feat(agent): 新增 提示词优化器

.gitignore

@@ -36,6 +36,7 @@ yarn.lock
 .claude
 
 CLAUDE.md
+.omc
 
 test-results
 playwright-report

package.json

@@ -65,8 +65,8 @@
     "@eslint/compat": "^1.4.1",
     "@eslint/js": "9.39.2",
     "@playwright/test": "^1.58.2",
-    "@rspack/cli": "^1.7.6",
-    "@rspack/core": "^1.6.8",
+    "@rspack/cli": "^1.7.11",
+    "@rspack/core": "^1.7.11",
     "@swc/helpers": "^0.5.17",
     "@testing-library/jest-dom": "^6.9.1",
     "@testing-library/react": "^16.3.0",
@@ -107,7 +107,7 @@
     "unocss": "66.5.4",
     "vitest": "^4.0.18"
   },
-  "packageManager": "pnpm@10.12.4",
+  "packageManager": "pnpm@10.33.0",
   "sideEffects": [
     "**/*.css",
     "**/*.scss",

pnpm-workspace.yaml

@@ -0,0 +1 @@
+minimumReleaseAge: 10080

src/app/service/agent/core/optimize_prompt.ts

@@ -0,0 +1,19 @@
+export const OPTIMIZE_PROMPT_SYSTEM = `Act as an expert prompt engineer operating inside the AI Agent feature of ScriptCat (a userscript manager). Your sole function is to transform whatever the user types into a clear, actionable LLM prompt for AI Agent. Transform the input into an actionable LLM prompt covering the relevant dimensions among: role, task, context, constraints, format, and tone. Preserve intent exactly. Inject context ONLY to resolve ambiguity. Do not embellish or over-engineer. Apply minimal edits to already precise inputs. Match input language. Output ONLY the raw prompt text wrapped with <optimized> tag with no preambles, commentary, or markdown fences.`;
+
+export function buildOptimizeUserPrompt(userInput: string): string {
+  return `"""
+${userInput}
+"""`;
+}
+
+/** 从 LLM 响应中提取 <optimized> 标签内容，fallback 到去除常见包裹后的全文 */
+export function extractOptimized(content: string): string {
+  const match = content.match(/<optimized>\s*([\s\S]*?)\s*<\/optimized>/i);
+  if (match) return match[1].trim();
+
+  // Fallback: 清理常见包裹
+  let text = content.trim();
+  text = text.replace(/^```[a-z]*\n?/i, "").replace(/\n?```$/, "");
+  text = text.replace(/<\/?optimized>/gi, "");
+  return text.trim();
+}

src/app/service/agent/core/types.ts

@@ -155,6 +155,7 @@ export type ChatStreamEvent =
       }>;
     }
   | { type: "compact_done"; summary: string; originalCount: number }
+  | { type: "optimize_done"; optimized: string }
   | {
       type: "sync";
       streamingMessage?: { content: string; thinking?: string; toolCalls: ToolCall[] };

src/app/service/agent/service_worker/agent.ts

@@ -151,6 +151,8 @@ export class AgentService {
     this.group.on("conversation", this.handleConversation.bind(this));
     // 流式聊天（UI 和 Sandbox 共用）
     this.group.on("conversationChat", this.handleConversationChat.bind(this));
+    // Prompt 优化（ephemeral 单次调用，复用 conversationChat 路径）
+    this.group.on("optimizePrompt", this.handleOptimizePrompt.bind(this));
     // 附加到后台运行中的会话
     this.group.on("attachToConversation", this.handleAttachToConversation.bind(this));
     // 获取正在运行的会话 ID 列表
@@ -394,6 +396,20 @@ export class AgentService {
     return this.chatService.handleConversationChat(params, sender);
   }
 
+  // Prompt 优化入口：无状态 ephemeral 调用，复用 ChatService.handleConversationChat
+  private async handleOptimizePrompt(params: { modelId: string; input: string }, sender: IGetSender) {
+    return this.chatService.handleConversationChat(
+      {
+        conversationId: `__optimize_${Date.now()}`,
+        message: "",
+        modelId: params.modelId,
+        optimizePrompt: true,
+        optimizeInput: params.input,
+      },
+      sender
+    );
+  }
+
   // 对内容做摘要/提取（供 tab 工具使用）
   // 优先使用摘要模型，fallback 到默认模型
   private async summarizeContent(content: string, prompt: string): Promise<string> {

src/app/service/agent/service_worker/chat_service.ts

@@ -26,6 +26,11 @@ import {
   buildCompactUserPrompt,
   extractSummary,
 } from "@App/app/service/agent/core/compact_prompt";
+import {
+  OPTIMIZE_PROMPT_SYSTEM,
+  buildOptimizeUserPrompt,
+  extractOptimized,
+} from "@App/app/service/agent/core/optimize_prompt";
 import { createTaskTools } from "@App/app/service/agent/core/tools/task_tools";
 import { createAskUserTool } from "@App/app/service/agent/core/tools/ask_user";
 import { createSubAgentTool } from "@App/app/service/agent/core/tools/sub_agent";
@@ -72,6 +77,9 @@ type ConversationChatParams = {
   // compact 模式
   compact?: boolean;
   compactInstruction?: string;
+  // optimize prompt 模式
+  optimizePrompt?: boolean;
+  optimizeInput?: string;
   // 后台运行模式
   background?: boolean;
 };
@@ -258,6 +266,12 @@ export class ChatService {
         return;
       }
 
+      // optimizePrompt
+      if (params.optimizePrompt) {
+        await this.handleOptimizePromptChat(params, sendEvent, abortController);
+        return;
+      }
+
       // 获取对话和模型
       const conv = await this.getConversation(params.conversationId);
       if (!conv) {
@@ -467,6 +481,46 @@ export class ChatService {
     sendEvent({ type: "done", usage: result.usage });
   }
 
+  /**
+   * Prompt 优化：直接调用 LLM，不走 tool loop，不持久化。
+   * 模仿 handleCompactChat 的调用方式。
+   */
+  private async handleOptimizePromptChat(
+    params: ConversationChatParams,
+    sendEvent: (event: ChatStreamEvent) => void,
+    abortController: AbortController
+  ): Promise<void> {
+    if (!params.optimizeInput?.trim()) {
+      sendEvent({ type: "error", message: "No input to optimize" });
+      return;
+    }
+
+    const model = await this.modelService.getModel(params.modelId);
+
+    const messages: ChatRequest["messages"] = [
+      { role: "system", content: OPTIMIZE_PROMPT_SYSTEM },
+      { role: "user", content: buildOptimizeUserPrompt(params.optimizeInput) },
+    ];
+
+    // 吞掉 content_delta / thinking_delta，避免把带 <optimized> 标签的原始流送给 UI
+    const silentSendEvent = (event: ChatStreamEvent) => {
+      if (event.type === "content_delta" || event.type === "thinking_delta") return;
+      sendEvent(event);
+    };
+
+    const result = await this.llmDeps.callLLM(
+      model,
+      { messages, cache: false },
+      silentSendEvent,
+      abortController.signal
+    );
+
+    const optimized = extractOptimized(result.content);
+
+    sendEvent({ type: "optimize_done", optimized } as any);
+    sendEvent({ type: "done", usage: result.usage });
+  }
+
   /**
    * 构建 session 级工具注册表。
    * 每个 chat 请求独立一个 SessionToolRegistry（parent = 全局 toolRegistry），

src/app/service/service_worker/resource.test.ts

@@ -0,0 +1,98 @@
+import { initTestEnv } from "@Tests/utils";
+import { ResourceService } from "./resource";
+import { vi, describe, it, expect, beforeEach } from "vitest";
+import type { Group } from "@Packages/message/server";
+import type { IMessageQueue } from "@Packages/message/message_queue";
+
+initTestEnv();
+
+// mock fetch
+const mockFetch = vi.fn();
+vi.stubGlobal("fetch", mockFetch);
+
+// 创建文本 blob 和二进制 blob 的辅助函数
+function textBlob(content: string, contentType = "text/plain") {
+  return new Blob([content], { type: contentType });
+}
+
+function binaryBlob(bytes: number[]) {
+  return new Blob([new Uint8Array(bytes)], { type: "application/octet-stream" });
+}
+
+function mockResponse(blob: Blob, status = 200, contentType?: string) {
+  return {
+    status,
+    blob: () => Promise.resolve(blob),
+    headers: new Headers(contentType ? { "content-type": contentType } : {}),
+  } as unknown as Response;
+}
+
+describe("ResourceService - loadByUrl", () => {
+  let service: ResourceService;
+
+  beforeEach(() => {
+    vi.clearAllMocks();
+    const mockGroup = {} as Group;
+    const mockMQ = {} as IMessageQueue;
+    service = new ResourceService(mockGroup, mockMQ);
+    // calculateHash 不影响核心逻辑，直接 mock
+    vi.spyOn(service, "calculateHash").mockResolvedValue({
+      md5: "mock-md5",
+      sha1: "",
+      sha256: "",
+      sha384: "",
+      sha512: "",
+    });
+  });
+
+  it("加载文本资源(require)时应设置 content", async () => {
+    const jsCode = "console.log('hello');";
+    mockFetch.mockResolvedValue(mockResponse(textBlob(jsCode), 200, "application/javascript; charset=utf-8"));
+
+    const res = await service.loadByUrl("https://example.com/lib.js", "require");
+
+    expect(res.url).toBe("https://example.com/lib.js");
+    expect(res.content).toBeTruthy();
+    expect(res.contentType).toBe("application/javascript");
+    expect(res.base64).toBeTruthy();
+    expect(res.type).toBe("require");
+  });
+
+  it("加载文本资源(resource)时应通过 blob.text() 设置 content", async () => {
+    const text = "plain text content";
+    mockFetch.mockResolvedValue(mockResponse(textBlob(text), 200, "text/plain"));
+
+    const res = await service.loadByUrl("https://example.com/data.txt", "resource");
+
+    expect(res.content).toBe(text);
+    expect(res.type).toBe("resource");
+  });
+
+  it("加载二进制资源时 content 应为空", async () => {
+    // 包含 null 字节的二进制数据，isText 会返回 false
+    const bytes = [0x89, 0x50, 0x4e, 0x47, 0x00, 0x00, 0x00, 0x00];
+    mockFetch.mockResolvedValue(mockResponse(binaryBlob(bytes), 200, "image/png"));
+
+    const res = await service.loadByUrl("https://example.com/img.png", "resource");
+
+    expect(res.content).toBe("");
+    expect(res.base64).toBeTruthy();
+    expect(res.contentType).toBe("image/png");
+  });
+
+  it("响应非200时应抛出异常", async () => {
+    mockFetch.mockResolvedValue(mockResponse(textBlob(""), 404));
+
+    await expect(service.loadByUrl("https://example.com/404", "require")).rejects.toThrow(
+      "resource response status not 200: 404"
+    );
+  });
+
+  it("没有 content-type 时应默认为 application/octet-stream", async () => {
+    mockFetch.mockResolvedValue(mockResponse(textBlob("data"), 200));
+
+    const res = await service.loadByUrl("https://example.com/noct", "resource");
+
+    expect(res.contentType).toBe("application/octet-stream");
+  });
+});

src/app/service/service_worker/resource.ts

@@ -264,7 +264,7 @@ export class ResourceService {
       throw new Error(`resource response status not 200: ${resp.status}`);
     }
     const data = await resp.blob();
-    const [hash, arrayBuffer, base64] = await Promise.all([
+    const [hash, uint8Array, base64] = await Promise.all([
       this.calculateHash(data),
       blobToUint8Array(data),
       blobToBase64(data),
@@ -280,7 +280,6 @@ export class ResourceService {
       type,
       createtime: Date.now(),
     };
-    const uint8Array = new Uint8Array(arrayBuffer);
     if (isText(uint8Array)) {
       if (type === "require" || type === "require-css") {
         resource.content = await readBlobContent(data, contentType); // @require和@require-css 是会转换成代码运行的，可以进行解码

src/locales/de-DE/translation.json

@@ -742,6 +742,10 @@
   "agent_mcp_no_prompts": "Keine Prompts verfügbar",
   "agent_mcp_loading": "Laden...",
   "agent_mcp_parameters": "Parameter",
+  "agent_optimize_prompt": "Prompt-Optimierer",
+  "agent_optimizing_prompt": "Optimierung...",
+  "agent_optimize_prompt_success": "Prompt optimiert",
+  "agent_optimize_prompt_failed": "Optimierung fehlgeschlagen",
   "agent_settings": "Einstellungen",
   "agent_settings_title": "Agent-Einstellungen",
   "agent_model_settings": "Modelleinstellungen",

src/locales/en-US/translation.json

@@ -746,6 +746,10 @@
   "agent_mcp_no_prompts": "No prompts available",
   "agent_mcp_loading": "Loading...",
   "agent_mcp_parameters": "Parameters",
+  "agent_optimize_prompt": "Prompt Optimizer",
+  "agent_optimizing_prompt": "Optimizing...",
+  "agent_optimize_prompt_success": "Prompt optimized",
+  "agent_optimize_prompt_failed": "Optimization failed",
   "agent_tasks": "Tasks",
   "agent_tasks_title": "Scheduled Tasks",
   "agent_tasks_create": "Create Task",

src/locales/ja-JP/translation.json

@@ -742,6 +742,10 @@
   "agent_mcp_no_prompts": "利用可能なプロンプトはありません",
   "agent_mcp_loading": "読み込み中...",
   "agent_mcp_parameters": "パラメータ",
+  "agent_optimize_prompt": "プロンプト最適化",
+  "agent_optimizing_prompt": "最適化中...",
+  "agent_optimize_prompt_success": "最適化完了",
+  "agent_optimize_prompt_failed": "最適化に失敗しました",
   "agent_settings": "設定",
   "agent_settings_title": "Agent 設定",
   "agent_model_settings": "モデル設定",

src/locales/ru-RU/translation.json

@@ -742,6 +742,10 @@
   "agent_mcp_no_prompts": "Нет доступных промптов",
   "agent_mcp_loading": "Загрузка...",
   "agent_mcp_parameters": "Параметры",
+  "agent_optimize_prompt": "Оптимизатор промптов",
+  "agent_optimizing_prompt": "Оптимизация...",
+  "agent_optimize_prompt_success": "Промпт оптимизирован",
+  "agent_optimize_prompt_failed": "Ошибка оптимизации",
   "agent_settings": "Настройки",
   "agent_settings_title": "Настройки Agent",
   "agent_model_settings": "Настройки модели",

src/locales/vi-VN/translation.json

@@ -742,6 +742,10 @@
   "agent_mcp_no_prompts": "Không có lời nhắc nào",
   "agent_mcp_loading": "Đang tải...",
   "agent_mcp_parameters": "Tham số",
+  "agent_optimize_prompt": "Trình tối ưu hóa câu lệnh",
+  "agent_optimizing_prompt": "Đang tối ưu hóa...",
+  "agent_optimize_prompt_success": "Đã tối ưu hóa",
+  "agent_optimize_prompt_failed": "Tối ưu hóa thất bại",
   "agent_settings": "Cài đặt",
   "agent_settings_title": "Cài đặt Agent",
   "agent_model_settings": "Cài đặt mô hình",

src/locales/zh-CN/translation.json

@@ -746,6 +746,10 @@
   "agent_mcp_no_prompts": "暂无可用提示词",
   "agent_mcp_loading": "加载中...",
   "agent_mcp_parameters": "参数",
+  "agent_optimize_prompt": "提示词优化器",
+  "agent_optimizing_prompt": "优化中...",
+  "agent_optimize_prompt_success": "优化完成",
+  "agent_optimize_prompt_failed": "优化失败",
   "agent_tasks": "定时任务",
   "agent_tasks_title": "定时任务管理",
   "agent_tasks_create": "创建任务",

src/locales/zh-TW/translation.json

@@ -742,6 +742,10 @@
   "agent_mcp_no_prompts": "暫無可用提示詞",
   "agent_mcp_loading": "載入中...",
   "agent_mcp_parameters": "參數",
+  "agent_optimize_prompt": "提示詞優化器",
+  "agent_optimizing_prompt": "優化中...",
+  "agent_optimize_prompt_success": "優化完成",
+  "agent_optimize_prompt_failed": "優化失敗",
   "agent_settings": "設定",
   "agent_settings_title": "Agent 設定",
   "agent_model_settings": "模型設定",