✨ 实现 /compact 命令和自动 compact 机制

- 新增模型上下文窗口映射表（model_context.ts），支持主流模型家族前缀匹配
- 新增 compact 摘要 prompt（compact_prompt.ts），参考 Claude Code 的 analysis + summary 结构
- AgentModelConfig 新增 contextWindow 字段，UI 表单支持配置
- handleConversationChat 新增 compact 分支，支持 /compact [指令] 手动压缩
- callLLMWithToolLoop 中当 inputTokens/contextWindow >= 80% 时自动触发 compact
- ChatStreamEvent 新增 compact_done 事件类型
- 新增 27 个单元测试覆盖映射表、摘要提取、手动/自动 compact 逻辑

Author: CodFrm

src/app/service/agent/compact_prompt.test.ts

@@ -0,0 +1,54 @@
+import { describe, expect, it } from "vitest";
+import { extractSummary, buildCompactUserPrompt, COMPACT_SYSTEM_PROMPT } from "./compact_prompt";
+
+describe("extractSummary", () => {
+  it("extracts content from <summary> tags", () => {
+    const response = `<analysis>Some analysis here</analysis>
+
+<summary>
+1. **Primary Request**: Build a feature
+2. **Key Decisions**: Used React
+</summary>`;
+    const result = extractSummary(response);
+    expect(result).toBe("1. **Primary Request**: Build a feature\n2. **Key Decisions**: Used React");
+  });
+
+  it("returns full content when no <summary> tag found", () => {
+    const response = "Just a plain summary without tags";
+    expect(extractSummary(response)).toBe("Just a plain summary without tags");
+  });
+
+  it("handles empty <summary> tags", () => {
+    expect(extractSummary("<summary></summary>")).toBe("");
+  });
+
+  it("handles multiline content inside <summary>", () => {
+    const response = `<summary>
+Line 1
+Line 2
+Line 3
+</summary>`;
+    expect(extractSummary(response)).toBe("Line 1\nLine 2\nLine 3");
+  });
+});
+
+describe("buildCompactUserPrompt", () => {
+  it("builds prompt without custom instruction", () => {
+    const prompt = buildCompactUserPrompt();
+    expect(prompt).toContain("Create a detailed summary");
+    expect(prompt).toContain("<summary>");
+    expect(prompt).not.toContain("Additional summarization instructions");
+  });
+
+  it("appends custom instruction when provided", () => {
+    const prompt = buildCompactUserPrompt("只保留代码相关内容");
+    expect(prompt).toContain("Additional summarization instructions from the user: 只保留代码相关内容");
+  });
+});
+
+describe("COMPACT_SYSTEM_PROMPT", () => {
+  it("is defined and non-empty", () => {
+    expect(COMPACT_SYSTEM_PROMPT).toBeTruthy();
+    expect(COMPACT_SYSTEM_PROMPT.length).toBeGreaterThan(0);
+  });
+});

src/app/service/agent/compact_prompt.ts

@@ -0,0 +1,40 @@
+export const COMPACT_SYSTEM_PROMPT = `You are a conversation summarizer. Your task is to create a detailed summary of the conversation, preserving all critical information needed to continue effectively.`;
+
+export function buildCompactUserPrompt(customInstruction?: string): string {
+  let prompt = `Create a detailed summary of the conversation so far.
+
+Before providing your final summary, wrap your analysis in <analysis> tags to organize your thoughts:
+
+1. Chronologically analyze each message. For each section identify:
+   - The user's explicit requests and intents
+   - Key decisions and outcomes
+   - Specific details: file names, code snippets, function signatures
+   - Errors encountered and how they were fixed
+   - Important user feedback or corrections
+
+2. Double-check for completeness.
+
+Your summary should include the following sections in <summary> tags:
+
+1. **Primary Request and Intent**: The user's core requests and success criteria
+2. **Key Decisions**: Important decisions made and their rationale
+3. **Current State**: What has been completed, files modified, artifacts produced
+4. **Errors and Fixes**: Problems encountered and their solutions
+5. **Pending Tasks**: Outstanding work items
+6. **Current Work**: What was being worked on immediately before this summary
+7. **Next Steps**: Specific actions needed to continue
+
+Be concise but complete — preserve all information that would prevent duplicate work or repeated mistakes.`;
+
+  if (customInstruction) {
+    prompt += `\n\nAdditional summarization instructions from the user: ${customInstruction}`;
+  }
+
+  return prompt;
+}
+
+/** 从 LLM 响应中提取 <summary> 标签内容 */
+export function extractSummary(content: string): string {
+  const match = content.match(/<summary>([\s\S]*?)<\/summary>/);
+  return match ? match[1].trim() : content.trim();
+}

src/app/service/agent/mcp_client.test.ts

@@ -409,6 +409,21 @@ describe("MCPClient", () => {
     });
   });
 
+  describe("请求超时", () => {
+    it("sendRequest 应传递 AbortSignal.timeout(60s)", async () => {
+      const client = new MCPClient(createConfig());
+
+      mockFetch.mockResolvedValueOnce(jsonResponse({ protocolVersion: "2025-03-26", capabilities: {} }));
+      mockFetch.mockResolvedValueOnce(new Response(null, { status: 200 }));
+      await client.initialize();
+
+      // 检查 initialize 的 fetch 调用带了 signal
+      expect(mockFetch.mock.calls[0][1].signal).toBeInstanceOf(AbortSignal);
+      // 检查 notification 的 fetch 调用也带了 signal
+      expect(mockFetch.mock.calls[1][1].signal).toBeInstanceOf(AbortSignal);
+    });
+  });
+
   describe("sendNotification 失败", () => {
     it("initialize 过程中通知失败应抛出", async () => {
       const client = new MCPClient(createConfig());

src/app/service/agent/mcp_client.ts

@@ -196,6 +196,7 @@ export class MCPClient {
       method: "POST",
       headers: this.buildHeaders(),
       body: JSON.stringify(body),
+      signal: AbortSignal.timeout(60_000),
     });
 
     // 存储 session ID
@@ -229,6 +230,7 @@ export class MCPClient {
       method: "POST",
       headers: this.buildHeaders(),
       body: JSON.stringify(body),
+      signal: AbortSignal.timeout(60_000),
     });
 
     // 存储 session ID

src/app/service/agent/model_context.test.ts

@@ -0,0 +1,66 @@
+import { describe, expect, it } from "vitest";
+import { getContextWindow, inferContextWindow, DEFAULT_CONTEXT_WINDOW } from "./model_context";
+
+describe("getContextWindow", () => {
+  it("returns user-configured contextWindow when provided", () => {
+    expect(getContextWindow({ model: "gpt-4o", contextWindow: 50_000 })).toBe(50_000);
+  });
+
+  it("matches GPT-4o prefix", () => {
+    expect(getContextWindow({ model: "gpt-4o" })).toBe(128_000);
+    expect(getContextWindow({ model: "gpt-4o-mini" })).toBe(128_000);
+  });
+
+  it("matches GPT-4.1 prefix before GPT-4", () => {
+    expect(getContextWindow({ model: "gpt-4.1-nano" })).toBe(1_047_576);
+  });
+
+  it("matches GPT-4 Turbo before GPT-4 base", () => {
+    expect(getContextWindow({ model: "gpt-4-turbo-preview" })).toBe(128_000);
+  });
+
+  it("matches GPT-4 base", () => {
+    expect(getContextWindow({ model: "gpt-4-0613" })).toBe(8_192);
+  });
+
+  it("matches Claude models", () => {
+    expect(getContextWindow({ model: "claude-sonnet-4-20250514" })).toBe(200_000);
+    expect(getContextWindow({ model: "claude-3-haiku" })).toBe(200_000);
+  });
+
+  it("matches Gemini models", () => {
+    expect(getContextWindow({ model: "gemini-2.0-flash" })).toBe(1_048_576);
+  });
+
+  it("matches DeepSeek models", () => {
+    expect(getContextWindow({ model: "deepseek-chat" })).toBe(64_000);
+  });
+
+  it("matches Qwen models", () => {
+    expect(getContextWindow({ model: "qwen-max" })).toBe(131_072);
+  });
+
+  it("matches Llama-4 before Llama base", () => {
+    expect(getContextWindow({ model: "llama-4-maverick" })).toBe(1_048_576);
+    expect(getContextWindow({ model: "llama-3.1-70b" })).toBe(131_072);
+  });
+
+  it("is case-insensitive", () => {
+    expect(getContextWindow({ model: "GPT-4O" })).toBe(128_000);
+    expect(getContextWindow({ model: "Claude-Sonnet-4" })).toBe(200_000);
+  });
+
+  it("returns default for unknown models", () => {
+    expect(getContextWindow({ model: "my-custom-model" })).toBe(DEFAULT_CONTEXT_WINDOW);
+  });
+});
+
+describe("inferContextWindow", () => {
+  it("returns prefix-matched value", () => {
+    expect(inferContextWindow("gpt-4o")).toBe(128_000);
+  });
+
+  it("returns default for unknown models", () => {
+    expect(inferContextWindow("unknown-model")).toBe(DEFAULT_CONTEXT_WINDOW);
+  });
+});

src/app/service/agent/model_context.ts

@@ -0,0 +1,62 @@
+// 模型上下文窗口大小映射表
+// [前缀, 上下文窗口大小]，按前缀长度降序排列以确保最精确匹配优先
+const MODEL_CONTEXT_PREFIXES: Array<[string, number]> = [
+  // OpenAI — GPT-5 系列
+  ["gpt-5", 400_000],
+  // OpenAI — GPT-4.1 系列
+  ["gpt-4.1", 1_047_576],
+  // OpenAI — GPT-4o 系列
+  ["gpt-4o", 128_000],
+  // OpenAI — GPT-4 Turbo
+  ["gpt-4-turbo", 128_000],
+  // OpenAI — GPT-4 基础
+  ["gpt-4", 8_192],
+  // OpenAI — GPT-3.5
+  ["gpt-3.5", 16_385],
+  // OpenAI — o 系列推理模型
+  ["o1", 200_000],
+  ["o3", 200_000],
+  ["o4", 200_000],
+  // Anthropic — Claude 系列（所有版本都是 200K）
+  ["claude", 200_000],
+  // Google — Gemini 系列
+  ["gemini", 1_048_576],
+  // Google — Gemma（本地部署）
+  ["gemma", 128_000],
+  // DeepSeek
+  ["deepseek", 64_000],
+  // Alibaba — Qwen 系列
+  ["qwen", 131_072],
+  ["qwq", 32_000],
+  // Meta — Llama 系列
+  ["llama-4", 1_048_576],
+  ["llama", 131_072],
+  // Mistral
+  ["mistral-nemo", 128_000],
+  ["mistral", 32_000],
+  // Microsoft — Phi
+  ["phi", 16_000],
+  // GLM
+  ["glm", 200_000],
+];
+
+export const DEFAULT_CONTEXT_WINDOW = 128_000;
+
+/** 获取模型的上下文窗口大小，优先使用用户配置，否则按前缀匹配 */
+export function getContextWindow(config: { model: string; contextWindow?: number }): number {
+  if (config.contextWindow) return config.contextWindow;
+  const modelLower = config.model.toLowerCase();
+  for (const [prefix, size] of MODEL_CONTEXT_PREFIXES) {
+    if (modelLower.startsWith(prefix)) return size;
+  }
+  return DEFAULT_CONTEXT_WINDOW;
+}
+
+/** 根据模型名称推断上下文窗口大小（不考虑用户配置） */
+export function inferContextWindow(model: string): number {
+  const modelLower = model.toLowerCase();
+  for (const [prefix, size] of MODEL_CONTEXT_PREFIXES) {
+    if (modelLower.startsWith(prefix)) return size;
+  }
+  return DEFAULT_CONTEXT_WINDOW;
+}

src/app/service/agent/skill_script_executor.test.ts

@@ -483,7 +483,7 @@ describe("SkillScriptExecutor 超时处理", () => {
     vi.useRealTimers();
   });
 
-  it("执行超过 30s 时应抛出带 errorCode=tool_timeout 的错误", async () => {
+  it("执行超过默认超时(300s)时应抛出带 errorCode=tool_timeout 的错误", async () => {
     vi.useFakeTimers();
 
     // sender.sendMessage 永不 resolve，模拟挂死的 SkillScript
@@ -497,8 +497,8 @@ describe("SkillScriptExecutor 超时处理", () => {
     // 先附加 catch 再推进时间，防止 rejection 在处理前被标记为 unhandled
     const errPromise = executor.execute({}).catch((e) => e);
 
-    // 推进 30s 触发超时
-    await vi.advanceTimersByTimeAsync(30_000);
+    // 推进 300s 触发超时
+    await vi.advanceTimersByTimeAsync(300_000);
 
     const err = await errPromise;
     expect(err).toBeInstanceOf(Error);
@@ -522,14 +522,14 @@ describe("SkillScriptExecutor 超时处理", () => {
     const executor = new SkillScriptExecutor(record, sender);
 
     const execPromise = executor.execute({}).catch(() => {});
-    await vi.advanceTimersByTimeAsync(30_000);
+    await vi.advanceTimersByTimeAsync(300_000);
     await execPromise;
 
     expect(capturedUuid).toMatch(/^skillscript-/);
     expect(getSkillScriptNameByUuid(capturedUuid)).toBe("");
   });
 
-  it("自定义 timeout 应覆盖默认 30s", async () => {
+  it("自定义 timeout 应覆盖默认 300s", async () => {
     vi.useFakeTimers();
 
     const sender = {
@@ -556,7 +556,7 @@ describe("SkillScriptExecutor 超时处理", () => {
     expect((err as any).errorCode).toBe("tool_timeout");
   });
 
-  it("30s 内完成的执行不应超时", async () => {
+  it("默认超时内完成的执行不应超时", async () => {
     vi.useFakeTimers();
 
     const sender = {

src/app/service/agent/skill_script_executor.ts

@@ -9,7 +9,7 @@ import { uuidv4 } from "@App/pkg/utils/uuid";
 export const SKILL_SCRIPT_UUID_PREFIX = "skillscript-";
 
 // Skill Script 默认超时（ms）
-const SKILL_SCRIPT_DEFAULT_TIMEOUT_MS = 30_000;
+const SKILL_SCRIPT_DEFAULT_TIMEOUT_MS = 300_000;
 
 // 全局的 Skill Script UUID → 工具信息映射，供 GM API 权限验证时使用
 // 直接携带 grants，避免运行时再查 repo（skill 的 Skill Script 不在 skillScriptRepo 中）

src/app/service/agent/tool_registry.ts

@@ -71,27 +71,30 @@ export class ToolRegistry {
 
     const results: ToolExecuteResult[] = [];
 
-    // 执行内置工具
-    for (const tc of builtinCalls) {
-      const tool = this.builtinTools.get(tc.name)!;
-      try {
-        let args: Record<string, unknown> = {};
-        if (tc.arguments) {
-          args = JSON.parse(tc.arguments);
-        }
-        const rawResult = await tool.executor.execute(args);
-
-        // 检查是否带附件
-        if (isToolResultWithAttachments(rawResult)) {
-          const attachments = await this.saveAttachments(rawResult.attachments);
-          results.push({ id: tc.id, result: rawResult.content, attachments });
-        } else {
-          results.push({ id: tc.id, result: typeof rawResult === "string" ? rawResult : JSON.stringify(rawResult) });
+    // 并行执行内置工具
+    const builtinResults = await Promise.all(
+      builtinCalls.map(async (tc): Promise<ToolExecuteResult> => {
+        const tool = this.builtinTools.get(tc.name)!;
+        try {
+          let args: Record<string, unknown> = {};
+          if (tc.arguments) {
+            args = JSON.parse(tc.arguments);
+          }
+          const rawResult = await tool.executor.execute(args);
+
+          // 检查是否带附件
+          if (isToolResultWithAttachments(rawResult)) {
+            const attachments = await this.saveAttachments(rawResult.attachments);
+            return { id: tc.id, result: rawResult.content, attachments };
+          } else {
+            return { id: tc.id, result: typeof rawResult === "string" ? rawResult : JSON.stringify(rawResult) };
+          }
+        } catch (e: any) {
+          return { id: tc.id, result: JSON.stringify({ error: e.message || "Tool execution failed" }) };
         }
-      } catch (e: any) {
-        results.push({ id: tc.id, result: JSON.stringify({ error: e.message || "Tool execution failed" }) });
-      }
-    }
+      })
+    );
+    results.push(...builtinResults);
 
     // 执行脚本工具
     if (scriptCalls.length > 0) {

src/app/service/agent/tools/ask_user.test.ts

@@ -37,7 +37,7 @@ describe("ask_user", () => {
     await expect(executor.execute({})).rejects.toThrow("question is required");
   });
 
-  it("should timeout after 5 minutes", async () => {
+  it("should resolve with timeout reason after 5 minutes", async () => {
     vi.useFakeTimers();
     const sendEvent = vi.fn();
     const resolvers = new Map<string, (answer: string) => void>();
@@ -48,7 +48,8 @@ describe("ask_user", () => {
     // Advance time past timeout
     vi.advanceTimersByTime(5 * 60 * 1000 + 1);
 
-    await expect(resultPromise).rejects.toThrow("User did not respond within 5 minutes");
+    const result = JSON.parse((await resultPromise) as string);
+    expect(result).toEqual({ answer: null, reason: "timeout" });
     expect(resolvers.size).toBe(0);
 
     vi.useRealTimers();