capabilities/speech-recognition/index.js';
+export * from './capabilities/speech-recognition/bailian/index.js';
+export * from './capabilities/speech-recognition/volcengine/index.js';
+export * from './capabilities/speech-recognition/siliconflow/index.js';
+export * from './capabilities/speech-recognition/groq/index.js';
+export * from './capabilities/translation/index.js';
+export * from './capabilities/translation/bailian/index.js';

package/dist/index.js

@@ -38,3 +38,12 @@ export * from './llm/index.js';
 // 开放能力模块协议与生成/LLM 执行内核解耦；ASR/OCR 等新种类无需改动根 client switch。
 export * from './capabilities/index.js';
 export * from './discovery/index.js';
+// Electron 主进程使用经典 Node 模块解析，无法可靠解析 package exports 子路径。
+// 把已经公开在子路径的非实时 ASR 模块同时从根入口导出，供正式宿主组合使用。
+export * from './capabilities/speech-recognition/index.js';
+export * from './capabilities/speech-recognition/bailian/index.js';
+export * from './capabilities/speech-recognition/volcengine/index.js';
+export * from './capabilities/speech-recognition/siliconflow/index.js';
+export * from './capabilities/speech-recognition/groq/index.js';
+export * from './capabilities/translation/index.js';
+export * from './capabilities/translation/bailian/index.js';

package/dist/llm/bigmodel/preset.js

@@ -1,5 +1,6 @@
 import { createLlmCapabilitiesForModel } from '../defaults.js';
 import { resolveLlmEndpointIdentity } from '../endpointProfiles.js';
+import { findLlmModelCatalogEntry } from '../modelCatalog.js';
 import { BIGMODEL_ENDPOINT_PROFILE_FAMILY, BIGMODEL_PROVIDER_FAMILY_ID, } from './profiles.js';
 export function createBigmodelProvider(options = {}) {
     const profileId = options.endpointProfile ?? 'cn';
@@ -36,7 +37,7 @@ export function createBigmodelModels(provider) {
     const profile = BIGMODEL_ENDPOINT_PROFILE_FAMILY.profiles.find(item => item.id === identity.endpointProfile);
     if (!profile)
         throw new Error('[llm_endpoint_profile_unknown] resolved BigModel endpoint profile is unavailable');
-    return profile.modelIds.map(modelId => ({
+    return profile.modelIds.filter(modelId => !findLlmModelCatalogEntry(modelId)?.deprecated).map(modelId => ({
         providerId: identity.providerId,
         providerFamilyId: identity.providerFamilyId,
         endpointProfile: identity.endpointProfile,

package/dist/llm/bigmodel/profiles.js

@@ -42,12 +42,6 @@ export const BIGMODEL_GLM_5_3_FLASH_PRICING = {
         inputPerMillionTokens: 0.8,
         outputPerMillionTokens: 2.8,
         cacheReadPerMillionTokens: 0.23,
-        promotion: {
-            inputPerMillionTokens: 0.4,
-            outputPerMillionTokens: 1.4,
-            cacheReadPerMillionTokens: 0.115,
-            observedAt: '2026-08-28',
-        },
     },
     global: {
         currency: 'USD',

package/dist/llm/chat.d.ts

@@ -27,6 +27,10 @@ export interface LlmChatCompletedInfo {
     output: string;
     reasoningOutput: string;
     usage: LlmUsageDto | null;
+    /** 发出请求到第一个字（含思考内容）的毫秒数；没有任何输出时为 null。 */
+    firstTokenMs: number | null;
+    /** 首字之后的输出速度（输出 token / 秒，含思考 token）；供应商没给用量时为 null。 */
+    outputTokensPerSecond: number | null;
     finishReason: string | null;
     truncated: boolean;
     toolCalls?: LlmStreamToolCall[];
@@ -47,6 +51,10 @@ export interface LlmChatStreamOutcome {
     output: string;
     reasoningOutput: string;
     usage: LlmUsageDto | null;
+    /** 发出请求到第一个字（含思考内容）的毫秒数；没有任何输出时为 null。 */
+    firstTokenMs: number | null;
+    /** 首字之后的输出速度（输出 token / 秒，含思考 token）；供应商没给用量时为 null。 */
+    outputTokensPerSecond: number | null;
     finishReason: string | null;
     truncated: boolean;
     toolCalls?: LlmStreamToolCall[];

package/dist/llm/chat.js

@@ -59,15 +59,27 @@ export async function runLlmChatStream(request, taskId, emit, runtime, hooks = {
             requestPayload: cloneJsonObject(requestPayload),
             processedRequest,
         });
+        // 首字时间从真正发出请求算起，不含前面的媒体预处理
+        const requestSentAtMs = Date.now();
+        let firstTokenAtMs = null;
         const output = await streamOpenAiCompatibleChat({
             endpoint,
             apiKey,
             providerId: identity.providerFamilyId,
             payload: requestPayload,
             signal: controller.signal,
-            emit,
+            emit: (event) => {
+                if (firstTokenAtMs === null && (event.type === 'Token' || event.type === 'ReasoningToken') && event.data)
+                    firstTokenAtMs = Date.now();
+                emit(event);
+            },
             transport: runtime.transport,
         });
+        const finishedAtMs = Date.now();
+        const firstTokenMs = firstTokenAtMs === null ? null : firstTokenAtMs - requestSentAtMs;
+        const streamingSeconds = firstTokenAtMs === null ? 0 : (finishedAtMs - firstTokenAtMs) / 1000;
+        const outputTokens = output.usage?.outputTokens ?? null;
+        const outputTokensPerSecond = outputTokens !== null && streamingSeconds > 0 ? outputTokens / streamingSeconds : null;
         if (isCancelled('llm', taskId)) {
             throw new Error(`[task_cancelled] LLM task cancelled: ${taskId}`);
         }
@@ -82,6 +94,8 @@ export async function runLlmChatStream(request, taskId, emit, runtime, hooks = {
             output: output.output,
             reasoningOutput: output.reasoningOutput,
             usage: output.usage,
+            firstTokenMs,
+            outputTokensPerSecond,
             finishReason: output.finishReason,
             truncated: output.truncated,
             toolCalls: output.toolCalls,
@@ -97,6 +111,8 @@ export async function runLlmChatStream(request, taskId, emit, runtime, hooks = {
             output: output.output,
             reasoningOutput: output.reasoningOutput,
             usage: output.usage,
+            firstTokenMs,
+            outputTokensPerSecond,
             finishReason: output.finishReason,
             truncated: output.truncated,
             toolCalls: output.toolCalls,
@@ -130,7 +146,7 @@ export async function runLlmChatStream(request, taskId, emit, runtime, hooks = {
 async function preprocessLlmRequest(request, runtime) {
     const body = { messages: request.messages };
     const params = request.metadata ?? {};
-    const processed = await preprocessRequestBody(request.providerFamilyId ?? request.providerId, '/v1/chat/completions', body, runtime, params);
+    const processed = await preprocessRequestBody(request.providerFamilyId ?? request.providerId, '/v1/chat/completions', body, runtime, params, undefined, undefined, undefined, 'llm');
     const processedBody = isJsonObject(processed) ? processed : body;
     const processedMessages = Array.isArray(processedBody.messages)
         ? processedBody.messages

package/dist/llm/chatTypes.d.ts

@@ -51,6 +51,12 @@ export interface LlmTraceDto {
     elapsedMs: number;
     inputChars: number;
     outputChars: number;
+    /** 发出请求到第一个字的毫秒数（含思考内容） */
+    firstTokenMs?: number | null;
+    /** 首字之后的输出 token 速度（个/秒） */
+    outputTokensPerSecond?: number | null;
+    inputTokens?: number | null;
+    outputTokens?: number | null;
 }
 export type LlmStreamEventDto = {
     type: 'Token';

package/dist/llm/defaults.d.ts

@@ -13,7 +13,7 @@ export declare const DEFAULT_DEEPSEEK_BASE_URL = "https://api.deepseek.com";
 export declare const DEFAULT_DEEPSEEK_MODEL_ID = "deepseek-flash";
 export declare const DEFAULT_PPIO_PROVIDER_ID = "ppio";
 export declare const DEFAULT_PPIO_BASE_URL = "https://api.ppio.com/openai";
-export declare const DEFAULT_PPIO_MODEL_ID = "deepseek/deepseek-v4-flash";
+export declare const DEFAULT_PPIO_MODEL_ID = "deepseek/deepseek-v4.1-flash";
 export declare const DEFAULT_LLM_REASONING_EFFORT: LlmReasoningEffort;
 export declare const DEEPSEEK_V4_CONTEXT_WINDOW = 1000000;
 export declare const DEEPSEEK_V4_MAX_OUTPUT_TOKENS = 384000;
@@ -24,6 +24,6 @@ export declare const DEFAULT_LLM_CAPABILITIES: LlmCapabilities;
  * 添加模型的两个入口（手动添加、获取模型列表）和内置模型清单都走这里，保证"同一个模型 ID
  * 无论从哪个入口进来，标出来的能力一致"。
  */
-export declare function createLlmCapabilitiesForModel(modelId: string): LlmCapabilities;
+export declare function createLlmCapabilitiesForModel(modelId: string, providerId?: string): LlmCapabilities;
 /** DeepSeek 官方要求思考模式默认开启，其余供应商默认关闭；具体档位见 `DEFAULT_LLM_REASONING_EFFORT`。 */
 export declare function createDefaultProviderReasoning(adapter?: string): LlmReasoningConfig;

package/dist/llm/defaults.js

@@ -13,7 +13,7 @@ export const DEFAULT_DEEPSEEK_BASE_URL = 'https://api.deepseek.com';
 export const DEFAULT_DEEPSEEK_MODEL_ID = 'deepseek-flash';
 export const DEFAULT_PPIO_PROVIDER_ID = 'ppio';
 export const DEFAULT_PPIO_BASE_URL = 'https://api.ppio.com/openai';
-export const DEFAULT_PPIO_MODEL_ID = 'deepseek/deepseek-v4-flash';
+export const DEFAULT_PPIO_MODEL_ID = 'deepseek/deepseek-v4.1-flash';
 export const DEFAULT_LLM_REASONING_EFFORT = 'high';
 export const DEEPSEEK_V4_CONTEXT_WINDOW = 1_000_000;
 export const DEEPSEEK_V4_MAX_OUTPUT_TOKENS = 384_000;
@@ -41,8 +41,8 @@ export const DEFAULT_LLM_CAPABILITIES = {
  * 添加模型的两个入口（手动添加、获取模型列表）和内置模型清单都走这里，保证"同一个模型 ID
  * 无论从哪个入口进来，标出来的能力一致"。
  */
-export function createLlmCapabilitiesForModel(modelId) {
-    const entry = findLlmModelCatalogEntry(modelId);
+export function createLlmCapabilitiesForModel(modelId, providerId) {
+    const entry = findLlmModelCatalogEntry(modelId, providerId);
     return entry
         ? applyLlmModelCatalogEntry(DEFAULT_LLM_CAPABILITIES, entry)
         : { ...DEFAULT_LLM_CAPABILITIES };

package/dist/llm/groq/preset.js

@@ -1,3 +1,4 @@
+import { tokenListPrice } from '../../pricing/tokenPricing.js';
 import { findProviderMetadata } from '../../providers/metadata.js';
 export const GROQ_PROVIDER_ID = 'groq';
 export const GROQ_BASE_URL = 'https://api.groq.com/openai/v1';
@@ -14,11 +15,26 @@ export const GROQ_PROVIDER_PRESET = {
     baseUrl: GROQ_BASE_URL,
     reasoning: { enabled: true, effort: 'medium' },
     reasoningConfigurable: true,
-    modelIds: [GROQ_DEFAULT_MODEL_ID],
+    // 2026-11-07 按官方模型页核对：自助账号可用的对话模型只有 GPT-OSS 两款正式版与 Qwen3.8-27B 预览版；
+    // Llama 系列为企业版，Prompt Guard / Safeguard 是安全分类器，Orpheus 是语音合成，均不进对话预设。
+    modelIds: [GROQ_DEFAULT_MODEL_ID, 'openai/gpt-oss-120b', 'qwen/qwen3.8-27b'],
+    // 托管限制以 Groq 模型页为准：Qwen3.8-27B 上下文 131,072、最大输出 16,384、每次最多 3 张图，支持 JSON Schema。
+    modelCapabilities: {
+        'qwen/qwen3.8-27b': {
+            image: true, video: false, audio: false, file: false,
+            toolCall: true, jsonOutput: true, structuredOutputMode: 'schema', reasoning: true,
+            contextWindow: 131_072, maxOutputTokens: 16_384,
+        },
+    },
+    modelPricing: {
+        [GROQ_DEFAULT_MODEL_ID]: tokenListPrice('USD', 0.075, 0.3),
+        'openai/gpt-oss-120b': tokenListPrice('USD', 0.15, 0.6),
+        'qwen/qwen3.8-27b': tokenListPrice('USD', 0.8, 4),
+    },
     websiteUrl: GROQ_PROVIDER_METADATA.websiteUrl,
     apiKeyUrl: GROQ_PROVIDER_METADATA.apiKeyUrl,
     docs: 'docs/model-adaptation/供应商/Groq.md',
-    note: 'GroqCloud（providerId=groq）不是 xAI Grok；GPT-OSS 不支持 reasoning_format。',
+    note: 'GroqCloud（providerId=groq）不是 xAI Grok；GPT-OSS 不支持 reasoning_format；Qwen3.8-27B 为预览模型，可能随时下线。',
 };
 /** 可直接交给统一能力发现器的默认模型；不会触发模型列表网络请求。 */
 export const GROQ_DEFAULT_MODEL_CONFIG = {

package/dist/llm/modelCatalog.d.ts

@@ -18,6 +18,11 @@ export interface LlmModelCatalogEntry {
     vendor: string;
     /** 规范化后仍与 `id` 不同的其他写法（例如同能力的价格变体） */
     aliases?: readonly string[];
+    /**
+     * 同一模型的不同快照能力不同时，按完整 Model ID（去掉厂商前缀与档位标记、小写）覆盖的字段。
+     * 例：方舟 doubao-seed-2-1-pro 的 260915 快照上下文 1024K，旧快照 260628 只有 256K。
+     */
+    snapshots?: Readonly<Record<string, Partial<Pick<LlmModelCatalogEntry, 'contextWindow' | 'maxOutputTokens' | 'input' | 'reasoning' | 'structuredOutputMode'>>>>;
     /** 本项目当前请求路径下真实可用的输入模态，取值依据见 `modelCatalogEntries.ts` 顶部注释 */
     input: {
         image: boolean;
@@ -38,6 +43,8 @@ export interface LlmModelCatalogEntry {
     /** 模型原厂已确认支持的协议；最终使用协议还要与接入供应商的端点能力取交集。 */
     apiProtocols?: readonly LlmApiProtocol[];
     note?: string;
+    /** 已过时的原因；填写后预设不再自动添加、能力发现不再列出，已有配置不受影响。 */
+    deprecated?: string;
     /** 该条目的资料出处，仓库内相对路径 */
     docs: string;
 }
@@ -47,10 +54,10 @@ export declare function modelSupportsLlmApiProtocol(entry: LlmModelCatalogEntry,
  *
  * 同一个模型经不同渠道进来写法不同：聚合网关带厂商前缀（`deepseek/deepseek-v4-flash`、
  * `xiaomimimo/mimo-v2.5-pro`），官方渠道带快照日期后缀（`doubao-seed-2-1-pro-260628`），
- * 部分网关还带 `:free` 之类的档位标记。这些都指向同一份能力。
+ * 部分网关还带 `:free` 之类的档位标记。规范化只收敛名称，原厂重定向等渠道差异仍按供应商查表。
  */
 export declare function normalizeCatalogModelId(modelId: string): string;
-export declare function findLlmModelCatalogEntry(modelId: string): LlmModelCatalogEntry | null;
+export declare function findLlmModelCatalogEntry(modelId: string, providerId?: string): LlmModelCatalogEntry | null;
 /**
  * 把目录条目盖到一份基础能力上。
  *

package/dist/llm/modelCatalog.js

@@ -1,4 +1,4 @@
-import { LLM_MODEL_CATALOG_ENTRIES } from './modelCatalogEntries.js';
+import { DEEPSEEK_OFFICIAL_FLASH_ALIASES, LLM_MODEL_CATALOG_ENTRIES } from './modelCatalogEntries.js';
 export function modelSupportsLlmApiProtocol(entry, protocol) {
     return (entry.apiProtocols ?? ['openai-compatible']).includes(protocol);
 }
@@ -7,7 +7,7 @@ export function modelSupportsLlmApiProtocol(entry, protocol) {
  *
  * 同一个模型经不同渠道进来写法不同：聚合网关带厂商前缀（`deepseek/deepseek-v4-flash`、
  * `xiaomimimo/mimo-v2.5-pro`），官方渠道带快照日期后缀（`doubao-seed-2-1-pro-260628`），
- * 部分网关还带 `:free` 之类的档位标记。这些都指向同一份能力。
+ * 部分网关还带 `:free` 之类的档位标记。规范化只收敛名称，原厂重定向等渠道差异仍按供应商查表。
  */
 export function normalizeCatalogModelId(modelId) {
     const trimmed = modelId.trim().toLowerCase();
@@ -23,9 +23,22 @@ const CATALOG_BY_KEY = new Map(LLM_MODEL_CATALOG_ENTRIES.flatMap(entry => [
     [entry.id, entry],
     ...(entry.aliases ?? []).map(alias => [normalizeCatalogModelId(alias), entry]),
 ]));
-export function findLlmModelCatalogEntry(modelId) {
+export function findLlmModelCatalogEntry(modelId, providerId) {
     const key = normalizeCatalogModelId(modelId);
-    return key ? CATALOG_BY_KEY.get(key) ?? null : null;
+    const provider = providerId?.trim().toLowerCase();
+    // 无前缀的旧名默认指原厂；显式供应商优先，聚合前缀不能继承原厂重定向。
+    if (provider === 'deepseek' || (!provider && !modelId.includes('/'))) {
+        const official = DEEPSEEK_OFFICIAL_FLASH_ALIASES.find(entry => entry.id === key);
+        if (official)
+            return official;
+    }
+    const entry = key ? CATALOG_BY_KEY.get(key) ?? null : null;
+    const snapshot = entry?.snapshots?.[withoutRouting(modelId)];
+    return entry && snapshot ? { ...entry, ...snapshot } : entry;
+}
+function withoutRouting(modelId) {
+    const trimmed = modelId.trim().toLowerCase();
+    return trimmed.slice(trimmed.lastIndexOf('/') + 1).split(':')[0];
 }
 /**
  * 把目录条目盖到一份基础能力上。

package/dist/llm/modelCatalogEntries.d.ts

@@ -13,3 +13,5 @@ import type { LlmModelCatalogEntry } from './modelCatalog.js';
  * 否则用户会得到"勾了但发过去没反应"的静默失效。
  */
 export declare const LLM_MODEL_CATALOG_ENTRIES: readonly LlmModelCatalogEntry[];
+/** 原厂旧名已转到新 Flash；网关同名仍保留上方的独立旧契约。 */
+export declare const DEEPSEEK_OFFICIAL_FLASH_ALIASES: readonly LlmModelCatalogEntry[];