feat(agent): stream admin debug preview

ai_knowledge_base_v2/apps/admin-web/package-lock.json

@@ -10,10 +10,12 @@
       "dependencies": {
         "@vitejs/plugin-vue": "^6.0.1",
         "element-plus": "^2.11.0",
+        "markdown-it": "^14.3.0",
         "vite": "^7.1.0",
         "vue": "^3.5.0"
       },
       "devDependencies": {
+        "@types/markdown-it": "^14.1.2",
         "typescript": "^5.9.0",
         "vue-tsc": "^3.1.0"
       }
@@ -877,6 +879,13 @@
       "integrity": "sha512-GhdPgy1el4/ImP05X05Uw4cw2/M93BCUmnEvWZNStlCzEKME4Fkk+YpoA5OiHNQmoS7Cafb8Xa3Pya8m1Qrzeg==",
       "license": "MIT"
     },
+    "node_modules/@types/linkify-it": {
+      "version": "5.0.0",
+      "resolved": "https://registry.npmjs.org/@types/linkify-it/-/linkify-it-5.0.0.tgz",
+      "integrity": "sha512-sVDA58zAw4eWAffKOaQH5/5j3XeayukzDk+ewSsnv3p4yJEZHCCzMDiZM8e0OUrRvmpGZ85jf4yDHkHsgBNr9Q==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/@types/lodash": {
       "version": "4.17.24",
       "resolved": "https://registry.npmjs.org/@types/lodash/-/lodash-4.17.24.tgz",
@@ -892,6 +901,24 @@
         "@types/lodash": "*"
       }
     },
+    "node_modules/@types/markdown-it": {
+      "version": "14.1.2",
+      "resolved": "https://registry.npmjs.org/@types/markdown-it/-/markdown-it-14.1.2.tgz",
+      "integrity": "sha512-promo4eFwuiW+TfGxhi+0x3czqTYJkG8qB17ZUJiVF10Xm7NLVRSLUsfRTU/6h1e24VvRnXCx+hG7li58lkzog==",
+      "dev": true,
+      "license": "MIT",
+      "dependencies": {
+        "@types/linkify-it": "^5",
+        "@types/mdurl": "^2"
+      }
+    },
+    "node_modules/@types/mdurl": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/@types/mdurl/-/mdurl-2.0.0.tgz",
+      "integrity": "sha512-RGdgjQUZba5p6QEFAVx2OGb8rQDL/cPRG7GiedRzMcJ1tYnUANBncjbSB1NRGwbvjcPeikRABz2nshyPk1bhWg==",
+      "dev": true,
+      "license": "MIT"
+    },
     "node_modules/@types/web-bluetooth": {
       "version": "0.0.21",
       "resolved": "https://registry.npmjs.org/@types/web-bluetooth/-/web-bluetooth-0.0.21.tgz",
@@ -1104,6 +1131,12 @@
       "dev": true,
       "license": "MIT"
     },
+    "node_modules/argparse": {
+      "version": "2.0.1",
+      "resolved": "https://registry.npmjs.org/argparse/-/argparse-2.0.1.tgz",
+      "integrity": "sha512-8+9WqebbFzpX9OR+Wa6O29asIogeRMzcGtAINdpMHHyAg10f05aSFVBbcEqGf/PXw1EjAZ+q2/bEBg3DvurK3Q==",
+      "license": "Python-2.0"
+    },
     "node_modules/async-validator": {
       "version": "4.2.5",
       "resolved": "https://registry.npmjs.org/async-validator/-/async-validator-4.2.5.tgz",
@@ -1238,6 +1271,25 @@
         "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
       }
     },
+    "node_modules/linkify-it": {
+      "version": "5.0.2",
+      "resolved": "https://registry.npmjs.org/linkify-it/-/linkify-it-5.0.2.tgz",
+      "integrity": "sha512-ONTm2jCMAVZjgQa/Fy1kScXsuOoF5NPTsoFBdE1KVIZ2vAh/r9+Bqo+0jINCBYnavTPQZz38QzFTme79ENoN3Q==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/puzrin"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/markdown-it"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "uc.micro": "^2.0.0"
+      }
+    },
     "node_modules/lodash": {
       "version": "4.18.1",
       "resolved": "https://registry.npmjs.org/lodash/-/lodash-4.18.1.tgz",
@@ -1270,6 +1322,51 @@
         "@jridgewell/sourcemap-codec": "^1.5.5"
       }
     },
+    "node_modules/markdown-it": {
+      "version": "14.3.0",
+      "resolved": "https://registry.npmjs.org/markdown-it/-/markdown-it-14.3.0.tgz",
+      "integrity": "sha512-RCEsPjR+sr0x+AuYp601tKTkgFG4YEPLCzHST3cQ/fhlJkqAkz1L2/Qbp1j9qw5SBwQHFBoW8+hoN5xssOF0Tw==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/puzrin"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/markdown-it"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "argparse": "^2.0.1",
+        "entities": "^4.5.0",
+        "linkify-it": "^5.0.2",
+        "mdurl": "^2.0.0",
+        "punycode.js": "^2.3.1",
+        "uc.micro": "^2.1.0"
+      },
+      "bin": {
+        "markdown-it": "bin/markdown-it.mjs"
+      }
+    },
+    "node_modules/markdown-it/node_modules/entities": {
+      "version": "4.5.0",
+      "resolved": "https://registry.npmjs.org/entities/-/entities-4.5.0.tgz",
+      "integrity": "sha512-V0hjH4dGPh9Ao5p0MoRY6BVqtwCjhz6vI5LT8AJ55H+4g9/4vbHx1I54fS0XuclLhDHArPQCiMjDxjaL8fPxhw==",
+      "license": "BSD-2-Clause",
+      "engines": {
+        "node": ">=0.12"
+      },
+      "funding": {
+        "url": "https://github.com/fb55/entities?sponsor=1"
+      }
+    },
+    "node_modules/mdurl": {
+      "version": "2.0.0",
+      "resolved": "https://registry.npmjs.org/mdurl/-/mdurl-2.0.0.tgz",
+      "integrity": "sha512-Lf+9+2r+Tdp5wXDXC4PcIBjTDtq4UKjCPMQhKIuzpJNW0b96kVqSwW0bT7FhRSfmAiFYgP+SCRvdrDozfh0U5w==",
+      "license": "MIT"
+    },
     "node_modules/memoize-one": {
       "version": "6.0.0",
       "resolved": "https://registry.npmjs.org/memoize-one/-/memoize-one-6.0.0.tgz",
@@ -1360,6 +1457,15 @@
         "node": "^10 || ^12 || >=14"
       }
     },
+    "node_modules/punycode.js": {
+      "version": "2.3.1",
+      "resolved": "https://registry.npmjs.org/punycode.js/-/punycode.js-2.3.1.tgz",
+      "integrity": "sha512-uxFIHU0YlHYhDQtV4R9J6a52SLx28BCjT+4ieh7IGbgwVJWO+km431c4yRlREUAsAmt/uMjQUyQHNEPf0M39CA==",
+      "license": "MIT",
+      "engines": {
+        "node": ">=6"
+      }
+    },
     "node_modules/rollup": {
       "version": "4.62.2",
       "resolved": "https://registry.npmjs.org/rollup/-/rollup-4.62.2.tgz",
@@ -1443,6 +1549,12 @@
         "node": ">=14.17"
       }
     },
+    "node_modules/uc.micro": {
+      "version": "2.1.0",
+      "resolved": "https://registry.npmjs.org/uc.micro/-/uc.micro-2.1.0.tgz",
+      "integrity": "sha512-ARDJmphmdvUk6Glw7y9DQ2bFkKBHwQHLi2lsaH6PPmz/Ka9sFOBsBluozhDltWmnv9u/cF6Rt87znRTPV+yp/A==",
+      "license": "MIT"
+    },
     "node_modules/vite": {
       "version": "7.3.6",
       "resolved": "https://registry.npmjs.org/vite/-/vite-7.3.6.tgz",

ai_knowledge_base_v2/apps/admin-web/package.json

@@ -11,10 +11,12 @@
   "dependencies": {
     "@vitejs/plugin-vue": "^6.0.1",
     "element-plus": "^2.11.0",
+    "markdown-it": "^14.3.0",
     "vite": "^7.1.0",
     "vue": "^3.5.0"
   },
   "devDependencies": {
+    "@types/markdown-it": "^14.1.2",
     "typescript": "^5.9.0",
     "vue-tsc": "^3.1.0"
   }

ai_knowledge_base_v2/apps/admin-web/src/components/AgentManagementView.vue

@@ -1,11 +1,12 @@
 <script setup lang="ts">
 import { ElMessage, ElMessageBox } from "element-plus";
-import { computed, onMounted, reactive, ref, watch } from "vue";
+import { computed, nextTick, onMounted, reactive, ref, watch } from "vue";
 
-import { api } from "../services/api";
+import { api, streamDebugAgent } from "../services/api";
-import type { AgentDebugResult, AgentRuntimeConfig, KnowledgeItem, ModelItem, PromptDetail, PromptHistoryItem } from "../types/api";
+import type { AgentRuntimeConfig, KnowledgeItem, ModelItem, PromptDetail, PromptHistoryItem } from "../types/api";
 import AgentGenerationParameters from "./AgentGenerationParameters.vue";
 import AdminPagination from "./AdminPagination.vue";
+import StreamingMarkdownMessage from "./StreamingMarkdownMessage.vue";
 
 const props = defineProps<{ previewKnowledgeId?: number | null }>();
 const emit = defineEmits<{ consumedPreview: [] }>();
@@ -15,6 +16,8 @@ const promptSaving = ref(false);
 const runtimeSaving = ref(false);
 const historyLoading = ref(false);
 const agentDebugging = ref(false);
+const agentDebugAbortController = ref<AbortController | null>(null);
+const agentPreviewChat = ref<HTMLElement | null>(null);
 const activeConfigTab = ref("prompt");
 const prompt = ref<PromptDetail | null>(null);
 const promptContent = ref("");
@@ -27,7 +30,7 @@ const historyDetailOpen = ref(false);
 const selectedHistory = ref<PromptDetail | null>(null);
 const historyDetailLoading = ref(false);
 const agentDebugTrace = ref<Record<string, any>[]>([]);
-const agentPreviewMessages = ref<{ role: "user" | "assistant" | "system"; content: string }[]>([
+const agentPreviewMessages = ref<{ role: "user" | "assistant" | "system"; content: string; streaming?: boolean }[]>([
   { role: "assistant", content: "选择模型和知识库后，可以在这里调试 Agent 的真实问答效果。" },
 ]);
 
@@ -254,15 +257,23 @@ async function restoreSelectedHistory() {
 }
 
 async function debugAgent() {
+  if (agentDebugging.value) return;
   if (!agentForm.modelId) return ElMessage.error("请选择调试模型");
   if (!agentForm.question.trim()) return ElMessage.error("请输入调试问题");
   if (!promptContent.value.trim()) return ElMessage.error("主提示词不能为空");
   const question = agentForm.question.trim();
   agentPreviewMessages.value.push({ role: "user", content: question });
+  agentPreviewMessages.value.push({ role: "assistant", content: "", streaming: true });
+  const assistantIndex = agentPreviewMessages.value.length - 1;
+  const currentAssistant = () => agentPreviewMessages.value[assistantIndex];
   agentForm.question = "";
   agentDebugging.value = true;
+  agentDebugTrace.value = [];
+  agentDebugAbortController.value = new AbortController();
+  scrollAgentPreview();
   try {
-    const result: AgentDebugResult = await api.debugAgent({
+    await streamDebugAgent(
+      {
         promptContent: promptContent.value,
         modelId: agentForm.modelId,
         knowledgeIds: agentForm.knowledgeIds,
@@ -273,21 +284,43 @@ async function debugAgent() {
         presencePenalty: agentForm.presencePenalty,
         frequencyPenalty: agentForm.frequencyPenalty,
         maxToken: agentForm.maxToken,
-    });
+      },
-    if (!result.ok) {
+      (chunk) => {
-      agentPreviewMessages.value.push({ role: "assistant", content: result.message });
+        currentAssistant().content += chunk;
-      return ElMessage.error(result.message);
+        scrollAgentPreview();
-    }
+      },
-    agentPreviewMessages.value.push({ role: "assistant", content: result.answer });
+      (result) => {
         agentDebugTrace.value = result.retrievalTrace || [];
-    ElMessage.success(result.message);
+        ElMessage.success(result.message || "Agent 调试完成");
+      },
+      agentDebugAbortController.value.signal,
+    );
+    if (!currentAssistant().content.trim()) currentAssistant().content = "模型未返回正式回答。";
   } catch (error) {
+    if (error instanceof DOMException && error.name === "AbortError") {
+      currentAssistant().content ||= "已停止生成";
+    } else {
       const message = errorMessage(error, "Agent 调试失败");
-    agentPreviewMessages.value.push({ role: "assistant", content: message });
+      currentAssistant().content ||= message;
       ElMessage.error(message);
-  } finally {
-    agentDebugging.value = false;
     }
+  } finally {
+    currentAssistant().streaming = false;
+    agentDebugging.value = false;
+    agentDebugAbortController.value = null;
+    scrollAgentPreview();
+  }
+}
+
+function stopDebugAgent() {
+  agentDebugAbortController.value?.abort();
+}
+
+function scrollAgentPreview() {
+  void nextTick(() => {
+    const element = agentPreviewChat.value;
+    if (element) element.scrollTop = element.scrollHeight;
+  });
 }
 
 function clearAgentPreview() {
@@ -295,15 +328,6 @@ function clearAgentPreview() {
   agentDebugTrace.value = [];
 }
 
-function agentMessageParts(content: string) {
-  const matched = content.match(/<think>([\s\S]*?)<\/think>/i);
-  if (!matched) return { reasoning: "", answer: content };
-  return {
-    reasoning: matched[1].trim(),
-    answer: content.replace(matched[0], "").trim() || "模型未返回正式回答。",
-  };
-}
-
 function changeTypeLabel(value?: string) {
   return ({ save: "手动保存", reset: "恢复默认", restore: "历史恢复", default: "系统默认" } as Record<string, string>)[value || ""] || "保存";
 }
@@ -443,25 +467,23 @@ function errorMessage(error: unknown, fallback: string) {
     <aside class="agent-preview-panel">
       <div class="agent-preview-head">
         <div><h3>调试预览</h3><small>{{ selectedKnowledgeSummary }}</small></div>
-        <div class="agent-preview-head-actions"><span>{{ selectedModelName }}</span><el-button link @click="clearAgentPreview">清空</el-button></div>
+        <div class="agent-preview-head-actions"><span>{{ selectedModelName }}</span><el-button link :disabled="agentDebugging" @click="clearAgentPreview">清空</el-button></div>
       </div>
-      <div class="agent-preview-chat">
+      <div ref="agentPreviewChat" class="agent-preview-chat">
         <article v-for="(message, index) in agentPreviewMessages" :key="`${message.role}-${index}`" class="agent-preview-message-row" :class="message.role">
           <div class="agent-preview-avatar">{{ message.role === "user" ? "你" : "AI" }}</div>
           <div class="agent-preview-bubble">
             <strong>{{ message.role === "user" ? "你" : "Agent" }}</strong>
             <template v-if="message.role === 'assistant'">
-              <details v-if="agentMessageParts(message.content).reasoning" class="agent-preview-thinking"><summary>思考过程</summary><pre>{{ agentMessageParts(message.content).reasoning }}</pre></details>
+              <StreamingMarkdownMessage :content="message.content" :streaming="message.streaming" />
-              <pre>{{ agentMessageParts(message.content).answer }}</pre>
             </template>
             <pre v-else>{{ message.content }}</pre>
           </div>
         </article>
-        <article v-if="agentDebugging" class="agent-preview-message-row assistant thinking"><div class="agent-preview-avatar">AI</div><div class="agent-preview-bubble"><strong>Agent</strong><pre>正在思考中...</pre></div></article>
       </div>
       <div class="agent-preview-composer">
-        <el-input v-model="agentForm.question" type="textarea" :rows="2" resize="none" placeholder="向 Agent 发送测试问题；Enter 发送，Shift+Enter 换行" @keydown.enter.exact.prevent="debugAgent" />
+        <el-input v-model="agentForm.question" type="textarea" :rows="2" resize="none" :disabled="agentDebugging" placeholder="向 Agent 发送测试问题；Enter 发送，Shift+Enter 换行" @keydown.enter.exact.prevent="debugAgent" />
-        <el-button type="primary" :loading="agentDebugging" @click="debugAgent">发送</el-button>
+        <el-button :type="agentDebugging ? 'danger' : 'primary'" @click="agentDebugging ? stopDebugAgent() : debugAgent()">{{ agentDebugging ? '停止' : '发送' }}</el-button>
       </div>
     </aside>
 

ai_knowledge_base_v2/apps/admin-web/src/components/StreamingMarkdownMessage.vue

@@ -0,0 +1,53 @@
+<script setup lang="ts">
+import MarkdownIt from "markdown-it";
+import { computed } from "vue";
+
+const props = defineProps<{
+  content: string;
+  streaming?: boolean;
+}>();
+
+const markdown = new MarkdownIt({ breaks: true, html: false, linkify: true });
+const answer = computed(() => stripStreamingReasoning(props.content).replace(/^\s+/, ""));
+const renderedContent = computed(() => answer.value.trim() ? markdown.render(answer.value) : "");
+
+function stripStreamingReasoning(content: string) {
+  const lower = content.toLowerCase();
+  let visible = "";
+  let cursor = 0;
+  let reasoningDepth = 0;
+  while (cursor < content.length) {
+    const tagStart = content.indexOf("<", cursor);
+    if (tagStart === -1) {
+      if (reasoningDepth === 0) visible += content.slice(cursor);
+      break;
+    }
+    if (reasoningDepth === 0) visible += content.slice(cursor, tagStart);
+    const tail = lower.slice(tagStart);
+    const openTag = tail.match(/^<think(?:\s[^>]*)?>/);
+    if (openTag) {
+      reasoningDepth += 1;
+      cursor = tagStart + openTag[0].length;
+      continue;
+    }
+    const closeTag = tail.match(/^<\/think\s*>/);
+    if (closeTag) {
+      reasoningDepth = Math.max(0, reasoningDepth - 1);
+      cursor = tagStart + closeTag[0].length;
+      continue;
+    }
+    if ("<think".startsWith(tail) || "</think>".startsWith(tail)) break;
+    if (reasoningDepth === 0) visible += "<";
+    cursor = tagStart + 1;
+  }
+  return visible;
+}
+</script>
+
+<template>
+  <div v-if="renderedContent" class="agent-preview-markdown" v-html="renderedContent"></div>
+  <div v-else-if="streaming" class="agent-preview-generation-state">
+    <span></span><span></span><span></span>
+    思考中
+  </div>
+</template>

ai_knowledge_base_v2/apps/admin-web/src/services/api.ts

@@ -2,6 +2,7 @@ import type {
   AdminProfile,
   AdminUser,
   AgentDebugResult,
+  AgentDebugStreamComplete,
   AgentGenerationConfig,
   AgentRuntimeConfig,
   AiLogRecord,
@@ -217,3 +218,48 @@ export const api = {
   clearFeishuCache: () =>
     request<{ cleared: number; message: string }>("/admin/feishu/cache/clear", { method: "POST", body: "{}" }),
 };
+
+export async function streamDebugAgent(
+  payload: Record<string, unknown>,
+  onChunk: (chunk: string) => void,
+  onComplete: (result: AgentDebugStreamComplete) => void,
+  signal?: AbortSignal,
+) {
+  const token = getToken();
+  const response = await fetch(`${API_BASE}/admin/agent/debug/stream`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      ...(token ? { Authorization: `Bearer ${token}` } : {}),
+    },
+    body: JSON.stringify(payload),
+    signal,
+  });
+  if (!response.ok || !response.body) {
+    const body = await readApiResponse<unknown>(response);
+    throw new Error(body.detail || body.message || "Agent 调试连接失败");
+  }
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder("utf-8");
+  let buffer = "";
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) break;
+    buffer += decoder.decode(value, { stream: true });
+    const events = buffer.split(/\r?\n\r?\n/);
+    buffer = events.pop() ?? "";
+    for (const event of events) {
+      const dataLines = event.split(/\r?\n/).filter((line) => line.startsWith("data:"));
+      if (!dataLines.length) continue;
+      const data = dataLines.map((line) => line.slice(5).trimStart()).join("\n").trim();
+      if (data === "[DONE]") return;
+      const parsed = JSON.parse(data) as AgentDebugStreamComplete & {
+        type?: string;
+        content?: string;
+      };
+      if (parsed.type === "error") throw new Error(parsed.message || "Agent 调试失败");
+      if (parsed.type === "content" && parsed.content) onChunk(parsed.content);
+      if (parsed.type === "complete") onComplete(parsed);
+    }
+  }
+}

ai_knowledge_base_v2/apps/admin-web/src/styles.css

@@ -1380,6 +1380,115 @@ textarea {
   line-height: 1.75;
 }
 
+.agent-preview-markdown {
+  overflow-wrap: anywhere;
+  color: #263832;
+  font-size: 14px;
+  line-height: 1.75;
+}
+
+.agent-preview-markdown > :first-child { margin-top: 0; }
+.agent-preview-markdown > :last-child { margin-bottom: 0; }
+.agent-preview-markdown p { margin: 0 0 10px; }
+
+.agent-preview-markdown h1,
+.agent-preview-markdown h2,
+.agent-preview-markdown h3,
+.agent-preview-markdown h4 {
+  margin: 16px 0 8px;
+  color: #1f352e;
+  line-height: 1.45;
+}
+
+.agent-preview-markdown h1 { font-size: 19px; }
+.agent-preview-markdown h2 { font-size: 17px; }
+.agent-preview-markdown h3,
+.agent-preview-markdown h4 { font-size: 15px; }
+
+.agent-preview-markdown ul,
+.agent-preview-markdown ol {
+  margin: 8px 0 10px;
+  padding-left: 22px;
+}
+
+.agent-preview-markdown li + li { margin-top: 5px; }
+.agent-preview-markdown strong { color: #1f352e; font-weight: 700; }
+
+.agent-preview-markdown blockquote {
+  margin: 10px 0;
+  padding: 7px 10px;
+  border-left: 3px solid #8cc7b6;
+  background: #eef7f4;
+  color: #526b63;
+}
+
+.agent-preview-markdown code {
+  padding: 2px 5px;
+  border-radius: 5px;
+  background: #edf5f2;
+  color: #0f735d;
+  font-family: "SFMono-Regular", Consolas, monospace;
+  font-size: 0.92em;
+}
+
+.agent-preview-markdown pre {
+  max-width: 100%;
+  overflow-x: auto;
+  margin: 10px 0;
+  padding: 10px;
+  border-radius: 8px;
+  background: #13231e;
+  color: #f4fffb;
+}
+
+.agent-preview-markdown pre code {
+  padding: 0;
+  background: transparent;
+  color: inherit;
+}
+
+.agent-preview-markdown table {
+  display: block;
+  max-width: 100%;
+  overflow-x: auto;
+  border-collapse: collapse;
+  margin: 10px 0;
+}
+
+.agent-preview-markdown th,
+.agent-preview-markdown td {
+  padding: 6px 8px;
+  border: 1px solid #dbe7e3;
+  white-space: nowrap;
+}
+
+.agent-preview-markdown a { color: #0f735d; font-weight: 700; }
+
+.agent-preview-generation-state {
+  display: flex;
+  align-items: center;
+  gap: 4px;
+  min-height: 28px;
+  color: #667a73;
+  font-size: 13px;
+}
+
+.agent-preview-generation-state span {
+  width: 5px;
+  height: 5px;
+  border-radius: 50%;
+  background: #0f735d;
+  animation: agent-generation-pulse 1.2s ease-in-out infinite;
+}
+
+.agent-preview-generation-state span:nth-child(2) { animation-delay: 0.14s; }
+.agent-preview-generation-state span:nth-child(3) { margin-right: 4px; animation-delay: 0.28s; }
+
+@keyframes agent-generation-pulse {
+  0%, 70%, 100% { opacity: 0.25; transform: translateY(0); }
+  35% { opacity: 1; transform: translateY(-2px); }
+}
+
 .agent-preview-thinking {
   margin-bottom: 10px;
   padding: 9px 10px;

ai_knowledge_base_v2/apps/admin-web/src/types/api.ts

@@ -157,6 +157,15 @@ export interface AgentDebugResult {
   retrievalLogId?: number;
 }
 
+export interface AgentDebugStreamComplete {
+  message: string;
+  modelName?: string;
+  retrieveCount?: number;
+  knowledgeIds?: string;
+  retrievalTrace?: Record<string, unknown>[];
+  retrievalLogId?: number;
+}
+
 export interface KnowledgeVersion {
   id: number;
   knowledgeId: number;

ai_knowledge_base_v2/apps/backend/app/api/admin_settings.py

@@ -1,6 +1,11 @@
 from __future__ import annotations
 
+import json
+from collections.abc import AsyncIterator
+
 from fastapi import APIRouter, Depends, HTTPException, Query, status
+from fastapi.encoders import jsonable_encoder
+from fastapi.responses import StreamingResponse
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
@@ -20,11 +25,10 @@ from app.schemas.admin import (
     SystemConfigSaveRequest,
 )
 from app.services.admin_service import OperationLogService
+from app.services.agent_debug_service import AgentDebugService
 from app.services.feishu_service import FeishuKnowledgeService
-from app.services.knowledge_agent_service import KnowledgeAgentService
 from app.services.knowledge_service import KnowledgeScope
 from app.services.model_service import ModelClientService
-from app.services.rag_service import PromptService, RagResult
 from app.services.secret_service import MASKED_SECRET, SENSITIVE_CONFIG_KEYS, SecretService
 
 router = APIRouter()
@@ -146,37 +150,11 @@ async def debug_agent(
     if model is None:
         raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="模型不存在")
 
-    rag_result = await KnowledgeAgentService.build_result(
+    rag_result = await AgentDebugService.build_result(db, payload)
-        db,
-        question=payload.question,
-        version_overrides=payload.knowledgeVersions,
-        preview_knowledge_ids=payload.knowledgeIds,
-    )
-    debug_messages = [
-        {"role": "system", "content": payload.promptContent.strip()},
-        *(rag_result.messages or []),
-    ]
-    rag_result = RagResult(
-        question=rag_result.question,
-        knowledge_scopes=rag_result.knowledge_scopes,
-        chunks=rag_result.chunks,
-        prompt=PromptService.render_messages(debug_messages),
-        allow_general_knowledge=rag_result.allow_general_knowledge,
-        retrieval_log_id=rag_result.retrieval_log_id,
-        tool_trace=rag_result.tool_trace,
-        messages=debug_messages,
-    )
     result = await ModelClientService.debug_model_async(
         model,
         rag_result,
-        {
+        AgentDebugService.overrides(payload),
-            "temperature": payload.temperature,
-            "top_p": payload.topP,
-            "top_k": payload.topK,
-            "presence_penalty": payload.presencePenalty,
-            "frequency_penalty": payload.frequencyPenalty,
-            "max_token": payload.maxToken,
-        },
     )
     result["retrievalTrace"] = rag_result.tool_trace or []
     result["retrievalLogId"] = rag_result.retrieval_log_id
@@ -185,6 +163,42 @@ async def debug_agent(
     return api_success(result)
 
 
+@router.post("/agent/debug/stream")
+def debug_agent_stream(
+    payload: AgentDebugRequest,
+    db: Session = Depends(get_db),
+    current_admin: Admin = Depends(get_current_admin),
+) -> StreamingResponse:
+    return StreamingResponse(
+        _debug_agent_stream(payload, db, current_admin),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache, no-transform",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+        },
+    )
+
+
+async def _debug_agent_stream(
+    payload: AgentDebugRequest,
+    db: Session,
+    current_admin: Admin,
+) -> AsyncIterator[str]:
+    async for event in AgentDebugService.stream(payload, db, current_admin):
+        yield _admin_sse_event(event)
+    yield _admin_sse_done()
+
+
+def _admin_sse_event(event: dict) -> str:
+    event = jsonable_encoder(event)
+    return f"data: {json.dumps(event, ensure_ascii=False)}\n\n"
+
+
+def _admin_sse_done() -> str:
+    return "data: [DONE]\n\n"
+
+
 @router.get("/agent/runtime-config")
 def get_agent_runtime_config(
     db: Session = Depends(get_db),

ai_knowledge_base_v2/apps/backend/app/services/agent_debug_service.py

@@ -0,0 +1,96 @@
+from __future__ import annotations
+
+import asyncio
+from collections.abc import AsyncIterator
+
+from sqlalchemy.orm import Session
+
+from app.models.admin import Admin
+from app.models.ai_config import ModelConfig
+from app.schemas.admin import AgentDebugRequest
+from app.services.admin_service import OperationLogService
+from app.services.knowledge_agent_service import KnowledgeAgentService
+from app.services.model_stream_service import ModelStreamService
+from app.services.rag_service import PromptService, RagResult
+
+
+class AgentDebugService:
+    @staticmethod
+    async def build_result(db: Session, payload: AgentDebugRequest) -> RagResult:
+        rag_result = await KnowledgeAgentService.build_result(
+            db,
+            question=payload.question,
+            version_overrides=payload.knowledgeVersions,
+            preview_knowledge_ids=payload.knowledgeIds,
+        )
+        debug_messages = [
+            {"role": "system", "content": payload.promptContent.strip()},
+            *(rag_result.messages or []),
+        ]
+        return RagResult(
+            question=rag_result.question,
+            knowledge_scopes=rag_result.knowledge_scopes,
+            chunks=rag_result.chunks,
+            prompt=PromptService.render_messages(debug_messages),
+            allow_general_knowledge=rag_result.allow_general_knowledge,
+            retrieval_log_id=rag_result.retrieval_log_id,
+            tool_trace=rag_result.tool_trace,
+            messages=debug_messages,
+        )
+
+    @staticmethod
+    def overrides(payload: AgentDebugRequest) -> dict:
+        return {
+            "temperature": payload.temperature,
+            "top_p": payload.topP,
+            "top_k": payload.topK,
+            "presence_penalty": payload.presencePenalty,
+            "frequency_penalty": payload.frequencyPenalty,
+            "max_token": payload.maxToken,
+        }
+
+    @classmethod
+    async def stream(
+        cls,
+        payload: AgentDebugRequest,
+        db: Session,
+        current_admin: Admin,
+    ) -> AsyncIterator[dict]:
+        model = db.get(ModelConfig, payload.modelId)
+        if model is None:
+            yield {"type": "error", "message": "模型不存在"}
+            return
+        try:
+            yield {"type": "status", "stage": "retrieving", "message": "思考中"}
+            rag_result = await cls.build_result(db, payload)
+            model_response = ModelStreamService.debug_stream_async(
+                model,
+                rag_result,
+                cls.overrides(payload),
+            )
+            async for chunk in model_response.chunks:
+                if chunk:
+                    yield {"type": "content", "content": chunk}
+            OperationLogService.write(
+                db,
+                admin_id=current_admin.id,
+                module="agent",
+                action="debug_stream",
+                target_id=model.id,
+            )
+            db.commit()
+            yield {
+                "type": "complete",
+                "message": "Agent 调试完成",
+                "modelName": model_response.model_name,
+                "retrieveCount": len(rag_result.chunks),
+                "knowledgeIds": rag_result.knowledge_ids,
+                "retrievalTrace": rag_result.tool_trace or [],
+                "retrievalLogId": rag_result.retrieval_log_id,
+            }
+        except asyncio.CancelledError:
+            db.rollback()
+            raise
+        except Exception as exc:
+            db.rollback()
+            yield {"type": "error", "message": str(exc) or "Agent 调试失败"}

ai_knowledge_base_v2/apps/backend/app/services/model_stream_service.py

@@ -17,6 +17,7 @@ from app.services.model_service import (
     _anthropic_headers,
     _auth_headers,
     _call_configured_model,
+    _copy_model_with_overrides,
     _decimal_to_float,
     _load_extra_params,
     _mock_answer,
@@ -117,6 +118,22 @@ class ModelStreamService:
             chunks=_stream_configured_model_async(model, rag_result),
         )
 
+    @staticmethod
+    def debug_stream_async(
+        model: ModelConfig,
+        rag_result: RagResult,
+        overrides: dict[str, Any],
+    ) -> AsyncStreamingModelResponse:
+        debug_model = _copy_model_with_overrides(model, overrides)
+        if not (debug_model.api_url or debug_model.base_url) or not debug_model.api_key:
+            raise ExternalServiceError("模型 Base URL/API URL 或 API Key 未配置", provider="model")
+        return AsyncStreamingModelResponse(
+            model_id=model.id,
+            model_name=model.model_name,
+            input_token=_rough_token_count(rag_result.prompt),
+            chunks=_stream_configured_model_async(debug_model, rag_result),
+        )
+
 
 def _get_enabled_model(db: Session) -> ModelConfig | None:
     return db.scalar(

ai_knowledge_base_v2/apps/backend/tests/test_agent_runtime_config.py

@@ -1,17 +1,22 @@
 import asyncio
 from decimal import Decimal
+import json
 from unittest.mock import patch
 
 from sqlalchemy import create_engine
 from sqlalchemy.orm import Session
 from sqlalchemy.pool import StaticPool
 
-from app.api.admin_settings import get_agent_runtime_config, save_agent_runtime_config
+from app.api.admin_settings import _debug_agent_stream, get_agent_runtime_config, save_agent_runtime_config
 from app.models import Base
 from app.models.admin import Admin
 from app.models.ai_config import ModelConfig
-from app.schemas.admin import AgentRuntimeConfigSaveRequest
+from app.schemas.admin import AgentDebugRequest, AgentRuntimeConfigSaveRequest
-from app.services.model_stream_service import _openai_stream_payload, _stream_configured_model_async
+from app.services.model_stream_service import (
+    AsyncStreamingModelResponse,
+    _openai_stream_payload,
+    _stream_configured_model_async,
+)
 from app.services.model_service import ModelClientService, _max_output_tokens
 from app.services.rag_service import RagResult
 
@@ -128,6 +133,66 @@ def test_disabled_stream_returns_one_complete_chunk():
     assert chunks == [answer]
 
 
+def test_agent_debug_stream_emits_status_content_and_trace():
+    async def chunks():
+        yield "<think>内部思考</think>"
+        yield "- **第一步**：停一下"
+
+    rag_result = RagResult(
+        question="怎么冷静",
+        knowledge_scopes=[],
+        chunks=[],
+        prompt="怎么冷静",
+        allow_general_knowledge=True,
+        retrieval_log_id=9,
+        tool_trace=[{"tool": "KnowledgeSearch", "status": "success"}],
+        messages=[{"role": "user", "content": "怎么冷静"}],
+    )
+    model_response = AsyncStreamingModelResponse(
+        model_id=1,
+        model_name="production-model",
+        input_token=10,
+        chunks=chunks(),
+    )
+
+    async def collect(payload, db, admin):
+        return [event async for event in _debug_agent_stream(payload, db, admin)]
+
+    with _database() as db:
+        admin = _admin()
+        model = _model()
+        db.add_all([admin, model])
+        db.commit()
+        payload = AgentDebugRequest(
+            promptContent="你是测试助手",
+            modelId=model.id,
+            knowledgeIds=[],
+            question="怎么冷静",
+            maxToken=8192,
+        )
+        with (
+            patch(
+                "app.services.agent_debug_service.KnowledgeAgentService.build_result",
+                return_value=rag_result,
+            ),
+            patch(
+                "app.services.agent_debug_service.ModelStreamService.debug_stream_async",
+                return_value=model_response,
+            ),
+        ):
+            events = asyncio.run(collect(payload, db, admin))
+
+    decoded = [
+        json.loads(event.removeprefix("data: "))
+        for event in events
+        if event != "data: [DONE]\n\n"
+    ]
+    assert [event["type"] for event in decoded] == ["status", "content", "content", "complete"]
+    assert decoded[1]["content"].startswith("<think>")
+    assert decoded[2]["content"] == "- **第一步**：停一下"
+    assert decoded[3]["retrievalTrace"][0]["tool"] == "KnowledgeSearch"
+
+
 def test_agent_debug_does_not_truncate_long_answer(monkeypatch):
     monkeypatch.setattr(
         "app.services.model_service._call_configured_model",