fix: 修复 session 历史丢失 tool_calls

.github/workflows/pr-ci.yml

@@ -18,6 +18,7 @@ jobs:
   test-python:
     name: Python 测试
     runs-on: windows-latest
+    timeout-minutes: 15
 
     steps:
       - name: 检出代码
@@ -57,18 +58,18 @@ jobs:
   test-e2e:
     name: E2E 测试
     runs-on: windows-latest
+    timeout-minutes: 20
     needs: test-python
 
     steps:
       - name: 检出代码
         uses: actions/checkout@v4
 
-      - name: 安装 uv（启用缓存）
+      - name: 安装 uv
         uses: astral-sh/setup-uv@v4
         with:
           version: "latest"
-          enable-cache: true
-          cache-suffix: "pr-ci"
+          enable-cache: false
 
       - name: 恢复 .venv 缓存
         id: cache-venv
@@ -93,19 +94,41 @@ jobs:
         uses: actions/cache@v4
         with:
           path: tests/e2e/node_modules
-          key: node-modules-${{ runner.os }}-${{ hashFiles('tests/e2e/package.json') }}
+          key: node-modules-${{ runner.os }}-${{ hashFiles('tests/e2e/package-lock.json') }}
 
       - name: 安装 E2E 依赖
         if: steps.cache-node-modules.outputs.cache-hit != 'true'
         working-directory: tests/e2e
         run: npm ci
 
+      - name: 验证 mock server 依赖
+        working-directory: tests/e2e
+        run: npm list ws
+
+      - name: 诊断网络状态
+        shell: pwsh
+        run: |
+          Write-Host "Checking if port 12393 is in use..."
+          $connection = Get-NetTCPConnection -LocalPort 12393 -ErrorAction SilentlyContinue
+          if ($connection) {
+            Write-Host "Port 12393 is in use by process: $($connection.OwningProcess)"
+            Get-Process -Id $connection.OwningProcess
+          } else {
+            Write-Host "Port 12393 is NOT in use"
+          }
+
+          Write-Host "`nTesting localhost connectivity..."
+          Test-NetConnection -ComputerName localhost -Port 12393 -WarningAction SilentlyContinue
+
+          Write-Host "`nHosts file content:"
+          Get-Content C:\Windows\System32\drivers\etc\hosts | Select-String -Pattern "localhost"
+
       - name: 缓存 Playwright 浏览器
         id: cache-playwright
         uses: actions/cache@v4
         with:
           path: ~\AppData\Local\ms-playwright
-          key: playwright-${{ runner.os }}-${{ hashFiles('tests/e2e/package.json') }}
+          key: playwright-${{ runner.os }}-${{ hashFiles('tests/e2e/package-lock.json') }}
 
       - name: 安装 Playwright
         if: steps.cache-playwright.outputs.cache-hit != 'true'

CLAUDE.md

@@ -117,7 +117,7 @@ worktree：[已就绪(路径/分支名) | 需要创建 → 先执行创建命令
 |----------|----------|------|
 | 文档更新 | 只改 .md，不涉及代码 | 直接改 |
 | 单文件小修 | ≤1 文件，逻辑明确 | 直接实施 → 自测 |
-| Bug 修复 | 有报错/复现，≤3 文件 | 读错题本 → 实施 → CR（独立 agent）→ P0/P1 归零 → push → 开 PR |
+| Bug 修复 | 有报错/复现，≤3 文件 | 读错题本 → **先 `git log --oneline \| grep -i fix` 数已有 fix 次数，≥2 次立即输出修复门禁** → 实施 → CR（独立 agent）→ P0/P1 归零 → push → 开 PR |
 | 多文件功能 | 跨文件但不涉及新 Module | 读错题本 → 方案确认（列验收标准+技术方案+风险点，用户确认后才能实施）→ 实施 → 自验证 → CR（独立 agent）→ P0/P1 归零 → push → 开 PR |
 | 新 Module | 不在 spine-now.md 冻结范围 | 走 Module 开发流程 5 步 |
 
@@ -147,10 +147,12 @@ worktree：[已就绪(路径/分支名) | 需要创建 → 先执行创建命令
 
 对同一功能/链路提交第 2 次 fix commit 前，**必须先输出修复门禁声明**，再写任何代码。不输出就动手 = DEV-83 违规。
 
+**跨 session 计数规则**：fix 次数必须从 `git log` 里数，不能因新对话而归零。新 session 处理同一链路时，必须先 `git log --oneline | grep -i fix` 检查历史，再判断是否需要输出修复门禁。
+
 ```
 --- 修复门禁 ---
 功能/链路：[一句话描述修的是什么]
-已有 fix 次数：[git log --oneline 计数]
+已有 fix 次数：[git log --oneline | grep -i fix 计数结果，不能凭记忆]
 每次 fix 摘要：[第 N 次改了什么 → 为什么没解决]
 平台验证：[已做最小复现 / 已搜 Issues / 不涉及平台 API]
 门禁结论：[继续当前路径(≤2次) | 强制停下换路径(≥3次)]

docs/error-books/_index.md

@@ -37,6 +37,7 @@
 | DEV-64 | 构建脚本数据源与运行时环境不一致 | 构建/打包 | 🟡×3 | build-packaging.md |
 | DEV-65 | 跨平台路径拼接用了宿主机 path API | 构建/打包 | 🟢 | build-packaging.md |
 | DEV-75 | 跨平台 API 降级只做初始化不做运行时兜底 | 构建/打包 | 🟢 | build-packaging.md |
+| DEV-108 | CI cache post-step 在实际 step 被跳过时找不到路径报错 | 构建/打包 | 🟢 | build-packaging.md |
 | DEV-60 | 隔离对象但共享有状态引用 | 流式/运行时 | 🟢 | streaming-runtime.md |
 | DEV-66 | 实时通道断线缓冲未区分消息时效性 | 流式/运行时 | 🟢 | streaming-runtime.md |
 | DEV-70 | 流式清洗逻辑未处理标签跨 chunk 拆分 | 流式/运行时 | 🟡×2 | streaming-runtime.md |
@@ -76,5 +77,6 @@
 | DEV-100 | 多问题批量修复不分类并行 | 代码习惯 | 🟢 | flow-code-habit.md |
 | DEV-101 | API 边界防御缺失 | 代码习惯 | 🟢 | flow-code-habit.md |
 | DEV-104 | 代码搬迁只验功能可用，未检查隐式依赖 | 代码习惯 | 🟢 | flow-code-habit.md |
+| DEV-107 | E2E 测试未隔离被测 App 的自启动后端，mock server 被杀死 | 代码习惯 | 🟢 | flow-code-habit.md |
 | REC-1~6 | 记录员典型错误（讨论没落盘等） | 通用/记录 | 🟢 | error-book-recorder.md |
 | DEV-94 | README 完成状态与实际启用条件/实现方案不一致 | 代码习惯 | 🟢 | flow-code-habit.md |

docs/error-books/build-packaging.md

@@ -22,6 +22,12 @@
 ✅ 路径拼接涉及目标平台时，用 `path.win32.join` / `path.posix.join` 显式选择，不依赖宿主机默认的 `path.join`。检查点：函数参数里有 `platform` → 路径 API 必须跟着走
 > 归因：`path.join` 的行为取决于 Node 运行的 OS 而非业务逻辑的目标 OS，混淆了"构建环境"与"目标环境"
 
+### DEV-108 CI cache post-step 在实际 step 被跳过时仍执行，找不到缓存路径报错 `🟢`
+
+❌ GitHub Actions 中 `setup-uv` 配置 `enable-cache: true`，但当 `.venv` 缓存命中时 `uv sync` 被跳过，uv 从未下载任何包，缓存目录不存在，post-step 自动保存缓存时报错 → 整个 job 标记为 failure
+✅ 若某 job 内的"真实执行步骤"可能被条件跳过（如 `if: cache-hit != 'true'`），则该步骤依赖的缓存动作必须设 `enable-cache: false` 或加 `if` 守卫，避免 post-step 找不到目录
+> 归因 C：新场景。CI 报错只出现在 post-step，与主流程失败现象混淆，排查成本高。PR #68
+
 ### DEV-75 跨平台 API 降级只做初始化不做运行时兜底 `🟢`
 
 ❌ 初始化时按平台降级（`setIgnoreMouseEvents(false)`），但运行时 IPC 仍无条件调用平台不支持的选项（`forward: true`），降级形同虚设

docs/error-books/flow-code-habit.md

@@ -174,6 +174,12 @@
 ✅ 写 API 端点时：①读 service 完整签名透传所有参数 ②数值字段加范围校验 ③错误语义映射正确状态码 ④错误映射写完后 grep 所有 raise/return error 逐条验证覆盖
 > 从 bot_civ DEV-27 搬运
 
+### DEV-107 E2E 测试未隔离被测 App 的自启动后端逻辑，导致 mock server 被杀死 `🟢`
+
+❌ Electron App 内 `startBackend()` 无条件执行 `netstat` 查杀端口 12393 进程（mock server），再尝试启动 Python 后端（CI 无 runtime），WebSocket 连接失败。测试只验证了"mock server 能启动"，没考虑 App 自身会攻击该端口
+✅ E2E 测试涉及"App 自带 server 启动逻辑"时，必须在 App 内加 `NODE_ENV=test` 守卫跳过真实后端启动。检查点：App 是否有端口清理/进程杀死逻辑？测试 mock 和 App 是否争用同一端口？
+> 归因 C：新场景。测试隔离思路只覆盖了"mock server 不崩溃"，未推演"App 启动后会对 mock 做什么"。PR #68
+
 ### DEV-104 代码搬迁只验功能可用，未检查隐式依赖在新上下文是否成立 `🟢`
 
 ❌ 把托盘菜单的 IPC handler 搬到独立设置窗口，只验证"功能能用"，没推演 `win` 的生命周期在新上下文是否仍然成立。主窗口销毁后设置窗口仍在，IPC 操作已销毁的 BrowserWindow 抛异常

frontend/desktop/main.js

@@ -1203,7 +1203,10 @@ if (!gotLock) {
   app.whenReady().then(() => {
     historyFilePath = resolveHistoryFilePath();
     loadHistoryFromDisk();
-    startBackend();
+    // NODE_ENV=test 时 mock-server.js 已在端口 12393 运行，跳过真实后端启动
+    if (process.env.NODE_ENV !== "test") {
+      startBackend();
+    }
     createWindow();
   });
 }

frontend/desktop/renderer/socket-client.js

@@ -1,7 +1,7 @@
 /**
  * WebSocket 客户端 — 连接灰风后端
  */
-const WS_URL = "ws://127.0.0.1:12393/ws";
+const WS_URL = "ws://localhost:12393/ws";
 let ws = null;
 let reconnectTimer = null;
 let pendingAudioMeta = null;

pyproject.toml

@@ -22,6 +22,7 @@ dependencies = [
     "pyautogui>=0.9.54",
     "pyperclip>=1.8.0",
     "pygetwindow>=0.0.9",
+    "playwright>=1.58.0",
 ]
 
 [project.optional-dependencies]

src/greywind/context_runtime/prompt_assembler.py

@@ -60,10 +60,14 @@ def assemble(
 
         # 最近对话
         for turn in recent_dialogue:
-            messages.append({
+            msg: Dict[str, Any] = {
                 "role": turn["role"],
                 "content": turn["content"],
-            })
+            }
+            # 恢复 tool_calls（如果有）
+            if turn.get("tool_calls"):
+                msg["tool_calls"] = turn["tool_calls"]
+            messages.append(msg)
 
         # 当前用户输入（可能附带截图）
         if screen_image_b64:

src/greywind/context_runtime/session_manager.py

@@ -28,13 +28,16 @@ def state(self) -> str:
     def state(self, value: str):
         self._state = value
 
-    def add_turn(self, role: str, content: str) -> None:
+    def add_turn(self, role: str, content: str, tool_calls: List[Dict[str, Any]] | None = None) -> None:
         """添加一轮对话"""
-        self._recent_dialogue.append({
+        entry: Dict[str, Any] = {
             "role": role,
             "content": content,
             "timestamp": time.time(),
-        })
+        }
+        if tool_calls:
+            entry["tool_calls"] = tool_calls
+        self._recent_dialogue.append(entry)
         if len(self._recent_dialogue) > self._max_recent * 2:
             self._recent_dialogue = self._recent_dialogue[-self._max_recent * 2:]
 

src/greywind/execution/playwright_provider.py

@@ -3,6 +3,7 @@
 
 import asyncio
 import base64
+import subprocess
 import uuid
 from typing import Optional, List, Any
 
@@ -11,6 +12,28 @@
 from .base import BrowserProvider, ActionResult, TabInfo
 
 
+async def _ensure_chromium_installed():
+    """检查并自动安装 Chromium 浏览器二进制"""
+    try:
+        from playwright.sync_api import sync_playwright
+        with sync_playwright() as p:
+            p.chromium.launch()
+        return True
+    except Exception:
+        logger.info("PlaywrightProvider: 正在安装 Chromium...")
+        try:
+            subprocess.run(
+                ["python", "-m", "playwright", "install", "chromium"],
+                capture_output=True,
+                check=True,
+            )
+            logger.info("PlaywrightProvider: Chromium 安装完成")
+            return True
+        except subprocess.CalledProcessError as e:
+            logger.error(f"PlaywrightProvider: Chromium 安装失败 — {e}")
+            return False
+
+
 class PlaywrightProvider(BrowserProvider):
     def __init__(self, screenshot_quality: int = 50, screenshot_width: int = 1280,
                  idle_timeout: int = 60, max_tabs: int = 10,
@@ -33,6 +56,12 @@ async def connect(self) -> bool:
         try:
             from playwright.async_api import async_playwright
             import os
+
+            # 首次启动：自动检查并安装 Chromium
+            if not await _ensure_chromium_installed():
+                logger.error("PlaywrightProvider: Chromium 不可用")
+                return False
+
             self._playwright = await async_playwright().start()
 
             # 持久化 user data dir：保留登录态，下次启动不用重新登录

src/greywind/persona/voice_pipeline.py

@@ -131,15 +131,19 @@ def _start_response(self, text, send_fn, send_audio_fn):
             self._respond(text, send_fn, send_audio_fn)
         )
 
-    async def interrupt(self):
+    async def interrupt(self) -> bool:
+        """打断当前响应。返回 True 表示确实取消了正在运行的任务。"""
         self._interrupted = True
+        was_running = False
         if self._response_task and not self._response_task.done():
+            was_running = True
             self._response_task.cancel()
             try:
                 await self._response_task
             except asyncio.CancelledError:
                 pass
         self._interrupted = False
+        return was_running
 
     async def _respond(self, user_text, send_fn, send_audio_fn):
         self._interrupted = False
@@ -209,7 +213,18 @@ async def _respond(self, user_text, send_fn, send_audio_fn):
                 )
 
                 if clean_response and not self._interrupted:
-                    self.session.add_turn("assistant", _sanitize_llm_text(clean_response))
+                    # 如果有 tool_calls，也保存到 session
+                    tool_calls_data = None
+                    if tool_calls:
+                        tool_calls_data = [
+                            {
+                                "id": tc.id,
+                                "type": "function",
+                                "function": {"name": tc.function.name, "arguments": tc.function.arguments},
+                            }
+                            for tc in tool_calls
+                        ]
+                    self.session.add_turn("assistant", _sanitize_llm_text(clean_response), tool_calls_data)
                 elif round_idx == max_rounds and not clean_response:
                     logger.warning("最后一轮（无 tools）LLM 返回空文本，可能模型不兼容 tool result 上下文")
 

src/greywind/server/ws_handler.py

@@ -158,7 +158,9 @@ async def send_audio(audio_bytes: bytes, payload: dict):
                 logger.info(f"桌面操控即时切换: enabled={enabled}")
 
             elif msg_type == "interrupt":
-                await pipeline.interrupt()
+                was_running = await pipeline.interrupt()
+                if was_running:
+                    await send_msg({"type": "status", "payload": {"state": "idle"}})
 
             elif msg_type == "clear_history":
                 pipeline.session.clear()

tests/e2e/electron.spec.ts

@@ -54,18 +54,27 @@ test.describe("Electron 窗口基础渲染", () => {
     }
   });
 
-  test("status-bar 在连接成功后显示"已连接"", async () => {
+  test('status-bar 在连接成功后显示"已连接"', async () => {
     const app = await electron.launch({
       args: [ELECTRON_APP_DIR],
       env: { ...process.env, NODE_ENV: "test" },
     });
 
+    // 捕获浏览器控制台日志
+    app.on("window", (page) => {
+      page.on("console", (msg) => {
+        console.log(`[Browser Console] ${msg.type()}: ${msg.text()}`);
+      });
+    });
+
     try {
       const win = await app.firstWindow();
       await win.waitForLoadState("domcontentloaded");
+      // 等待更长时间让 WebSocket 连接
+      await win.waitForTimeout(2000);
       // WebSocket 连接 + health check 完成后 status-bar 变"已连接"
       await expect(win.locator("#status-bar")).toContainText("已连接", {
-        timeout: 8000,
+        timeout: 15000,
       });
     } finally {
       await app.close();