chore(ai): normalize DeepSeek Flash model and pricing

apps/pac-docs/content/docs/algorithms/overview.mdx

@@ -153,7 +153,7 @@ icon: Sparkles
 | **禁词 + 不承诺** | 不许"保证治好" / 报价 / 写死时间(诊所没排班接口,时间只能给"示例") |
 | **失败降级** | 大模型出错或违规 → 退回安全模板,客服永远有东西可用 |
 
-> 📐 **真理源**:提示词组装 `skill-composer.ts` + `skills/*/SKILL.md`(按 frontmatter 声明式匹配);安全规则**单一源** `safety-rules.ts`(机器闸 + prompt 同引,输出后扫禁词 / 加粗时间 / 临床码,命中即兜底);模型默认 `deepseek-v4-flash`;每次落 `agent_invocations`(档位 / 模型 / token / cost / 用了哪些块)。话术块是 MD 文件,**运营可改文案不动代码**。
+> 📐 **真理源**:提示词组装 `skill-composer.ts` + `skills/*/SKILL.md`(按 frontmatter 声明式匹配);安全规则**单一源** `safety-rules.ts`(机器闸 + prompt 同引,输出后扫禁词 / 加粗时间 / 临床码,命中即兜底);模型默认 `deepseek-flash`;每次落 `agent_invocations`(档位 / 模型 / token / cost / 用了哪些块)。话术块是 MD 文件,**运营可改文案不动代码**。
 
 > **本质**:话术不是"一个大模板",而是按患者特征拼装 + 大模型润色 + 安全兜底。
 

apps/pac-docs/content/docs/algorithms/script-generation.mdx

@@ -79,7 +79,7 @@ system prompt = `base common`(`shared/skills/_base/common.md`,三档共用:角
 
 ## provider 与落账
 
-- **模型**:三档默认 `deepseek-v4-flash`(`?model=` 可覆盖)。
+- **模型**:三档默认 `deepseek-flash`(`?model=` 可覆盖)。
 - **落账**:每次经 `AiCallRunner` → `invocation-recorder` 两段式落 `agent_invocations`(tier 嵌在 `promptVersion`、`modelProvider`/`modelName`、token、`costYuan`、source);深度档额外 `attachJudge` 落自检质量分。
 - **写库**:`plan_scripts` upsert(`content` Markdown、`status='ready'`、`source`、`agentInvocationId`)。
 

apps/pac-docs/content/docs/deployment.mdx

@@ -85,7 +85,8 @@ DW_CLICKHOUSE_PASSWORD=<密码>                  # env 名须等于 manifest 的
 # ─── AI(DeepSeek)───
 DEEPSEEK_API_KEY=<key>
 DEEPSEEK_BASE_URL=https://api.deepseek.com
-AI_DEFAULT_MODEL=deepseek-v4-flash
+# 旧 deepseek-v4-flash 配置兼容归一；Flash 成本按高峰价估算，非实际账单
+AI_DEFAULT_MODEL=deepseek-flash
 
 # ─── 定时任务 ───
 # 注:增量 / 滞后监控 / 兜底扫描的代码默认是"永不"哨兵,必须显式设 env 才会调度

apps/pac-service/.env.example

@@ -84,10 +84,10 @@ DEEPSEEK_API_KEY=
 DEEPSEEK_BASE_URL=https://api.deepseek.com
 
 # 默认主力模型
-#   local:      deepseek-v4-flash  (省钱开发)
-#   staging:    deepseek-v4-flash
-#   production: deepseek-v4-pro    (质量优先)
-AI_DEFAULT_MODEL=deepseek-v4-flash
+#   local:      deepseek-flash  (省钱开发)
+#   staging:    deepseek-flash
+#   production: deepseek-flash  (Pro 可单独评估后配置)
+AI_DEFAULT_MODEL=deepseek-flash
 
 AI_REQUEST_TIMEOUT_SEC=60
 
@@ -107,7 +107,8 @@ GEMINI_API_KEY=
 # GEMINI_LIVE_MODEL=gemini-3.1-flash-live-preview
 
 # 价格表(¥/M tokens)— 调价时改这里重启;不配走默认
-# AI_PRICE_TABLE_JSON={"deepseek-v4-pro":{"inHit":0.5,"inMiss":3.6,"out":25},"deepseek-v4-flash":{"inHit":0.07,"inMiss":0.5,"out":2}}
+# Flash 默认按官方人民币高峰价保守估算，空闲时段实际为一半；自定义表会整体替换默认表。
+# AI_PRICE_TABLE_JSON={"deepseek-flash":{"inHit":0.04,"inMiss":2,"out":8}}
 
 
 # ─── CORS / iframe ───────────────────────────────────────────────────

apps/pac-service/src/cli/ai-gen-script.cli.ts

@@ -6,7 +6,7 @@
  *   pnpm ai:gen-script -- --pid=F1001                  # 用 patient externalId(自动取该 patient active plan)
  *   pnpm ai:gen-script -- --pid=F1001 --bust-cache     # 跳过缓存重新调
  *   pnpm ai:gen-script -- --pid=F1001 --dry-run        # 不写 PlanScript 表
- *   pnpm ai:gen-script -- --pid=F1001 --model=deepseek-v4-flash  # 切模型
+ *   pnpm ai:gen-script -- --pid=F1001 --model=deepseek-flash  # 切模型
  *
  * 输入完全确定(不重新 pull 患者数据),仅用于验证 prompt + model 改动效果。
  */
@@ -61,7 +61,7 @@ Usage:
   pnpm ai:gen-script -- --pid=<patient-external-id>
   pnpm ai:gen-script -- --pid=F1001 --bust-cache       # 跳过缓存,真实调 LLM
   pnpm ai:gen-script -- --pid=F1001 --dry-run          # 不写 PlanScript
-  pnpm ai:gen-script -- --pid=F1001 --model=deepseek-v4-flash
+  pnpm ai:gen-script -- --pid=F1001 --model=deepseek-flash
 
 Options:
   --plan-id=<uuid>    直接指定 plan id
@@ -70,7 +70,7 @@ Options:
   --tenant=<str>      默认 t_demo
   --bust-cache        强制重跑(不走 prompt cache)
   --dry-run           只生成不写库
-  --model=<id>        覆盖默认模型(deepseek-v4-pro / deepseek-v4-flash)
+  --model=<id>        覆盖默认模型(deepseek-v4-pro / deepseek-flash)
 `);
 }
 

apps/pac-service/src/config/ai-models.ts

+/** 官方保留的旧 Flash 别名；统一请求名和调用记录，不改历史数据。 */
+export function canonicalModelId(modelId: string): string {
+  return modelId === 'deepseek-v4-flash' || modelId === 'deepseek-v4-flash-vision-exp'
+    ? 'deepseek-flash'
+    : modelId;
+}

apps/pac-service/src/config/configuration.ts

+import { canonicalModelId } from './ai-models';
+
 export interface AppConfig {
   nodeEnv: 'development' | 'test' | 'production';
   port: number;
@@ -15,7 +17,7 @@ export interface AppConfig {
     deepseekApiKey: string;
     /// DeepSeek base URL(默认官方;私有 / 镜像 endpoint 时覆盖)
     deepseekBaseUrl: string;
-    /// 默认主力模型(deepseek-v4-pro / deepseek-v4-flash)
+    /// 默认主力模型(deepseek-v4-pro / deepseek-flash)
     defaultModel: string;
     /// Gemini API key(Google AI Studio;海外需可达 generativelanguage.googleapis.com)
     geminiApiKey: string;
@@ -75,7 +77,7 @@ export function loadConfig(): AppConfig {
     ai: {
       deepseekApiKey: process.env.DEEPSEEK_API_KEY ?? '',
       deepseekBaseUrl: process.env.DEEPSEEK_BASE_URL ?? 'https://api.deepseek.com',
-      defaultModel: process.env.AI_DEFAULT_MODEL ?? 'deepseek-v4-flash',
+      defaultModel: canonicalModelId(process.env.AI_DEFAULT_MODEL ?? 'deepseek-flash'),
       geminiApiKey: process.env.GEMINI_API_KEY ?? '',
       geminiBaseUrl: process.env.GEMINI_BASE_URL ?? '',
       geminiDefaultModel: process.env.GEMINI_DEFAULT_MODEL ?? 'gemini-3.5-flash',
@@ -108,8 +110,8 @@ export function loadConfig(): AppConfig {
 /**
  * AI 价格表 — 从 env JSON 读,失败兜底默认值。
  *
- * 单位:¥/M tokens(已按汇率换算,DeepSeek 公开价 USD * 7.2)。
- *   inHit:  输入命中 prompt cache 的部分(便宜 ~7x)
+ * 单位:¥/M tokens。DeepSeek Flash 使用官方人民币高峰价作保守估算，非实际账单。
+ *   inHit:  输入命中 prompt cache 的部分
  *   inMiss: 输入未命中(全价)
  *   out:    输出(含 reasoning,DeepSeek 统一计费)
  *
@@ -120,14 +122,16 @@ export function loadConfig(): AppConfig {
  *   4. 新调用按新价计
  */
 function parsePriceTable(raw: string | undefined): Record<string, { inHit: number; inMiss: number; out: number }> {
-  // ⚠️ 价格核对于 2026-06(官方页):¥/M tokens。美元价按 ×7.2 折 RMB。
+  // 除下方单独更新的 Flash 外，其他价格沿用 2026-06 配置：¥/M tokens，美元价按 ×7.2 折 RMB。
   //   用各家**标准价**(不含限时促销,促销到期不会少报);要按实付/促销算 → 用 AI_PRICE_TABLE_JSON 覆盖。
-  //   DeepSeek:api-docs.deepseek.com/quick_start/pricing(V4-pro $0.003625/0.435/0.87、V4-flash $0.0028/0.14/0.28)
+  //   Flash 于 2026-10-23 单独核对官方人民币价：高峰 0.04/2/8，空闲为一半。
+  //   https://api-docs.deepseek.com/zh-cn/quick_start/pricing/
+  //   仍用固定单价保守估算，不模拟节假日峰谷账单；其他模型价格保持原配置。
   //   Gemini 3.5 Flash:$0.15/1.5/9.0(cache/in/out)
   //   Qwen3.7-Max(百炼,RMB 直接):缓存输入 1.2 / 输入 12 / 输出 36;现有限时 5 折 = 6/18(促销价走 env)
   const DEFAULT = {
     'deepseek-v4-pro':   { inHit: 0.03, inMiss: 3.13, out: 6.26 },
-    'deepseek-v4-flash': { inHit: 0.02, inMiss: 1.01, out: 2.02 },
+    'deepseek-flash':   { inHit: 0.04, inMiss: 2, out: 8 },
     'gemini-3.5-flash':  { inHit: 1.08, inMiss: 10.8, out: 64.8 },
     'gemini-2.5-flash':  { inHit: 1.08, inMiss: 10.8, out: 64.8 },
     // Qwen3.7-Max 是旗舰最贵档(不是便宜模型):标准价 ¥/M 直接填

apps/pac-service/src/modules/ai/calls/draft-plan-script/tiers/deep/calls.ts

@@ -56,7 +56,7 @@ export class DeepPlanCall implements AiCall<ScriptContext, DeepPlanZ> {
   readonly kind = 'script' as const;
   readonly callKey = 'draft_plan_script_plan';
   readonly promptVersion = 'draft_plan_script@2026-06-16-deep-plan-v9'; // v9: 去掉钦定段式(给自由度),只留从果推因/层层递进原则;v8: 从果推因+后果不吓唬+3-6段
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = DeepPlanSchema;
   buildPrompt(ctx: ScriptContext) {
     return { system: PLAN_SYSTEM, prompt: buildPlanPrompt(ctx) };
@@ -68,7 +68,7 @@ export class DeepWriteCall implements AiCall<DeepWriteInput, DeepWriteZ> {
   readonly kind = 'script' as const;
   readonly callKey = 'draft_plan_script_write';
   readonly promptVersion = 'draft_plan_script@2026-07-29-deep-write-v19'; // v19: 自报家门改「我是{诊断医生}医生的助理X」(业务口径:非客服身份)。v18: 低龄约束(<18 不主推种植)+ 替牙期(≤12)说「早期矫治」。v17: 高龄硬约束(≥75 缺牙禁推种植/手术)。v16: 自报家门去人名 → 占位【回访客服】(渲染期按登录人回填,修串名 bug)。v15: 去掉钦定段式(给自由度);v14: 禁内部代码/术语(K08等)说大白话
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = DeepWriteSchema;
   constructor(private readonly skillRegistry: DraftPlanScriptSkillRegistry) {}
   buildPrompt(input: DeepWriteInput) {
@@ -86,7 +86,7 @@ export class DeepVerifyCall implements AiCall<DeepVerifyInput, DeepVerifyZ> {
   readonly kind = 'judge' as const;
   readonly callKey = 'draft_plan_script_verify';
   readonly promptVersion = 'draft_plan_script@2026-06-16-deep-verify-v10'; // v10: 递进检查去钦定段式(给自由度);v9: 加④患者听得懂(无代码/术语/含糊)
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = DeepVerifySchema;
   buildPrompt(input: DeepVerifyInput) {
     return { system: VERIFY_SYSTEM, prompt: buildVerifyPrompt(input) };

apps/pac-service/src/modules/ai/calls/draft-plan-script/tiers/stable/stable.call.ts

@@ -77,7 +77,7 @@ export class DraftPlanScriptCall
   readonly kind = 'script' as const;
   readonly callKey = 'draft_plan_script';
   readonly promptVersion = DRAFT_PLAN_SCRIPT_PROMPT_VERSION;
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = DraftPlanScriptSchema;
   readonly safetyRules = safetyRules;
 

apps/pac-service/src/modules/ai/calls/draft-plan-script/tiers/standard/standard.call.ts

@@ -30,7 +30,7 @@ export class StandardScriptCall implements AiCall<DraftPlanScriptInput, DeepDraf
   readonly kind = 'script' as const;
   readonly callKey = 'draft_plan_script';
   readonly promptVersion = STANDARD_PROMPT_VERSION;
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = StandardScriptSchema;
   // 显式标 SafetyRule<DeepDraft>,让 runner.run 的 TOutput 推成 DeepDraft(否则会被结构型窄化丢 tone)
   readonly safetyRules: ReadonlyArray<SafetyRule<DeepDraft>> = SCRIPT_SAFETY_RULES_SECTIONS;

apps/pac-service/src/modules/ai/calls/draft-wecom-script/calls.ts

@@ -52,7 +52,7 @@ export class WecomWriteCall implements AiCall<WecomWriteInput, WecomWriteZ> {
   readonly callKey = 'draft_wecom_script_write';
   // ⚠️ 换 promptVersion 会让缓存整体失效(这是想要的):流程和形态都变了,旧缓存不该再命中
   readonly promptVersion = 'draft_wecom_script@2026-08-05-single-v2';
-  readonly defaultModelId = 'deepseek-v4-flash';
+  readonly defaultModelId = 'deepseek-flash';
   readonly outputSchema = WecomWriteSchema;
   constructor(private readonly skillRegistry: DraftPlanScriptSkillRegistry) {}
   buildPrompt(input: WecomWriteInput) {

apps/pac-service/src/modules/ai/core/ai-provider.service.ts

@@ -5,6 +5,7 @@ import { createGoogleGenerativeAI, type GoogleGenerativeAIProvider } from '@ai-s
 import { createOpenAICompatible, type OpenAICompatibleProvider } from '@ai-sdk/openai-compatible';
 import type { LanguageModel } from 'ai';
 import type { AppConfig } from '../../../config/configuration';
+import { canonicalModelId } from '../../../config/ai-models';
 
 /**
  * AiProviderService — 模型工厂。
@@ -31,7 +32,7 @@ export class AiProviderService {
 
   constructor(private readonly config: ConfigService<AppConfig, true>) {
     const ai = this.config.get('ai', { infer: true });
-    this.defaultModel = ai.defaultModel;
+    this.defaultModel = canonicalModelId(ai.defaultModel);
     this.geminiDefaultModel = ai.geminiDefaultModel;
     this.qwenDefaultModel = ai.qwenDefaultModel;
 
@@ -105,14 +106,14 @@ export class AiProviderService {
   /**
    * 解析 modelId → vendor 实例 + 规范化具体 modelId(落账 modelName / 计费 priceTable 用规范化后的)。
    * 命名约定:
-   *   "deepseek-v4-pro" / "deepseek-v4-flash" / "deepseek"(裸键)→ deepseek
+   *   "deepseek-v4-pro" / "deepseek-flash" / "deepseek"(裸键)→ deepseek
    *   "gemini-2.5-flash" / "gemini"(裸键)                      → gemini
    *   "qwen-max"(将来)                                         → qwen
-   * 裸键(无 "-" 版本)→ 解析到 config 配的默认型号;带版本 → 原样直传 SDK。
+   * 裸键(无 "-" 版本)→ 解析到 config 配的默认型号;具体型号 → 直传 SDK（旧 Flash 别名归一为 deepseek-flash）。
    */
   resolve(modelId: string): { model: LanguageModel; provider: string; modelId: string } {
     if (modelId === 'deepseek' || modelId.startsWith('deepseek-')) {
-      const canonical = modelId === 'deepseek' ? this.defaultModel : modelId;
+      const canonical = canonicalModelId(modelId === 'deepseek' ? this.defaultModel : modelId);
       return { model: this.deepseek(canonical), provider: 'deepseek', modelId: canonical };
     }
     if (modelId === 'gemini' || modelId.startsWith('gemini-')) {

apps/pac-service/src/modules/ai/core/cost.ts

+import { canonicalModelId } from '../../../config/ai-models';
+
 /**
  * 成本估算 —— 纯函数，`AiCallRunner` 与助手共用**同一份**口径。
  *
@@ -30,7 +32,11 @@ export function estimateCostYuan(
   completionTokens: number,
   cachedInputTokens = 0,
 ): CostResult {
-  const p = priceTable[modelId];
+  // 新旧 Flash 名使用同一计价；兼容部署时仍按旧名称配置的自定义价目表。
+  const canonical = canonicalModelId(modelId);
+  const p = priceTable[canonical] ?? priceTable[modelId] ?? (canonical === 'deepseek-flash'
+    ? priceTable['deepseek-v4-flash'] ?? priceTable['deepseek-v4-flash-vision-exp']
+    : undefined);
   const price = p ?? priceTable['deepseek-v4-pro'] ?? FALLBACK_PRICE;
   // 防御:cached > prompt 不该发生,clamp
   const hit = Math.min(Math.max(0, cachedInputTokens), Math.max(0, promptTokens));

apps/pac-service/src/modules/ai/orchestrators/plan-script.orchestrator.ts

@@ -270,7 +270,7 @@ export class PlanScriptOrchestrator {
       yield {
         type: 'start',
         invocationId: '',
-        modelId: options.modelIdOverride ?? 'deepseek-v4-flash',
+        modelId: options.modelIdOverride ?? 'deepseek-flash',
         promptVersion: 'draft_plan_script@deep-pipeline',
       };
       let r: DeepScriptResult;

apps/pac-service/tests/deepseek-model-config.spec.ts

+import { loadConfig } from '../src/config/configuration';
+import { AiProviderService } from '../src/modules/ai/core/ai-provider.service';
+import { estimateCostYuan } from '../src/modules/ai/core/cost';
+import { ConfigService } from '@nestjs/config';
+import type { AppConfig } from '../src/config/configuration';
+
+jest.mock('@ai-sdk/deepseek', () => ({ createDeepSeek: () => (id: string) => ({ modelId: id }) }));
+jest.mock('@ai-sdk/google', () => ({ createGoogleGenerativeAI: () => (id: string) => ({ modelId: id }) }));
+jest.mock('@ai-sdk/openai-compatible', () => ({ createOpenAICompatible: () => (id: string) => ({ modelId: id }) }));
+
+describe('DeepSeek Flash 型号与计价迁移', () => {
+  const originalEnv = process.env;
+  beforeEach(() => {
+    process.env = {
+      DATABASE_URL: 'postgresql://unused', REDIS_URL: 'redis://unused',
+      JWT_SECRET: 'test-only', JWT_REFRESH_SECRET: 'test-only',
+      DEEPSEEK_API_KEY: 'test-only', GEMINI_API_KEY: 'test-only', QWEN_API_KEY: 'test-only',
+    };
+  });
+  afterEach(() => { process.env = originalEnv; });
+
+  function provider() {
+    const config = loadConfig();
+    return new AiProviderService(new ConfigService<AppConfig, true>(config));
+  }
+
+  it('默认配置、SDK 请求和落账型号一致', () => {
+    expect(loadConfig().ai.defaultModel).toBe('deepseek-flash');
+    const resolved = provider().resolve('deepseek');
+    expect(resolved.modelId).toBe('deepseek-flash');
+    expect(resolved.model).toMatchObject({ modelId: 'deepseek-flash' });
+    expect(resolved.provider).toBe('deepseek');
+  });
+
+  it.each(['deepseek-v4-flash', 'deepseek-v4-flash-vision-exp'])('兼容环境变量和显式旧别名 %s', alias => {
+    process.env.AI_DEFAULT_MODEL = alias;
+    expect(loadConfig().ai.defaultModel).toBe('deepseek-flash');
+    expect(provider().getDefaultModelId()).toBe('deepseek-flash');
+    expect(provider().resolve(alias).model).toMatchObject({ modelId: 'deepseek-flash' });
+  });
+
+  it('保留显式 Pro 配置与其他供应商路由', () => {
+    process.env.AI_DEFAULT_MODEL = 'deepseek-v4-pro';
+    expect(provider().resolve('deepseek').modelId).toBe('deepseek-v4-pro');
+    expect(provider().resolve('qwen').provider).toBe('qwen');
+    expect(provider().resolve('gemini').provider).toBe('gemini');
+  });
+
+  it.each(['deepseek-flash', 'deepseek-v4-flash', 'deepseek-v4-flash-vision-exp'])('%s 按 Flash 价估算，不误落 Pro 兜底', model => {
+    const table = loadConfig().ai.priceTable;
+    expect(estimateCostYuan(table, model, 1_000_000, 1_000_000, 500_000))
+      .toEqual({ yuan: 9.02, priceMissing: false });
+  });
+
+  it('沿用旧名称的自定义价格，不悄悄替换用户配置', () => {
+    process.env.AI_PRICE_TABLE_JSON = JSON.stringify({ 'deepseek-v4-flash': { inHit: 1, inMiss: 2, out: 3 } });
+    expect(estimateCostYuan(loadConfig().ai.priceTable, 'deepseek-flash', 1_000_000, 1_000_000))
+      .toEqual({ yuan: 5, priceMissing: false });
+  });
+
+  it('自定义价格同时有新旧名称时，以新名称为准', () => {
+    const table = {
+      'deepseek-flash': { inHit: 1, inMiss: 2, out: 3 },
+      'deepseek-v4-flash': { inHit: 7, inMiss: 8, out: 9 },
+    };
+    expect(estimateCostYuan(table, 'deepseek-v4-flash', 1_000_000, 1_000_000))
+      .toEqual({ yuan: 5, priceMissing: false });
+  });
+});