diff --git a/packages/services/conversation-service/src/infrastructure/agents/coordinator/agent-loop.ts b/packages/services/conversation-service/src/infrastructure/agents/coordinator/agent-loop.ts
index 6690d80..6ad4479 100644
--- a/packages/services/conversation-service/src/infrastructure/agents/coordinator/agent-loop.ts
+++ b/packages/services/conversation-service/src/infrastructure/agents/coordinator/agent-loop.ts
@@ -31,10 +31,6 @@ import {
   isAgentInvocationTool,
   getToolsForClaudeAPI,
 } from '../tools/coordinator-tools';
-import {
-  MAX_FOLLOWUP_LENGTH,
-  smartTruncate,
-} from '../schemas/coordinator-response.schema';
 
 const logger = new Logger('AgentLoop');
 
@@ -180,7 +176,6 @@ export async function* agentLoop(
         messages: messages as any,
         tools: getToolsForClaudeAPI(additionalTools) as any,
         max_tokens: 2048,
-        ...(params.outputConfig ? { output_config: params.outputConfig } : {}),
       } as any);
       break; // success
     } catch (error: any) {
@@ -235,14 +230,11 @@ export async function* agentLoop(
 
         if (delta.type === 'text_delta') {
           currentTextContent += delta.text;
-          // Structured Output 模式下不直接 yield text（JSON 片段不能展示给用户）
-          if (!params.outputConfig) {
-            yield {
-              type: 'text',
-              content: delta.text,
-              timestamp: Date.now(),
-            };
-          }
+          yield {
+            type: 'text',
+            content: delta.text,
+            timestamp: Date.now(),
+          };
         } else if (delta.type === 'input_json_delta') {
           // Tool input being streamed — accumulate silently
         }
@@ -288,43 +280,6 @@ export async function* agentLoop(
     }
   } catch (error) {
     const errMsg = error instanceof Error ? error.message : String(error);
-
-    // ---- Structured Output 验证失败恢复 ----
-    // SDK 的 Zod 验证可能失败（如 intent 枚举值不匹配），
-    // 但模型的 answer 文本通常是正确的。从已累积的 blocks 中提取内容。
-    if (errMsg.includes('Failed to parse structured output') && assistantBlocks.length > 0) {
-      logger.warn(`Structured output validation failed, recovering from accumulated text`);
-      const accumulatedText = assistantBlocks
-        .filter(b => b.type === 'text' && 'text' in b)
-        .map(b => (b as any).text)
-        .join('');
-
-      if (accumulatedText) {
-        try {
-          // 手动 JSON.parse 跳过 Zod 验证 — intent 值不影响用户看到的内容
-          const parsed = JSON.parse(accumulatedText);
-          if (parsed.answer) {
-            yield { type: 'text', content: parsed.answer, timestamp: Date.now() };
-            if (parsed.followUp) {
-              if (/？|\?/.test(parsed.followUp)) {
-                yield { type: 'text', content: '\n\n' + parsed.followUp, timestamp: Date.now() };
-              } else {
-                yield { type: 'text', content: parsed.followUp, timestamp: Date.now() };
-              }
-            }
-            logger.debug(`[Turn ${currentTurn + 1}] Recovered intent: ${parsed.intent} (validation bypassed)`);
-          } else {
-            yield { type: 'text', content: accumulatedText, timestamp: Date.now() };
-          }
-        } catch {
-          // JSON 也解析不了 → 原始文本兜底
-          yield { type: 'text', content: accumulatedText, timestamp: Date.now() };
-        }
-      }
-      return; // 流已结束，无法继续正常流程
-    }
-
-    // 其他流错误 → 向用户报错
     logger.error(`Stream processing error: ${errMsg}`);
     yield {
       type: 'error',
@@ -341,37 +296,6 @@ export async function* agentLoop(
     finalMessage = await stream.finalMessage();
   } catch (error) {
     const errMsg = error instanceof Error ? error.message : String(error);
-
-    // finalMessage 也可能因 structured output 验证失败而抛错
-    if (errMsg.includes('Failed to parse structured output') && assistantBlocks.length > 0) {
-      logger.warn(`finalMessage structured output validation failed, recovering`);
-      const accumulatedText = assistantBlocks
-        .filter(b => b.type === 'text' && 'text' in b)
-        .map(b => (b as any).text)
-        .join('');
-
-      if (accumulatedText) {
-        try {
-          const parsed = JSON.parse(accumulatedText);
-          if (parsed.answer) {
-            yield { type: 'text', content: parsed.answer, timestamp: Date.now() };
-            if (parsed.followUp) {
-              if (/？|\?/.test(parsed.followUp)) {
-                yield { type: 'text', content: '\n\n' + parsed.followUp, timestamp: Date.now() };
-              } else {
-                yield { type: 'text', content: parsed.followUp, timestamp: Date.now() };
-              }
-            }
-          } else {
-            yield { type: 'text', content: accumulatedText, timestamp: Date.now() };
-          }
-        } catch {
-          yield { type: 'text', content: accumulatedText, timestamp: Date.now() };
-        }
-      }
-      return;
-    }
-
     logger.error(`Failed to get final message: ${errMsg}`);
     yield {
       type: 'error',
@@ -418,44 +342,6 @@ export async function* agentLoop(
       .map(b => b.text)
       .join('');
 
-    // ---- Structured Output 解析：从 JSON 中提取 answer + followUp ----
-    // 长度控制由 提示词 + Schema描述 + LLM-Judge 三层负责
-    // 这里只做 JSON 解析 + 安全网（2000字极端情况保护）
-    if (params.outputConfig && responseText) {
-      try {
-        const parsed = JSON.parse(responseText);
-        if (parsed.answer) {
-          const SAFETY_NET = 2000; // 极端情况安全网，正常不会触发
-          let answer = parsed.answer;
-          if (answer.length > SAFETY_NET) {
-            answer = smartTruncate(answer, SAFETY_NET);
-            logger.warn(
-              `[Turn ${currentTurn + 1}] Answer hit safety net: ${parsed.answer.length} → ${answer.length} chars (intent=${parsed.intent})`,
-            );
-          }
-
-          yield { type: 'text', content: answer, timestamp: Date.now() };
-
-          // followUp：含？→ 跟进问题；否则直接追加（模型可能将内容分到 followUp）
-          if (parsed.followUp) {
-            if (/？|\?/.test(parsed.followUp)) {
-              const followUp = smartTruncate(parsed.followUp, MAX_FOLLOWUP_LENGTH);
-              yield { type: 'text', content: '\n\n' + followUp, timestamp: Date.now() };
-            } else {
-              // 非问题的 followUp 直接追加（防止内容丢失）
-              yield { type: 'text', content: parsed.followUp, timestamp: Date.now() };
-            }
-          }
-        } else {
-          yield { type: 'text', content: responseText, timestamp: Date.now() };
-        }
-        logger.debug(`[Turn ${currentTurn + 1}] Structured output intent: ${parsed.intent}`);
-      } catch {
-        logger.warn(`[Turn ${currentTurn + 1}] Structured output parse failed, falling back to raw text`);
-        yield { type: 'text', content: responseText, timestamp: Date.now() };
-      }
-    }
-
     // --- Evaluation Gate (optional, zero-config safe) ---
     if (params.evaluationGate) {
       try {
diff --git a/packages/services/conversation-service/src/infrastructure/agents/coordinator/coordinator-agent.service.ts b/packages/services/conversation-service/src/infrastructure/agents/coordinator/coordinator-agent.service.ts
index b346264..ee28808 100644
--- a/packages/services/conversation-service/src/infrastructure/agents/coordinator/coordinator-agent.service.ts
+++ b/packages/services/conversation-service/src/infrastructure/agents/coordinator/coordinator-agent.service.ts
@@ -21,10 +21,6 @@ import {
   CoordinatorPromptConfig,
 } from '../prompts/coordinator-system-prompt';
 
-// Structured Output
-import { zodOutputFormat } from '@anthropic-ai/sdk/helpers/zod';
-import { CoordinatorResponseSchema } from '../schemas/coordinator-response.schema';
-
 // Specialist Services
 import { PolicyExpertService } from '../specialists/policy-expert.service';
 import { AssessmentExpertService } from '../specialists/assessment-expert.service';
@@ -290,7 +286,6 @@ export class CoordinatorAgentService implements OnModuleInit {
         currentTurnCount: 0,
         currentCostUsd: 0,
         evaluationGate: evaluationGateCallback,
-        outputConfig: { format: zodOutputFormat(CoordinatorResponseSchema) as any },
       };
 
       // 6. Create tool executor
diff --git a/packages/services/conversation-service/src/infrastructure/agents/schemas/coordinator-response.schema.ts b/packages/services/conversation-service/src/infrastructure/agents/schemas/coordinator-response.schema.ts
index 70e922c..f5ed6ee 100644
--- a/packages/services/conversation-service/src/infrastructure/agents/schemas/coordinator-response.schema.ts
+++ b/packages/services/conversation-service/src/infrastructure/agents/schemas/coordinator-response.schema.ts
@@ -1,52 +1,17 @@
 /**
- * Coordinator Response Schema — Structured Output
+ * Coordinator Response Schema — 已弃用
  *
- * 强制 Coordinator 输出结构化 JSON，包含意图分类和简洁回答。
- * 通过 Anthropic API 的 output_config 实现硬约束。
+ * 历史说明：
+ * 此文件曾定义 Zod schema 用于 Anthropic API 的 output_config (Structured Output)，
+ * 强制模型输出 JSON { intent, answer, followUp }。
+ *
+ * 移除原因（2026-02）：
+ * 1. Structured Output 阻塞了流式输出（JSON 片段无法展示给用户）
+ * 2. Zod 验证 intent 枚举频繁崩溃，已出现 4 次 hotfix
+ * 3. followUp 字段导致模型将内容分割丢失
+ * 4. intent 分类仅用于日志，对用户无价值
+ * 5. 回答精准度完全由 System Prompt (Layer 1) + LLM-Judge (Layer 3) 控制
+ *
+ * 保留此文件（空）以避免 git blame 丢失历史上下文。
+ * 如果未来需要重新引入结构化输出，可参考 git 历史。
  */
-
-import { z } from 'zod';
-
-export const CoordinatorResponseSchema = z.object({
-  intent: z.enum([
-    'factual_question',       // 直接事实问题："X的条件是什么"
-    'yes_no_question',        // 是非判断问题："我能不能申请X"
-    'comparison_question',    // 对比选择问题："A和B哪个好"
-    'assessment_request',     // 评估请求："帮我评估一下"
-    'objection_expression',   // 情绪/犹豫表达："太贵了"/"怕被拒"
-    'detailed_consultation',  // 复杂政策咨询：明确要求详细了解
-    'casual_chat',            // 闲聊/打招呼："你好"
-  ]),
-  answer: z.string().describe('直接回答用户的文本，简洁精准，默认100字以内'),
-  followUp: z.string().optional().describe('直接对用户提出的跟进问题（必须以？结尾，必须是用户能看到的自然语言问题，禁止填写内部策略备注）'),
-});
-
-export type CoordinatorResponse = z.infer<typeof CoordinatorResponseSchema>;
-
-/** followUp 问题最大长度 */
-export const MAX_FOLLOWUP_LENGTH = 80;
-
-/**
- * 智能截断：在句子边界处截断，避免截断在句子中间
- */
-export function smartTruncate(text: string, maxLen: number): string {
-  if (text.length <= maxLen) return text;
-
-  const truncated = text.substring(0, maxLen);
-
-  // 在截断范围内找最后一个句子结束符
-  const sentenceEnders = ['。', '！', '？', '；', '. ', '! ', '? '];
-  let lastEnd = -1;
-  for (const ender of sentenceEnders) {
-    const idx = truncated.lastIndexOf(ender);
-    if (idx > lastEnd) lastEnd = idx;
-  }
-
-  // 如果在后半段找到句子边界，在那里截断
-  if (lastEnd > maxLen * 0.5) {
-    return text.substring(0, lastEnd + 1);
-  }
-
-  // 没有好的边界，硬截断
-  return truncated + '...';
-}
diff --git a/packages/services/conversation-service/src/infrastructure/agents/types/agent.types.ts b/packages/services/conversation-service/src/infrastructure/agents/types/agent.types.ts
index f50c919..c53e4a7 100644
--- a/packages/services/conversation-service/src/infrastructure/agents/types/agent.types.ts
+++ b/packages/services/conversation-service/src/infrastructure/agents/types/agent.types.ts
@@ -280,8 +280,6 @@ export interface AgentLoopParams {
     turnCount: number,
     agentsUsed: string[],
   ) => Promise<import('../coordinator/evaluation-gate.service').GateResult>;
-  /** Structured Output — 传入 Claude API 的 output_config */
-  outputConfig?: { format: Record<string, unknown> };
 }
 
 /** Claude API 消息格式 */