feat(rvw): harden json parsing and finalize 0316 rollout

Stabilize RVW editorial and methodology JSON parsing in production with layered repair and fallback handling, then publish the paired frontend task-level language selector updates. Also reset deployment checklist, record the 0316 deployment summary, and refresh the SAE runtime status with latest backend/frontend IPs. Made-with: Cursor
2026-03-16 00:24:33 +08:00
parent 707f783229
commit c3554fd61d
17 changed files with 704 additions and 239 deletions
--- a/backend/src/modules/admin/services/tenantService.ts
+++ b/backend/src/modules/admin/services/tenantService.ts
@@ -385,6 +385,7 @@ class TenantService {
    backgroundImageUrl?: string;
    modules: string[];
    isReviewOnly: boolean;
+    editorialDefaultStandard: 'zh' | 'en';
  } | null> {
    // 根据 code 查找租户
    const tenant = await prisma.tenants.findUnique({
@@ -414,6 +415,7 @@ class TenantService {
      backgroundImageUrl: tenant.login_background_url || undefined,
      modules,
      isReviewOnly,
+      editorialDefaultStandard: tenant.journal_language === 'ZH' ? 'zh' : 'en',
    };
  }
 }
--- a/backend/src/modules/rvw/controllers/reviewController.ts
+++ b/backend/src/modules/rvw/controllers/reviewController.ts
@@ -190,19 +190,28 @@ export async function createTask(
 export async function runReview(
  request: FastifyRequest<{
    Params: { taskId: string };
-    Body: { agents: AgentType[] };
+    Body: { agents: AgentType[]; editorialBaseStandard?: 'zh' | 'en' };
  }>,
  reply: FastifyReply
 ) {
  try {
    const userId = getUserId(request);
    const { taskId } = request.params;
-    const { agents } = request.body;
+    const { agents, editorialBaseStandard } = request.body;
+    const normalizedEditorialBaseStandard =
+      editorialBaseStandard === 'zh' || editorialBaseStandard === 'en'
+        ? editorialBaseStandard
+        : undefined;

-    logger.info('[RVW:Controller] 运行审查', { taskId, agents });
+    logger.info('[RVW:Controller] 运行审查', { taskId, agents, editorialBaseStandard: normalizedEditorialBaseStandard });

    // ✅ 返回 jobId（Platform-Only架构）
-    const { jobId } = await reviewService.runReview({ taskId, agents, userId });
+    const { jobId } = await reviewService.runReview({
+      taskId,
+      agents,
+      editorialBaseStandard: normalizedEditorialBaseStandard,
+      userId,
+    });

    return reply.send({
      success: true,
@@ -232,17 +241,31 @@ export async function batchRunReview(
    Body: {
      taskIds: string[];
      agents: AgentType[];
+      editorialBaseStandard?: 'zh' | 'en';
    };
  }>,
  reply: FastifyReply
 ) {
  try {
    const userId = getUserId(request);
-    const { taskIds, agents } = request.body;
+    const { taskIds, agents, editorialBaseStandard } = request.body;
+    const normalizedEditorialBaseStandard =
+      editorialBaseStandard === 'zh' || editorialBaseStandard === 'en'
+        ? editorialBaseStandard
+        : undefined;

-    logger.info('[RVW:Controller] 批量运行审查', { taskCount: taskIds.length, agents });
+    logger.info('[RVW:Controller] 批量运行审查', {
+      taskCount: taskIds.length,
+      agents,
+      editorialBaseStandard: normalizedEditorialBaseStandard,
+    });

-    const result = await reviewService.batchRunReview({ taskIds, agents, userId });
+    const result = await reviewService.batchRunReview({
+      taskIds,
+      agents,
+      editorialBaseStandard: normalizedEditorialBaseStandard,
+      userId,
+    });

    return reply.send({
      success: true,
--- a/backend/src/modules/rvw/services/editorialService.ts
+++ b/backend/src/modules/rvw/services/editorialService.ts
@@ -16,6 +16,8 @@ import { EditorialReview } from '../types/index.js';
 import { parseJSONFromLLMResponse } from './utils.js';
 import { composeRvwSystemPrompt, getRvwProtocol, sanitizeRvwBusinessPrompt } from './promptProtocols.js';

+const EDITORIAL_FALLBACK_MODEL: ModelType = 'qwen3-72b';
+
 function isValidEditorialReview(result: unknown): result is EditorialReview {
  if (!result || typeof result !== 'object') return false;
  const data = result as Record<string, unknown>;
@@ -25,6 +27,14 @@ function isValidEditorialReview(result: unknown): result is EditorialReview {
  return true;
 }

+function buildContentObservability(content: string): { length: number; preview: string } {
+  const compact = (content || '').replace(/\s+/g, ' ').trim();
+  return {
+    length: content?.length ?? 0,
+    preview: compact.slice(0, 500),
+  };
+}
+
 async function repairEditorialToJson(
  rawContent: string,
  modelType: ModelType
@@ -48,11 +58,44 @@ async function repairEditorialToJson(
  });

  const repairedContent = repaired.content ?? '';
-  const parsed = parseJSONFromLLMResponse<EditorialReview>(repairedContent);
-  if (!isValidEditorialReview(parsed)) {
-    throw new Error('稿约规范性评估结果结构化修复失败（JSON字段不完整）');
+  try {
+    const parsed = parseJSONFromLLMResponse<EditorialReview>(repairedContent);
+    if (!isValidEditorialReview(parsed)) {
+      throw new Error('稿约规范性评估结果结构化修复失败（JSON字段不完整）');
+    }
+    return parsed;
+  } catch {
+    // 二次兜底：使用更严格、更短的指令，强制输出单一 JSON 对象
+    logger.warn('[RVW:Editorial] 第一次结构化修复失败，进入超严格兜底修复');
+    const strictMessages = [
+      {
+        role: 'system' as const,
+        content:
+          'You are a JSON formatter. Output ONE valid JSON object only. No markdown, no prose, no code fence.\n' +
+          '{' +
+          '"overall_score":0,' +
+          '"summary":"",' +
+          '"items":[{"criterion":"","status":"pass","score":0,"issues":[],"suggestions":[]}]' +
+          '}\n' +
+          'Rules: status must be pass|warning|fail; scores must be numbers 0-100; issues/suggestions must be arrays.',
+      },
+      {
+        role: 'user' as const,
+        content: `Convert the following editorial review text into that JSON schema:\n\n${rawContent}\n\nReturn JSON object only.`,
+      },
+    ];
+
+    const strictRepaired = await llmAdapter.chat(strictMessages, {
+      temperature: 0,
+      maxTokens: 2500,
+    });
+    const strictContent = strictRepaired.content ?? '';
+    const strictParsed = parseJSONFromLLMResponse<EditorialReview>(strictContent);
+    if (!isValidEditorialReview(strictParsed)) {
+      throw new Error('稿约规范性评估结果结构化修复失败（JSON字段不完整）');
+    }
+    return strictParsed;
  }
-  return parsed;
 }

 /**
@@ -94,38 +137,58 @@ export async function reviewEditorialStandards(
      logger.info('[RVW:Editorial] 使用 DRAFT 版本 Prompt（调试模式）', { userId });
    }

-    // 2. 构建消息
-    const messages = [
-      { role: 'system' as const, content: composeRvwSystemPrompt('editorial', businessPrompt) },
-      { role: 'user' as const, content: `请对以下稿件进行稿约规范性评估。\n\n稿件内容如下：\n${text}` },
-    ];
+    // 2. 按模型序列尝试（主模型 -> 备用模型）
+    const candidateModels = Array.from(new Set([modelType, EDITORIAL_FALLBACK_MODEL]));
+    let lastError: Error | null = null;

-    // 3. 调用LLM
-    logger.info('[RVW:Editorial] 开始稿约规范性评估', { modelType });
-    const llmAdapter = LLMFactory.getAdapter(modelType);
-    const response = await llmAdapter.chat(messages, {
-      temperature: 0.3, // 较低温度以获得更稳定的评估
-      maxTokens: 8000,  // 确保完整输出
-    });
-    const editContent = response.content ?? '';
-    logger.info('[RVW:Editorial] 评估完成', { 
-      modelType, 
-      responseLength: editContent.length 
-    });
+    for (const candidateModel of candidateModels) {
+      try {
+        const messages = [
+          { role: 'system' as const, content: composeRvwSystemPrompt('editorial', businessPrompt) },
+          { role: 'user' as const, content: `请对以下稿件进行稿约规范性评估。\n\n稿件内容如下：\n${text}` },
+        ];

-    // 4. 解析 JSON（失败则自动进入结构化修复）
-    try {
-      const result = parseJSONFromLLMResponse<EditorialReview>(editContent);
-      if (!isValidEditorialReview(result)) {
-        throw new Error('LLM返回的数据格式不正确');
+        logger.info('[RVW:Editorial] 开始稿约规范性评估', {
+          modelType: candidateModel,
+          fallback: candidateModel !== modelType,
+        });
+        const llmAdapter = LLMFactory.getAdapter(candidateModel);
+        const response = await llmAdapter.chat(messages, {
+          temperature: 0.2,
+          maxTokens: 8000,
+        });
+        const editContent = response.content ?? '';
+        logger.info('[RVW:Editorial] 评估完成', {
+          modelType: candidateModel,
+          responseLength: editContent.length,
+        });
+
+        try {
+          const result = parseJSONFromLLMResponse<EditorialReview>(editContent);
+          if (!isValidEditorialReview(result)) {
+            throw new Error('LLM返回的数据格式不正确');
+          }
+          return result;
+        } catch (parseError) {
+          logger.warn('[RVW:Editorial] 原始响应解析失败，进入修复流程', {
+            modelType: candidateModel,
+            reason: parseError instanceof Error ? parseError.message : 'Unknown parse error',
+            ...buildContentObservability(editContent),
+          });
+          return await repairEditorialToJson(editContent, candidateModel);
+        }
+      } catch (candidateError) {
+        const err = candidateError instanceof Error ? candidateError : new Error(String(candidateError));
+        lastError = err;
+        logger.warn('[RVW:Editorial] 模型评估失败，尝试下一候选模型', {
+          modelType: candidateModel,
+          fallback: candidateModel !== modelType,
+          error: err.message,
+        });
      }
-      return result;
-    } catch (parseError) {
-      logger.warn('[RVW:Editorial] 原始响应解析失败，进入修复流程', {
-        reason: parseError instanceof Error ? parseError.message : 'Unknown parse error',
-      });
-      return await repairEditorialToJson(editContent, modelType);
    }
+
+    throw lastError ?? new Error('稿约规范性评估失败（所有候选模型均失败）');
  } catch (error) {
    logger.error('[RVW:Editorial] 稿约规范性评估失败', {
      error: error instanceof Error ? error.message : 'Unknown error',
--- a/backend/src/modules/rvw/services/methodologyService.ts
+++ b/backend/src/modules/rvw/services/methodologyService.ts
@@ -17,6 +17,7 @@ import { MethodologyCheckpoint, MethodologyIssue, MethodologyPart, MethodologyRe
 import { parseJSONFromLLMResponse } from './utils.js';
 import { composeRvwSystemPrompt, getRvwProtocol, sanitizeRvwBusinessPrompt } from './promptProtocols.js';

+const METHODOLOGY_FALLBACK_MODEL: ModelType = 'qwen3-72b';
 const METHODOLOGY_CONCLUSIONS = ['直接接收', '小修', '大修', '拒稿'] as const;
 type MethodologyConclusion = typeof METHODOLOGY_CONCLUSIONS[number];
 const METHODOLOGY_CHECKPOINT_ITEMS = [
@@ -228,6 +229,14 @@ function isValidMethodologyReview(result: unknown): result is MethodologyReview
  return true;
 }

+function buildContentObservability(content: string): { length: number; preview: string } {
+  const compact = (content || '').replace(/\s+/g, ' ').trim();
+  return {
+    length: content?.length ?? 0,
+    preview: compact.slice(0, 500),
+  };
+}
+
 function normalizeMethodologyCheckpoints(input: unknown): MethodologyCheckpoint[] {
  const normalizedMap = new Map<number, MethodologyCheckpoint>();
  if (Array.isArray(input)) {
@@ -319,48 +328,93 @@ function aggregateMethodologySections(sections: SectionReviewResult[]): Methodol
 }

 async function reviewMethodologySection(
-  llmAdapter: ReturnType<typeof LLMFactory.getAdapter>,
+  modelType: ModelType,
  businessPrompt: string,
  text: string,
  section: MethodologySectionDef
 ): Promise<SectionReviewResult> {
-  const messages = [
-    { role: 'system' as const, content: `${businessPrompt}\n\n${buildSectionProtocol(section)}` },
-    { role: 'user' as const, content: `请仅评估“${section.part}”（检查点 ${section.start}-${section.end}），并按协议返回 JSON。\n\n稿件内容如下：\n${text}` },
-  ];
-  const response = await llmAdapter.chat(messages, {
-    temperature: 0.2,
-    maxTokens: 2800,
-  });
-  const content = response.content ?? '';
-  try {
-    const parsed = parseJSONFromLLMResponse<SectionReviewResult>(content);
-    if (!isValidSectionReview(parsed, section)) {
-      throw new Error('section json invalid');
+  const candidateModels = Array.from(new Set([modelType, METHODOLOGY_FALLBACK_MODEL]));
+  let lastError: Error | null = null;
+
+  for (const candidateModel of candidateModels) {
+    try {
+      const llmAdapter = LLMFactory.getAdapter(candidateModel);
+      const messages = [
+        { role: 'system' as const, content: `${businessPrompt}\n\n${buildSectionProtocol(section)}` },
+        { role: 'user' as const, content: `请仅评估“${section.part}”（检查点 ${section.start}-${section.end}），并按协议返回 JSON。\n\n稿件内容如下：\n${text}` },
+      ];
+      const response = await llmAdapter.chat(messages, {
+        temperature: 0.2,
+        maxTokens: 2800,
+      });
+      const content = response.content ?? '';
+      try {
+        const parsed = parseJSONFromLLMResponse<SectionReviewResult>(content);
+        if (!isValidSectionReview(parsed, section)) {
+          throw new Error('section json invalid');
+        }
+        return normalizeSectionReview(parsed, section);
+      } catch {
+        const repairMessages = [
+          {
+            role: 'system' as const,
+            content: `你是 JSON 结构化助手。把输入文本转成目标 JSON。\n\n${buildSectionProtocol(section)}`,
+          },
+          {
+            role: 'user' as const,
+            content: `请将以下方法学评估文本重组为目标 JSON（仅检查点 ${section.start}-${section.end}）：\n\n${content}`,
+          },
+        ];
+        const repaired = await llmAdapter.chat(repairMessages, {
+          temperature: 0.1,
+          maxTokens: 1800,
+        });
+        const repairedContent = repaired.content ?? '';
+        try {
+          const repairedParsed = parseJSONFromLLMResponse<SectionReviewResult>(repairedContent);
+          if (!isValidSectionReview(repairedParsed, section)) {
+            throw new Error('section repair invalid');
+          }
+          return normalizeSectionReview(repairedParsed, section);
+        } catch {
+          // 二次兜底：严格 JSON 输出
+          const strictMessages = [
+            {
+              role: 'system' as const,
+              content:
+                'You are a JSON formatter. Output ONE valid JSON object only. No markdown, no prose.\n' +
+                `Must include: part="${section.part}", score(0-100), issues[], checkpoints[] for ids ${section.start}-${section.end}.`,
+            },
+            {
+              role: 'user' as const,
+              content: `Convert the following section review text into strict JSON:\n\n${content}`,
+            },
+          ];
+          const strictRepaired = await llmAdapter.chat(strictMessages, {
+            temperature: 0,
+            maxTokens: 1500,
+          });
+          const strictContent = strictRepaired.content ?? '';
+          const strictParsed = parseJSONFromLLMResponse<SectionReviewResult>(strictContent);
+          if (!isValidSectionReview(strictParsed, section)) {
+            throw new Error('section strict repair invalid');
+          }
+          return normalizeSectionReview(strictParsed, section);
+        }
+      }
+    } catch (error) {
+      const err = error instanceof Error ? error : new Error(String(error));
+      lastError = err;
+      logger.warn('[RVW:Methodology] 分段评估模型失败，尝试下一候选模型', {
+        section: section.part,
+        modelType: candidateModel,
+        fallback: candidateModel !== modelType,
+        error: err.message,
+      });
    }
-    return normalizeSectionReview(parsed, section);
-  } catch {
-    const repairMessages = [
-      {
-        role: 'system' as const,
-        content: `你是 JSON 结构化助手。把输入文本转成目标 JSON。\n\n${buildSectionProtocol(section)}`,
-      },
-      {
-        role: 'user' as const,
-        content: `请将以下方法学评估文本重组为目标 JSON（仅检查点 ${section.start}-${section.end}）：\n\n${content}`,
-      },
-    ];
-    const repaired = await llmAdapter.chat(repairMessages, {
-      temperature: 0.1,
-      maxTokens: 1800,
-    });
-    const repairedContent = repaired.content ?? '';
-    const repairedParsed = parseJSONFromLLMResponse<SectionReviewResult>(repairedContent);
-    if (!isValidSectionReview(repairedParsed, section)) {
-      throw new Error('section repair invalid');
-    }
-    return normalizeSectionReview(repairedParsed, section);
  }
+
+  throw lastError ?? new Error('section evaluate failed');
 }

 async function reviewMethodologyLegacy(
@@ -368,23 +422,42 @@ async function reviewMethodologyLegacy(
  text: string,
  modelType: ModelType
 ): Promise<MethodologyReview> {
-  const llmAdapter = LLMFactory.getAdapter(modelType);
-  const messages = [
-    { role: 'system' as const, content: composeRvwSystemPrompt('methodology', businessPrompt) },
-    { role: 'user' as const, content: `请对以下稿件进行方法学评估。\n\n稿件内容如下：\n${text}` },
-  ];
-  const response = await llmAdapter.chat(messages, {
-    temperature: 0.3,
-    maxTokens: 5000,
-  });
-  const methContent = response.content ?? '';
-  try {
-    const result = parseJSONFromLLMResponse<MethodologyReview>(methContent);
-    if (!isValidMethodologyReview(result)) throw new Error('invalid json');
-    return normalizeMethodologyReview(result);
-  } catch {
-    return repairMethodologyToJson(methContent, modelType);
+  const candidateModels = Array.from(new Set([modelType, METHODOLOGY_FALLBACK_MODEL]));
+  let lastError: Error | null = null;
+  for (const candidateModel of candidateModels) {
+    try {
+      const llmAdapter = LLMFactory.getAdapter(candidateModel);
+      const messages = [
+        { role: 'system' as const, content: composeRvwSystemPrompt('methodology', businessPrompt) },
+        { role: 'user' as const, content: `请对以下稿件进行方法学评估。\n\n稿件内容如下：\n${text}` },
+      ];
+      const response = await llmAdapter.chat(messages, {
+        temperature: 0.2,
+        maxTokens: 5000,
+      });
+      const methContent = response.content ?? '';
+      try {
+        const result = parseJSONFromLLMResponse<MethodologyReview>(methContent);
+        if (!isValidMethodologyReview(result)) throw new Error('invalid json');
+        return normalizeMethodologyReview(result);
+      } catch {
+        logger.warn('[RVW:Methodology] legacy原始响应解析失败，进入修复流程', {
+          modelType: candidateModel,
+          ...buildContentObservability(methContent),
+        });
+        return repairMethodologyToJson(methContent, candidateModel);
+      }
+    } catch (error) {
+      const err = error instanceof Error ? error : new Error(String(error));
+      lastError = err;
+      logger.warn('[RVW:Methodology] legacy模型失败，尝试下一候选模型', {
+        modelType: candidateModel,
+        fallback: candidateModel !== modelType,
+        error: err.message,
+      });
+    }
  }
+  throw lastError ?? new Error('legacy evaluate failed');
 }

 async function repairMethodologyToJson(
@@ -411,13 +484,37 @@ async function repairMethodologyToJson(
  });

  const repairedContent = repaired.content ?? '';
-  const parsed = parseJSONFromLLMResponse<MethodologyReview>(repairedContent);
-
-  if (!isValidMethodologyReview(parsed)) {
-    throw new Error('方法学评估结果结构化修复失败（JSON字段不完整）');
+  try {
+    const parsed = parseJSONFromLLMResponse<MethodologyReview>(repairedContent);
+    if (!isValidMethodologyReview(parsed)) {
+      throw new Error('方法学评估结果结构化修复失败（JSON字段不完整）');
+    }
+    return normalizeMethodologyReview(parsed);
+  } catch {
+    logger.warn('[RVW:Methodology] 第一次结构化修复失败，进入超严格兜底修复');
+    const strictMessages = [
+      {
+        role: 'system' as const,
+        content:
+          'You are a JSON formatter. Output ONE valid JSON object only. No markdown, no prose.\n' +
+          '{"overall_score":0,"summary":"","conclusion":"小修","checkpoints":[{"id":1,"item":"","status":"pass","finding":"","suggestion":""}],"parts":[{"part":"科研设计评估","score":0,"issues":[]}]}',
+      },
+      {
+        role: 'user' as const,
+        content: `Convert the following methodology review text into strict JSON:\n\n${rawContent}`,
+      },
+    ];
+    const strictRepaired = await llmAdapter.chat(strictMessages, {
+      temperature: 0,
+      maxTokens: 2500,
+    });
+    const strictContent = strictRepaired.content ?? '';
+    const strictParsed = parseJSONFromLLMResponse<MethodologyReview>(strictContent);
+    if (!isValidMethodologyReview(strictParsed)) {
+      throw new Error('方法学评估结果结构化修复失败（JSON字段不完整）');
+    }
+    return normalizeMethodologyReview(strictParsed);
  }
-
-  return normalizeMethodologyReview(parsed);
 }

 /**
@@ -466,7 +563,6 @@ export async function reviewMethodology(
    }
    businessPrompt = sanitizeRvwBusinessPrompt('methodology', businessPrompt);

-    const llmAdapter = LLMFactory.getAdapter(modelType);
    logger.info('[RVW:Methodology] 开始分治并行评估', {
      modelType,
      sections: METHODOLOGY_SECTION_DEFS.map(section => `${section.part}(${section.start}-${section.end})`),
@@ -474,7 +570,7 @@ export async function reviewMethodology(

    const settled = await Promise.allSettled(
      METHODOLOGY_SECTION_DEFS.map(section =>
-        reviewMethodologySection(llmAdapter, businessPrompt, text, section)
+        reviewMethodologySection(modelType, businessPrompt, text, section)
      )
    );
    const sectionResults: SectionReviewResult[] = [];
--- a/backend/src/modules/rvw/services/reviewService.ts
+++ b/backend/src/modules/rvw/services/reviewService.ts
@@ -93,7 +93,9 @@ export async function createTask(
  logger.info('[RVW] 任务已创建', { taskId: task.id, status: task.status });

  // 2. 生成 OSS 存储 Key 并上传文件
-  const storageKey = generateRvwStorageKey(tenantId, userId, task.id, filename);
+  // 单租户历史数据 tenantId 可能为空，存储层使用固定命名空间兜底
+  const storageTenantId = tenantId ?? 'public';
+  const storageKey = generateRvwStorageKey(storageTenantId, userId, task.id, filename);
  let updatedTask = task;
  
  try {
@@ -181,7 +183,7 @@ async function extractDocumentAsync(taskId: string, file: Buffer, filename: stri
 * @returns jobId 供前端轮询状态
 */
 export async function runReview(params: RunReviewParams): Promise<{ jobId: string }> {
-  const { taskId, agents, userId } = params;
+  const { taskId, agents, editorialBaseStandard, userId } = params;

  // 验证智能体选择
  validateAgentSelection(agents);
@@ -236,6 +238,7 @@ export async function runReview(params: RunReviewParams): Promise<{ jobId: strin
    taskId,
    userId,
    agents,
+    editorialBaseStandard,
    extractedText: task.extractedText,
    modelType: (task.modelUsed || 'deepseek-v3') as ModelType,
    __expireInSeconds: 15 * 60, // 15min: 串行(5min)+并行(5min)+提取+余量，长文档可达8-10min
@@ -259,7 +262,7 @@ export async function batchRunReview(params: BatchRunParams): Promise<{
  success: string[]; 
  failed: { taskId: string; error: string }[] 
 }> {
-  const { taskIds, agents, userId } = params;
+  const { taskIds, agents, editorialBaseStandard, userId } = params;

  // 验证智能体选择
  validateAgentSelection(agents);
@@ -283,7 +286,7 @@ export async function batchRunReview(params: BatchRunParams): Promise<{
    const batch = taskIds.slice(i, i + MAX_CONCURRENT);
    
    const results = await Promise.allSettled(
-      batch.map(taskId => runReview({ taskId, agents, userId }))
+      batch.map(taskId => runReview({ taskId, agents, editorialBaseStandard, userId }))
    );

    results.forEach((result, index) => {
--- a/backend/src/modules/rvw/services/utils.ts
+++ b/backend/src/modules/rvw/services/utils.ts
@@ -6,71 +6,111 @@
 import { MethodologyReview, MethodologyStatus } from '../types/index.js';
 import { jsonrepair } from 'jsonrepair';

+function tryParseJsonCandidate<T>(candidate: string): T | null {
+  const normalized = candidate.trim().replace(/^\uFEFF/, '');
+  if (!normalized) return null;
+  try {
+    return JSON.parse(normalized) as T;
+  } catch {
+    try {
+      const repaired = jsonrepair(normalized);
+      return JSON.parse(repaired) as T;
+    } catch {
+      return null;
+    }
+  }
+}
+
+function extractBalancedJsonCandidates(content: string): string[] {
+  const text = content || '';
+  const candidates: string[] = [];
+  const stack: string[] = [];
+  let start = -1;
+  let inString = false;
+  let escaped = false;
+
+  for (let i = 0; i < text.length; i++) {
+    const ch = text[i];
+
+    if (inString) {
+      if (escaped) {
+        escaped = false;
+      } else if (ch === '\\') {
+        escaped = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+
+    if (ch === '"') {
+      inString = true;
+      continue;
+    }
+
+    if (ch === '{' || ch === '[') {
+      if (stack.length === 0) start = i;
+      stack.push(ch);
+      continue;
+    }
+
+    if (ch === '}' || ch === ']') {
+      if (stack.length === 0) continue;
+      const open = stack[stack.length - 1];
+      if ((open === '{' && ch === '}') || (open === '[' && ch === ']')) {
+        stack.pop();
+        if (stack.length === 0 && start >= 0) {
+          candidates.push(text.slice(start, i + 1));
+          start = -1;
+        }
+      } else {
+        // 栈失配时重置，继续寻找下一个合法片段
+        stack.length = 0;
+        start = -1;
+      }
+    }
+  }
+
+  return candidates;
+}
+
 /**
 * 从LLM响应中解析JSON
 * 支持多种格式：纯JSON、```json代码块、混合文本
 */
 export function parseJSONFromLLMResponse<T>(content: string): T {
-  try {
-    // 1. 尝试直接解析
-    return JSON.parse(content) as T;
-  } catch {
-    // 1.1 先尝试 jsonrepair（处理尾逗号、引号缺失等常见脏 JSON）
-    try {
-      const repaired = jsonrepair(content);
-      return JSON.parse(repaired) as T;
-    } catch {
-      // 继续后续提取策略
-    }
+  // 1) 直接解析 + jsonrepair
+  const direct = tryParseJsonCandidate<T>(content);
+  if (direct !== null) return direct;

-    // 2. 尝试提取```json代码块
-    const jsonMatch = content.match(/```json\s*\n?([\s\S]*?)\n?```/);
-    if (jsonMatch) {
-      try {
-        return JSON.parse(jsonMatch[1].trim()) as T;
-      } catch {
-        // 尝试修复代码块 JSON
-        try {
-          const repaired = jsonrepair(jsonMatch[1].trim());
-          return JSON.parse(repaired) as T;
-        } catch {
-          // 继续尝试其他方法
-        }
-      }
-    }
-
-    // 3. 尝试提取{}或[]包裹的内容
-    const objectMatch = content.match(/(\{[\s\S]*\})/);
-    if (objectMatch) {
-      try {
-        return JSON.parse(objectMatch[1]) as T;
-      } catch {
-        try {
-          const repaired = jsonrepair(objectMatch[1]);
-          return JSON.parse(repaired) as T;
-        } catch {
-          // 继续尝试其他方法
-        }
-      }
-    }
-
-    const arrayMatch = content.match(/(\[[\s\S]*\])/);
-    if (arrayMatch) {
-      try {
-        return JSON.parse(arrayMatch[1]) as T;
-      } catch {
-        try {
-          const repaired = jsonrepair(arrayMatch[1]);
-          return JSON.parse(repaired) as T;
-        } catch {
-          // 失败
-        }
-      }
-    }
-
-    // 4. 所有尝试都失败
-    throw new Error('无法从LLM响应中解析JSON');
+  // 2) 提取 Markdown 代码块（```json / ```）
+  const fenceRegex = /```(?:json)?\s*\n?([\s\S]*?)\n?```/gi;
+  for (const match of content.matchAll(fenceRegex)) {
+    const parsed = tryParseJsonCandidate<T>(match[1] || '');
+    if (parsed !== null) return parsed;
  }
+
+  // 3) 平衡括号提取，逐候选尝试
+  const balancedCandidates = extractBalancedJsonCandidates(content);
+  for (const candidate of balancedCandidates) {
+    const parsed = tryParseJsonCandidate<T>(candidate);
+    if (parsed !== null) return parsed;
+  }
+
+  // 4) 最后兜底：贪婪正则对象 / 数组（兼容极端场景）
+  const objectMatch = content.match(/(\{[\s\S]*\})/);
+  if (objectMatch) {
+    const parsed = tryParseJsonCandidate<T>(objectMatch[1]);
+    if (parsed !== null) return parsed;
+  }
+  const arrayMatch = content.match(/(\[[\s\S]*\])/);
+  if (arrayMatch) {
+    const parsed = tryParseJsonCandidate<T>(arrayMatch[1]);
+    if (parsed !== null) return parsed;
+  }
+
+  // 5) 所有尝试都失败
+  throw new Error('无法从LLM响应中解析JSON');
 }

 /**
--- a/backend/src/modules/rvw/types/index.ts
+++ b/backend/src/modules/rvw/types/index.ts
@@ -125,6 +125,7 @@ export interface ForensicsResult {
 export interface RunReviewParams {
  taskId: string;
  agents: AgentType[];  // 可选1个或2个
+  editorialBaseStandard?: 'zh' | 'en';
  userId: string;
 }

@@ -134,6 +135,7 @@ export interface RunReviewParams {
 export interface BatchRunParams {
  taskIds: string[];
  agents: AgentType[];
+  editorialBaseStandard?: 'zh' | 'en';
  userId: string;
 }

--- a/backend/src/modules/rvw/workers/reviewWorker.ts
+++ b/backend/src/modules/rvw/workers/reviewWorker.ts
@@ -50,6 +50,7 @@ interface ReviewJob {
  taskId: string;
  userId: string;
  agents: AgentType[];
+  editorialBaseStandard?: 'zh' | 'en';
  extractedText: string;
  modelType: ModelType;
 }
@@ -115,7 +116,7 @@ export async function registerReviewWorker() {

  // 注册审查Worker
  jobQueue.process<ReviewJob>('rvw_review_task', async (job: Job<ReviewJob>) => {
-    const { taskId, userId, agents, extractedText, modelType } = job.data;
+    const { taskId, userId, agents, editorialBaseStandard, extractedText, modelType } = job.data;
    const startTime = Date.now();

    logger.info('[reviewWorker] Processing review job', {
@@ -123,6 +124,7 @@ export async function registerReviewWorker() {
      taskId,
      userId,
      agents,
+      editorialBaseStandard,
      textLength: extractedText.length,
      useSkillsArchitecture: USE_SKILLS_ARCHITECTURE,
    });
@@ -188,6 +190,7 @@ export async function registerReviewWorker() {
          taskId,
          userId,
          agents,
+          editorialBaseStandard,
          extractedText,
          existingTask?.filePath || '',
          existingTask?.fileName || 'unknown.docx',
@@ -229,7 +232,13 @@ export async function registerReviewWorker() {
          logger.info('[reviewWorker] Running editorial review (legacy)', { taskId });
          console.log('  🔍 运行稿约规范性智能体...');
          
-          editorialResult = await reviewEditorialStandards(extractedText, modelType, userId);
+          editorialResult = await reviewEditorialStandards(
+            extractedText,
+            modelType,
+            userId,
+            undefined,
+            editorialBaseStandard
+          );
          
          logger.info('[reviewWorker] Editorial review completed', {
            taskId,
@@ -461,6 +470,7 @@ async function executeWithSkills(
  taskId: string,
  userId: string,
  agents: AgentType[],
+  taskEditorialBaseStandard: 'zh' | 'en' | undefined,
  extractedText: string,
  filePath: string,
  fileName: string,
@@ -524,28 +534,43 @@ async function executeWithSkills(
      },
    });
    if (cfg) {
+      const resolvedEditorialBaseStandard =
+        taskEditorialBaseStandard ??
+        (cfg.editorialBaseStandard === 'zh' ? 'zh' : inferredEditorialBase);
      tenantRvwConfig = {
        ...cfg,
-        editorialBaseStandard: cfg.editorialBaseStandard === 'zh' ? 'zh' : inferredEditorialBase,
+        editorialBaseStandard: resolvedEditorialBaseStandard,
      };
      logger.info('[reviewWorker] 已加载租户审稿配置', {
        taskId,
        tenantId: taskWithTenant.tenantId,
+        taskEditorialBaseStandard,
+        resolvedEditorialBaseStandard,
        hasMethodologyPrompt: !!cfg.methodologyExpertPrompt,
        hasEditorialPrompt: !!cfg.editorialExpertPrompt,
        hasDataPrompt: !!cfg.dataForensicsExpertPrompt,
        hasClinicalPrompt: !!cfg.clinicalExpertPrompt,
      });
    } else {
+      const resolvedEditorialBaseStandard = taskEditorialBaseStandard ?? inferredEditorialBase;
      tenantRvwConfig = {
-        editorialBaseStandard: inferredEditorialBase,
+        editorialBaseStandard: resolvedEditorialBaseStandard,
      };
      logger.info('[reviewWorker] 未找到租户审稿配置，回退系统基线', {
        taskId,
        tenantId: taskWithTenant.tenantId,
-        editorialBaseStandard: inferredEditorialBase,
+        taskEditorialBaseStandard,
+        editorialBaseStandard: resolvedEditorialBaseStandard,
      });
    }
+  } else if (taskEditorialBaseStandard) {
+    tenantRvwConfig = {
+      editorialBaseStandard: taskEditorialBaseStandard,
+    };
+    logger.info('[reviewWorker] 使用任务级稿约基线语言', {
+      taskId,
+      editorialBaseStandard: taskEditorialBaseStandard,
+    });
  }

  // 构建上下文（V4.0：注入租户配置实现 Hybrid Prompt）