feat: CDP 接管开关 + 扫码登录面板 + Docker 部署

- runner: enable_cdp_mode 从硬编码 False 改为系统设置 cdp_enabled。服务器部署下爬虫接管已开启远程调试的 Chrome（默认 9222），复用其 profile 登录态；本机桌面默认仍为关，行为不变
- qrlogin: 新增 CDP 扫码登录。Chrome 在服务器上跑于 Xvfb，show_qrcode 依赖的 PIL 桌面看图程序不存在，二维码无处可显示；改为经 CDP 从页面取出二维码交给 WebUI 渲染。刻意复用 browser.contexts[0]（新建 context 是无痕 profile，扫了也白扫），且绝不调用 browser.close()（会连带关掉操作者自己的 Chrome）
- webui: 设置页新增扫码面板，替换原本跳到采集页看终端二维码的入口
- Dockerfile / .dockerignore / docker-compose.yml: 服务器部署。host 网络是必需而非图省事——容器里 127.0.0.1:9222 必须落到宿主机回环
- UPSTREAM.md: 补充 gitcode 镜像，用于 GitHub 大包传输必断时补历史

.dockerignore

@@ -0,0 +1,29 @@
+# Keep the build context to what the server actually runs.
+.git
+.github
+.venv
+venv
+
+# The WebUI sources are not needed -- only the bundle they produce, which lands
+# in api/webui and is therefore NOT excluded.
+webui/node_modules
+webui/src
+webui/dist
+
+# Runtime state: per-run crawler output and the login browser profile. Mounted
+# as a volume instead, so it survives image rebuilds.
+data
+browser_data
+
+# Secrets come from the environment via compose, never baked into a layer.
+.env
+
+tests
+docs
+__pycache__
+**/__pycache__
+*.pyc
+*.pyo
+.pytest_cache
+*.db
+*.log

Dockerfile

@@ -0,0 +1,43 @@
+# Server deployment image.
+#
+# No browser is bundled on purpose. On this deployment the crawler attaches over
+# CDP to the Chrome already running on the host (see the 接管已有 Chrome setting),
+# so shipping a second copy of Chromium would only add hundreds of megabytes and
+# a login state that nothing uses. The Playwright Python package is still needed
+# -- that is what speaks CDP -- hence PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD.
+FROM python:3.11-slim
+
+ENV PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 \
+    TZ=Asia/Shanghai
+
+# asyncmy compiles a Cython extension, so a toolchain has to exist at build time.
+# It is left installed: purging it risks taking libmysqlclient with it, and a
+# slightly larger image is cheaper than a runtime that fails months later.
+RUN apt-get update \
+ && apt-get install -y --no-install-recommends \
+      build-essential pkg-config default-libmysqlclient-dev \
+ && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Requirements first: this layer only rebuilds when the pins actually change.
+ARG PIP_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple
+COPY requirements.txt ./
+RUN pip install --no-cache-dir -i "$PIP_INDEX" -r requirements.txt
+
+COPY . .
+
+# The WebUI bundle lives at api/webui, which is gitignored -- it is built on the
+# workstation (`cd webui && npm run build`) and shipped inside the build context.
+# Without this check a missing bundle is invisible until someone opens the page
+# and gets the bare JSON stub from serve_frontend().
+RUN test -f api/webui/index.html \
+ || (echo "ERROR: api/webui/index.html is missing. Run 'cd webui && npm run build' before building the image." >&2; exit 1)
+
+EXPOSE 18051
+
+# api.main reads MC_HOST / MC_PORT from the environment; compose supplies both.
+CMD ["python", "-m", "api.main"]

UPSTREAM.md

@@ -25,9 +25,14 @@ webui/src/components/{monitor,settings,auth}/    新视图
 webui/src/components/layout/{PlatformSwitcher,UnwiredPlatformNotice}.tsx
 webui/src/{hooks/useMonitor.ts,hooks/usePlatform.ts,store/platformStore.ts,lib/monitorFormat.ts,types/monitor.ts}
 docs/监控功能使用说明.md
-tests/test_{auth,settings,platforms,monitor_*}.py
+tests/test_{auth,settings,platforms,qrlogin,monitor_*}.py
+Dockerfile / .dockerignore / docker-compose.yml     服务器部署用
 ```
 
+其中 `api/monitor/qrlogin.py` + `webui/.../QrLoginPanel.tsx` 是**服务器专用**的扫码登录：
+那台机器上 Chrome 跑在 Xvfb 里，`show_qrcode` 调的 PIL `Image.show()` 需要桌面看图程序，
+服务器没有，二维码会无处可去。所以改成用 CDP 把二维码从页面里读出来交给前端 `<img>` 显示。
+
 ### 2. 加法改动（低冲突）
 
 只在既有文件里**新增**内容，不改动原有行：
@@ -90,9 +95,25 @@ tests/test_{auth,settings,platforms,monitor_*}.py
 git stash                      # 或先 commit 到自己的分支（推荐）
 git fetch origin main
 git rebase origin/main         # 冲突只会出现在上表第 3、4 类文件里
-./.venv/Scripts/python.exe -m pytest tests/ -q     # 486 个测试就是回归网
+./.venv/Scripts/python.exe -m pytest tests/ -q     # 502 个测试就是回归网
 ```
 
+### 直连 GitHub 不通时（本机常见）
+
+本机到 `github.com` 时通时不通，**大包传输必断**（`Recv failure: Connection was reset`
+或 `unexpected disconnect while reading sideband packet`），所以 `git clone` / `--unshallow`
+这类一次性拉全量的操作基本必失败。可用的替代源：
+
+```bash
+# gitcode 的 GitHub 镜像，国内直连，比 GitHub 本身还新一天以内
+git remote add gitcode https://gitcode.com/gh_mirrors/me/MediaCrawler.git
+git fetch --no-tags --unshallow gitcode     # 本仓库当初就是这样补全历史的，约 2 秒
+```
+
+注意 `git fetch` 只写 `refs/remotes/*`，**不会动本地 `main`**；
+但拉镜像会把 `upstream/main` 指到镜像的 tip（可能比 GitHub 晚一天），
+等 GitHub 通了再 `git fetch upstream` 正回来即可。
+
 ### 强烈建议：先把改动提交掉
 
 当前状态是**未提交**的（25 个上游文件被改 + 31 个新文件）。在 `main` 分支上裸着工作区，

api/main.py

@@ -65,6 +65,7 @@ async def lifespan(_app: FastAPI):
     happen whether or not anyone has the UI open.
     """
     from .monitor.db import dispose_engine, init_db
+    from .monitor.qrlogin import shutdown as shutdown_qrlogin
     from .monitor.scheduler import monitor_scheduler
 
     await init_db()
@@ -89,6 +90,9 @@ async def lifespan(_app: FastAPI):
         yield
     finally:
         await monitor_scheduler.stop()
+        # Drops the tab a QR login may have opened and stops the Playwright
+        # client; leaving them would strand a driver process on every restart.
+        await shutdown_qrlogin()
         await dispose_engine()
 
 

api/monitor/app_settings.py

@@ -204,6 +204,18 @@ SETTING_SPECS: List[SettingSpec] = [
         maximum=23,
         affects_new_runs=False,
     ),
+    SettingSpec(
+        name="cdp_enabled",
+        scope=SCOPE_SYSTEM,
+        type=TYPE_BOOL,
+        label="接管已有 Chrome（CDP）",
+        help=(
+            "开启后爬虫不再自己启动浏览器，而是接管本机已开放远程调试端口的 Chrome"
+            "（默认 127.0.0.1:9222），复用它的登录态与扩展。"
+            "服务器部署请开启；本机桌面使用请保持关闭。"
+        ),
+        default=False,
+    ),
 ]
 
 SPECS_BY_NAME = {spec.name: spec for spec in SETTING_SPECS}

api/monitor/qrlogin.py

@@ -0,0 +1,268 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2025 [email protected]
+#
+# This file is part of MediaCrawler project.
+# Repository: https://github.com/NanmiCoder/MediaCrawler/blob/main/api/monitor/qrlogin.py
+# GitHub: https://github.com/NanmiCoder
+# Licensed under NON-COMMERCIAL LEARNING LICENSE 1.1
+#
+# 声明：本代码仅供学习和研究目的使用。使用者应遵守以下原则：
+# 1. 不得用于任何商业用途。
+# 2. 使用时应遵守目标平台的使用条款和robots.txt规则。
+# 3. 不得进行大规模爬取或对平台造成运营干扰。
+# 4. 应合理控制请求频率，避免给目标平台带来不必要的负担。
+# 5. 不得用于任何非法或不当的用途。
+#
+# 详细许可条款请参阅项目根目录下的LICENSE文件。
+# 使用本代码即表示您同意遵守上述原则和LICENSE中的所有条款。
+
+"""Show a login QR code to the operator through the WebUI.
+
+Why this module exists: on a server deployment Chrome runs under Xvfb, so there
+is no display to look at, and the helper the crawler normally uses to present the
+QR (`show_qrcode` in `tools/crawler_util.py`) calls PIL's ``Image.show()`` -- it
+needs a desktop image viewer that such a machine does not have, so the code would
+go nowhere and the operator would be stuck. Instead the QR is read straight out
+of the page over CDP and handed to the WebUI, which renders it as an ``<img>``.
+
+Three details that are easy to get wrong:
+
+* **Reuse the browser's default context.** ``browser.new_context()`` would give
+  an incognito-like profile, so the scan would land in a cookie jar the crawler
+  never reads and every run would still look logged out. The real profile -- the
+  one the crawler attaches to -- is ``browser.contexts[0]``.
+* **Never call ``browser.close()``.** On a CDP connection that tears down the
+  operator's own Chrome and takes every unrelated tab with it. Only the page this
+  module opened is closed, and the Playwright client is stopped to drop the
+  socket.
+* **A scan is not a login until the cookie changes.** The page can be showing a
+  QR for an account that is in fact already signed in, so success is decided by
+  ``web_session`` appearing or changing against the value captured at start,
+  never by anything the page displays.
+"""
+
+import asyncio
+import os
+import time
+from typing import Any, Dict, Optional
+
+import config
+from playwright.async_api import async_playwright
+from tools import utils
+
+from .platforms import PLATFORM_XHS
+
+# How long a QR stays valid before the session is written off. The platform
+# rotates the code well before this; the limit exists so an abandoned attempt
+# cannot pin a browser tab open indefinitely.
+QR_TTL_SECONDS = 180
+
+STATUS_IDLE = "idle"
+STATUS_WAITING = "waiting"
+STATUS_SUCCESS = "success"
+STATUS_EXPIRED = "expired"
+STATUS_ERROR = "error"
+
+# Only xhs is wired: it is the only platform whose monitor pipeline works, and
+# pretending otherwise would offer the operator a button that cannot succeed.
+LOGIN_URL: Dict[str, str] = {PLATFORM_XHS: "https://www.xiaohongshu.com"}
+QR_SELECTOR: Dict[str, str] = {PLATFORM_XHS: "xpath=//img[@class='qrcode-img']"}
+LOGIN_BUTTON_SELECTOR: Dict[str, str] = {
+    PLATFORM_XHS: "xpath=//*[@id='app']/div[1]/div[2]/div[1]/ul/div[1]/button"
+}
+SESSION_COOKIE: Dict[str, str] = {PLATFORM_XHS: "web_session"}
+
+_IDLE_SNAPSHOT: Dict[str, Any] = {
+    "status": STATUS_IDLE,
+    "platform": None,
+    "image": "",
+    "message": "",
+    "elapsed": 0,
+    "expires_in": 0,
+}
+
+_lock = asyncio.Lock()
+_current: Optional["QrLoginSession"] = None
+
+
+def _cdp_url() -> str:
+    """Where to reach the browser's DevTools endpoint.
+
+    ``MC_CDP_URL`` wins so a deployment can point at another host without a code
+    change; otherwise the port comes from the same config the crawler itself
+    reads, so the two can never drift apart.
+    """
+    return os.getenv("MC_CDP_URL") or f"http://127.0.0.1:{config.CDP_DEBUG_PORT}"
+
+
+def _login_url(platform: str) -> str:
+    if platform == PLATFORM_XHS and getattr(config, "XHS_INTERNATIONAL", False):
+        return "https://www.rednote.com"
+    return LOGIN_URL[platform]
+
+
+class QrLoginSession:
+    """One live QR-login attempt against the CDP browser."""
+
+    def __init__(self, platform: str, playwright: Any, page: Any, baseline: str) -> None:
+        self.platform = platform
+        self.status = STATUS_WAITING
+        self.message = "请用小红书 App 扫描二维码"
+        self.image = ""
+        self.started_at = time.time()
+        self._playwright = playwright
+        self._page = page
+        # The `web_session` value present *before* the scan. An account already
+        # signed in has a non-empty baseline, which is why success is "changed",
+        # not merely "present".
+        self._baseline = baseline
+
+    @property
+    def elapsed(self) -> float:
+        return time.time() - self.started_at
+
+    async def refresh(self) -> None:
+        """Poll the browser once for a completed scan."""
+        if self.status != STATUS_WAITING:
+            return
+        if self.elapsed > QR_TTL_SECONDS:
+            self.status = STATUS_EXPIRED
+            self.message = "二维码已超时，请重新获取"
+            return
+        try:
+            cookies = await self._page.context.cookies()
+        except Exception:
+            # The operator may have closed the tab we opened.
+            self.status = STATUS_ERROR
+            self.message = "二维码所在页面已被关闭，请重新获取"
+            return
+        token = {c["name"]: c["value"] for c in cookies}.get(
+            SESSION_COOKIE[self.platform], ""
+        )
+        if token and token != self._baseline:
+            self.status = STATUS_SUCCESS
+            self.message = "登录成功，登录态已写入浏览器 profile"
+
+    def snapshot(self) -> Dict[str, Any]:
+        return {
+            "status": self.status,
+            "platform": self.platform,
+            "image": self.image,
+            "message": self.message,
+            "elapsed": int(self.elapsed),
+            "expires_in": max(0, int(QR_TTL_SECONDS - self.elapsed)),
+        }
+
+    async def close(self) -> None:
+        """Drop our page and the Playwright client, leaving Chrome untouched."""
+        try:
+            await self._page.close()
+        except Exception:
+            pass
+        try:
+            await self._playwright.stop()
+        except Exception:
+            pass
+
+
+async def _read_qr(page: Any, platform: str) -> str:
+    """Pull the QR image out of the page, opening the login dialog if needed."""
+    image = await utils.find_login_qrcode(page, selector=QR_SELECTOR[platform])
+    if image:
+        return image
+    # The dialog does not always open on its own. This is the same fallback the
+    # crawler's own QR flow performs before giving up.
+    await asyncio.sleep(0.5)
+    try:
+        await page.locator(LOGIN_BUTTON_SELECTOR[platform]).click(timeout=5000)
+    except Exception:
+        return ""
+    return await utils.find_login_qrcode(page, selector=QR_SELECTOR[platform])
+
+
+async def _reset_locked() -> None:
+    global _current
+    if _current is not None:
+        await _current.close()
+        _current = None
+
+
+async def start(platform: str = PLATFORM_XHS) -> Dict[str, Any]:
+    """Open a login page in the CDP browser and return its QR code."""
+    if platform not in LOGIN_URL:
+        raise ValueError(f"平台 {platform} 尚未接入扫码登录（目前仅支持小红书）")
+
+    async with _lock:
+        await _reset_locked()
+
+        playwright = await async_playwright().start()
+        try:
+            browser = await playwright.chromium.connect_over_cdp(_cdp_url(), timeout=15000)
+        except Exception as exc:
+            await playwright.stop()
+            raise RuntimeError(
+                f"连接浏览器失败（{_cdp_url()}）。请确认服务器上的 Chrome 以 "
+                f"--remote-debugging-port 启动。原始错误：{exc}"
+            ) from exc
+
+        if not browser.contexts:
+            await playwright.stop()
+            raise RuntimeError(
+                "浏览器没有可用上下文。CDP 已连上，但读不到 profile —— "
+                "请确认 Chrome 不是以无痕模式启动的。"
+            )
+
+        # contexts[0] is the real profile. See the module docstring.
+        context = browser.contexts[0]
+        page = await context.new_page()
+        try:
+            await page.goto(_login_url(platform), wait_until="domcontentloaded", timeout=30000)
+            image = await _read_qr(page, platform)
+            cookies = await context.cookies()
+        except Exception as exc:
+            try:
+                await page.close()
+            except Exception:
+                pass
+            await playwright.stop()
+            raise RuntimeError(f"打开登录页失败：{exc}") from exc
+
+        baseline = {c["name"]: c["value"] for c in cookies}.get(
+            SESSION_COOKIE[platform], ""
+        )
+
+        session = QrLoginSession(platform, playwright, page, baseline)
+        session.image = image
+        if not image:
+            if baseline:
+                session.status = STATUS_SUCCESS
+                session.message = "浏览器已经是登录状态，无需扫码"
+            else:
+                session.status = STATUS_ERROR
+                session.message = "页面上没找到二维码，请确认站点结构没有变化"
+
+        global _current
+        _current = session
+        return session.snapshot()
+
+
+async def status() -> Dict[str, Any]:
+    async with _lock:
+        if _current is None:
+            return dict(_IDLE_SNAPSHOT)
+        await _current.refresh()
+        return _current.snapshot()
+
+
+async def cancel() -> Dict[str, Any]:
+    async with _lock:
+        await _reset_locked()
+    snapshot = dict(_IDLE_SNAPSHOT)
+    snapshot["message"] = "已取消"
+    return snapshot
+
+
+async def shutdown() -> None:
+    """Release the browser tab at application shutdown."""
+    async with _lock:
+        await _reset_locked()

api/monitor/runner.py

@@ -162,6 +162,11 @@ async def execute_task(task_id: int, trigger: str = "manual") -> IngestResult:
         urls = build_target_urls(task.mode, targets)
         cookie = await get_cookie(session, platform)
         strategy = await _strategy_settings(session, platform)
+        # System-wide switch. On a headless server the crawler must attach to the
+        # Chrome already listening on the debug port -- that browser is where the
+        # operator scanned the login QR, so its profile is the login. Default False
+        # keeps desktop runs launching a private browser exactly as before.
+        cdp_enabled = await app_settings.get_value(session, "cdp_enabled", fallback=False)
 
         run = MonitorRun(
             task_id=task.id,
@@ -209,8 +214,9 @@ async def execute_task(task_id: int, trigger: str = "manual") -> IngestResult:
         # Isolate this run's output: the crawler names files by date only, so
         # otherwise same-day runs would append into one shared file.
         save_data_path=str(out_dir),
-        # Unattended runs must not try to attach to the user's desktop Chrome.
+        # Attach to the browser already running on CDP_DEBUG_PORT when the
-        enable_cdp_mode=False,
+        # operator enabled it; otherwise launch a private, throwaway browser.
+        enable_cdp_mode=cdp_enabled,
         # Only injecting web_session is not enough to sign requests from a cold
         # browser profile.
         inject_all_cookies=True,

api/routers/monitor.py

@@ -23,7 +23,7 @@ from typing import Any, Dict, List, Optional
 
 from fastapi import APIRouter, HTTPException, Query, Response
 
-from ..monitor import notify, report, service
+from ..monitor import notify, qrlogin, report, service
 from ..monitor.db import get_session
 from ..monitor.platforms import PLATFORM_XHS
 from ..monitor.settings import (
@@ -243,6 +243,43 @@ async def clear_cookie_endpoint(platform: str = Query(default=PLATFORM_XHS)):
         return {"message": "Cookie cleared"}
 
 
+# ---------------------------------------------------------------------------
+# QR login
+# ---------------------------------------------------------------------------
+
+# These drive the browser already listening on the CDP debug port, which is the
+# same browser -- and therefore the same profile -- that monitor runs attach to.
+# Scanning once is what makes later unattended runs logged in.
+
+
+@router.post("/login/qr")
+async def start_qr_login(platform: str = Query(default=PLATFORM_XHS)):
+    """Open the login page in the CDP browser and return its QR code.
+
+    A server deployment has no display (Chrome sits under Xvfb), so the code is
+    surfaced here for the operator to scan instead of in a desktop window that
+    does not exist.
+    """
+    try:
+        return await qrlogin.start(platform)
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    except RuntimeError as exc:
+        raise HTTPException(status_code=502, detail=str(exc))
+
+
+@router.get("/login/qr")
+async def get_qr_login():
+    """Poll the live session: waiting -> success / expired / error."""
+    return await qrlogin.status()
+
+
+@router.delete("/login/qr")
+async def cancel_qr_login():
+    """Drop our tab and stop polling."""
+    return await qrlogin.cancel()
+
+
 # ---------------------------------------------------------------------------
 # Report
 # ---------------------------------------------------------------------------

docker-compose.yml

@@ -0,0 +1,24 @@
+services:
+  mediacrawler:
+    build: .
+    image: mediacrawler:latest
+    container_name: mediacrawler
+    restart: unless-stopped
+
+    # host networking is a requirement, not a convenience: the crawler attaches
+    # to the operator's Chrome at 127.0.0.1:9222, and inside a bridge network
+    # that loopback is the container's own, where no browser is listening.
+    # It also puts the app port directly on the host, so `ports:` is not used.
+    network_mode: host
+
+    env_file:
+      - .env
+    environment:
+      MC_HOST: 0.0.0.0
+      MC_PORT: "18051"
+      TZ: Asia/Shanghai
+
+    volumes:
+      # Per-run crawler output, read back by the monitor layer. Without this the
+      # data would live inside the container and vanish on every rebuild.
+      - ./data:/app/data

tests/test_qrlogin.py

@@ -0,0 +1,190 @@
+# -*- coding: utf-8 -*-
+"""Tests for CDP-driven QR login.
+
+The behaviour worth pinning down is not "does it call Playwright" but the two
+decisions that silently produce a logged-out crawler if they regress:
+
+* the QR must be read from the browser's **default** context, because a new
+  context is an incognito-like profile whose cookies the crawler never sees;
+* success must be decided by ``web_session`` *changing*, not merely existing --
+  an account already signed in has a value before the operator ever scans.
+"""
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from api.monitor import qrlogin
+
+
+@pytest.fixture(autouse=True)
+def _reset_module_state():
+    """Each test starts from "nothing on screen" and leaves it that way."""
+    qrlogin._current = None
+    yield
+    qrlogin._current = None
+
+
+def _fake_stack(qr_image: str = "data:image/png;base64,AAAA", cookies=None):
+    """A Playwright/Chrome stand-in wired the way the real one behaves."""
+    context = MagicMock()
+    context.cookies = AsyncMock(return_value=list(cookies or []))
+    page = MagicMock()
+    page.goto = AsyncMock()
+    page.close = AsyncMock()
+    context.new_page = AsyncMock(return_value=page)
+
+    browser = MagicMock()
+    browser.contexts = [context]
+    # Reaching for a fresh context is the bug this guards against, so make it
+    # blow up loudly rather than quietly returning a throwaway profile.
+    browser.new_context = AsyncMock(
+        side_effect=AssertionError("must reuse browser.contexts[0], not a new context")
+    )
+
+    playwright = MagicMock()
+    playwright.chromium.connect_over_cdp = AsyncMock(return_value=browser)
+    playwright.stop = AsyncMock()
+
+    manager = MagicMock()
+    manager.start = AsyncMock(return_value=playwright)
+
+    return manager, playwright, browser, context, page
+
+
+def _patch(monkeypatch, manager, qr_image="data:image/png;base64,AAAA"):
+    monkeypatch.setattr(qrlogin, "async_playwright", lambda: manager)
+    monkeypatch.setattr(
+        qrlogin.utils, "find_login_qrcode", AsyncMock(return_value=qr_image)
+    )
+
+
+@pytest.mark.asyncio
+async def test_idle_before_any_session():
+    snapshot = await qrlogin.status()
+
+    assert snapshot["status"] == qrlogin.STATUS_IDLE
+    assert snapshot["image"] == ""
+
+
+@pytest.mark.asyncio
+async def test_unwired_platform_is_rejected():
+    """Only xhs is wired; anything else must fail loudly, not show a dead button."""
+    with pytest.raises(ValueError):
+        await qrlogin.start("dy")
+
+
+@pytest.mark.asyncio
+async def test_start_returns_the_qr_and_reuses_the_default_context(monkeypatch):
+    manager, playwright, browser, context, _page = _fake_stack()
+    _patch(monkeypatch, manager)
+
+    snapshot = await qrlogin.start(qrlogin.PLATFORM_XHS)
+
+    assert snapshot["status"] == qrlogin.STATUS_WAITING
+    assert snapshot["image"] == "data:image/png;base64,AAAA"
+    context.new_page.assert_awaited_once()
+    browser.new_context.assert_not_called()
+    playwright.chromium.connect_over_cdp.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_success_requires_the_cookie_to_change():
+    session = qrlogin.QrLoginSession(
+        qrlogin.PLATFORM_XHS,
+        MagicMock(),
+        _page_with_cookies([{"name": "web_session", "value": "before"}]),
+        baseline="before",
+    )
+
+    await session.refresh()
+    assert session.status == qrlogin.STATUS_WAITING
+
+    session._page = _page_with_cookies([{"name": "web_session", "value": "after"}])
+    await session.refresh()
+    assert session.status == qrlogin.STATUS_SUCCESS
+
+
+@pytest.mark.asyncio
+async def test_an_already_signed_in_profile_needs_no_scan(monkeypatch):
+    """No QR on the page plus a session cookie means someone is already logged in."""
+    manager, _playwright, _browser, _context, _page = _fake_stack(
+        qr_image="", cookies=[{"name": "web_session", "value": "existing"}]
+    )
+    _patch(monkeypatch, manager, qr_image="")
+
+    snapshot = await qrlogin.start(qrlogin.PLATFORM_XHS)
+
+    assert snapshot["status"] == qrlogin.STATUS_SUCCESS
+    assert "登录" in snapshot["message"]
+
+
+@pytest.mark.asyncio
+async def test_no_qr_and_no_session_is_an_error(monkeypatch):
+    manager, _playwright, _browser, _context, _page = _fake_stack(qr_image="", cookies=[])
+    _patch(monkeypatch, manager, qr_image="")
+
+    snapshot = await qrlogin.start(qrlogin.PLATFORM_XHS)
+
+    assert snapshot["status"] == qrlogin.STATUS_ERROR
+
+
+@pytest.mark.asyncio
+async def test_session_expires(monkeypatch):
+    session = qrlogin.QrLoginSession(
+        qrlogin.PLATFORM_XHS, MagicMock(), _page_with_cookies([]), baseline=""
+    )
+    session.started_at -= qrlogin.QR_TTL_SECONDS + 1
+
+    await session.refresh()
+
+    assert session.status == qrlogin.STATUS_EXPIRED
+
+
+@pytest.mark.asyncio
+async def test_closed_tab_is_reported_rather_than_crashing():
+    page = MagicMock()
+    page.context.cookies = AsyncMock(side_effect=RuntimeError("Target closed"))
+    session = qrlogin.QrLoginSession(qrlogin.PLATFORM_XHS, MagicMock(), page, baseline="")
+
+    await session.refresh()
+
+    assert session.status == qrlogin.STATUS_ERROR
+
+
+@pytest.mark.asyncio
+async def test_unreachable_browser_is_reported(monkeypatch):
+    """A server whose Chrome is not listening must say so, not 500 anonymously."""
+    playwright = MagicMock()
+    playwright.chromium.connect_over_cdp = AsyncMock(
+        side_effect=OSError("Connection refused")
+    )
+    playwright.stop = AsyncMock()
+    manager = MagicMock()
+    manager.start = AsyncMock(return_value=playwright)
+    monkeypatch.setattr(qrlogin, "async_playwright", lambda: manager)
+
+    with pytest.raises(RuntimeError) as excinfo:
+        await qrlogin.start(qrlogin.PLATFORM_XHS)
+
+    assert "连接浏览器失败" in str(excinfo.value)
+    playwright.stop.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_cancel_closes_the_tab_and_resets(monkeypatch):
+    manager, _playwright, _browser, _context, page = _fake_stack()
+    _patch(monkeypatch, manager)
+    await qrlogin.start(qrlogin.PLATFORM_XHS)
+
+    snapshot = await qrlogin.cancel()
+
+    assert snapshot["status"] == qrlogin.STATUS_IDLE
+    page.close.assert_awaited_once()
+    assert (await qrlogin.status())["status"] == qrlogin.STATUS_IDLE
+
+
+def _page_with_cookies(cookies):
+    page = MagicMock()
+    page.context.cookies = AsyncMock(return_value=list(cookies))
+    return page

webui/src/components/monitor/QrLoginPanel.tsx

@@ -0,0 +1,154 @@
+import { useEffect, useState } from 'react'
+import { useQueryClient } from '@tanstack/react-query'
+import { AlertTriangle, CheckCircle2, Loader2, QrCode, RefreshCw, X } from 'lucide-react'
+
+import { Badge } from '@/components/ui/badge'
+import { Button } from '@/components/ui/button'
+import { useCancelQrLogin, useQrLoginStatus, useStartQrLogin } from '@/hooks/useMonitor'
+import { useCurrentPlatform } from '@/hooks/usePlatform'
+
+/**
+ * Scan-to-login for a host with no display.
+ *
+ * The crawler's own QR flow prints the code into the terminal via PIL's
+ * `Image.show()`, which needs a desktop image viewer. On a server Chrome runs
+ * under Xvfb and there is no such viewer, so the backend reads the code out of
+ * that same browser over CDP and hands it here instead.
+ *
+ * That browser is the one monitor runs attach to, which is the point: scanning
+ * once leaves the login in the profile that every later unattended run reuses.
+ */
+export function QrLoginPanel() {
+  const { capability, platform } = useCurrentPlatform()
+  const queryClient = useQueryClient()
+  const [polling, setPolling] = useState(false)
+
+  const { data: state } = useQrLoginStatus(polling)
+  const start = useStartQrLogin()
+  const cancel = useCancelQrLogin()
+
+  const label = capability?.label ?? platform
+  const status = state?.status ?? 'idle'
+
+  // Stop polling the moment the outcome is known, and refresh the cookie panel:
+  // a completed scan is what makes it start reporting a healthy login.
+  useEffect(() => {
+    if (status === 'waiting' || status === 'idle') return
+    setPolling(false)
+    if (status === 'success') {
+      queryClient.invalidateQueries({ queryKey: ['monitorCookie'] })
+    }
+  }, [status, queryClient])
+
+  const begin = () => {
+    setPolling(true)
+    start.mutate()
+  }
+
+  const busy = start.isPending || cancel.isPending
+
+  return (
+    <div className="rounded-lg glass-panel float-panel p-4 space-y-3">
+      <div className="flex items-center justify-between gap-3">
+        <div className="flex items-center gap-2">
+          <QrCode className="w-4 h-4 text-cyber-neon-cyan" />
+          <span className="font-mono text-xs tracking-wider text-cyber-text-primary">
+            {label}扫码登录
+          </span>
+          {status === 'waiting' && (
+            <Badge variant="warning" className="text-[10px]">
+              等待扫码
+            </Badge>
+          )}
+          {status === 'success' && (
+            <Badge variant="success" className="text-[10px]">
+              已登录
+            </Badge>
+          )}
+        </div>
+
+        {status === 'waiting' && (
+          <Button
+            variant="ghost"
+            size="sm"
+            disabled={busy}
+            onClick={() => {
+              setPolling(false)
+              cancel.mutate()
+            }}
+          >
+            <X className="w-3 h-3" />
+          </Button>
+        )}
+      </div>
+
+      {status === 'waiting' && state?.image && (
+        <div className="flex items-start gap-4">
+          {/* The code is a data: URL straight from the page, so nothing is
+              fetched from a third party and no file is written server-side. */}
+          <img
+            src={state.image}
+            alt="登录二维码"
+            className="w-44 h-44 rounded-md border border-cyber-border-DEFAULT bg-white p-1"
+          />
+          <div className="space-y-2 pt-1">
+            <p className="text-[11px] font-mono text-cyber-text-secondary leading-relaxed">
+              用<span className="text-cyber-neon-cyan">{label} App</span>扫码。
+              扫码成功后登录态会写入服务器上那台 Chrome 的 profile，
+              之后定时监控无需再登录。
+            </p>
+            <p className="text-[11px] font-mono text-cyber-text-muted">
+              剩余 <span className="text-cyber-neon-cyan">{state.expires_in}</span> 秒
+            </p>
+          </div>
+        </div>
+      )}
+
+      {status === 'waiting' && !state?.image && (
+        <p className="flex items-center gap-2 text-[11px] font-mono text-cyber-text-muted">
+          <Loader2 className="w-3 h-3 animate-spin" />
+          正在从浏览器取二维码…
+        </p>
+      )}
+
+      {status === 'success' && (
+        <p className="flex items-start gap-2 text-[11px] font-mono text-cyber-neon-green leading-relaxed">
+          <CheckCircle2 className="w-3 h-3 mt-0.5 shrink-0" />
+          {state?.message || '登录成功'}
+        </p>
+      )}
+
+      {(status === 'error' || status === 'expired') && (
+        <div className="space-y-2">
+          <p className="flex items-start gap-2 text-[11px] font-mono text-cyber-neon-orange leading-relaxed">
+            <AlertTriangle className="w-3 h-3 mt-0.5 shrink-0" />
+            {state?.message || '获取二维码失败'}
+          </p>
+          <Button variant="outline" size="sm" disabled={busy} onClick={begin}>
+            <RefreshCw className="w-3 h-3 mr-1" />
+            重新获取
+          </Button>
+        </div>
+      )}
+
+      {status === 'idle' && (
+        <div className="space-y-2">
+          <p className="text-[11px] font-mono text-cyber-text-muted leading-relaxed">
+            走 CDP 接管服务器上已开启远程调试的 Chrome，把二维码取回来显示在这里。
+            需要先在「系统设置」里打开
+            <span className="text-cyber-neon-cyan">接管已有 Chrome（CDP）</span>，
+            并确保那台 Chrome 正以 <span className="text-cyber-neon-cyan">9222</span> 端口运行。
+          </p>
+          <Button size="sm" disabled={busy} onClick={begin}>
+            {start.isPending ? (
+              <Loader2 className="w-3 h-3 mr-1 animate-spin" />
+            ) : (
+              <QrCode className="w-3 h-3 mr-1" />
+            )}
+            获取二维码
+          </Button>
+        </div>
+      )}
+    </div>
+  )
+}

webui/src/components/settings/SettingsView.tsx

@@ -3,6 +3,7 @@ import { KeyRound, QrCode, Save } from 'lucide-react'
 
 import { Button } from '@/components/ui/button'
 import { CookiePanel } from '@/components/monitor/CookiePanel'
+import { QrLoginPanel } from '@/components/monitor/QrLoginPanel'
 import { Section, SettingField } from '@/components/settings/SettingFields'
 import { useSettings, useUpdateSettings } from '@/hooks/useMonitor'
 import { useCurrentPlatform } from '@/hooks/usePlatform'
@@ -66,17 +67,19 @@ export function SettingsView({ onNavigate }: { onNavigate?: (view: AppView) => v
           </p>
         </div>
 
-        <Section title="登录态" description={`${label}的登录 Cookie，定时监控必须持久化登录态。`}>
+        <Section
+          title="登录态"
+          description={`${label}的登录态。定时监控必须持久化，扫码一次最省事。`}
+        >
+          {/* QR first: it is the one that keeps working unattended, since the
+              scan lands in the very browser profile the monitor runs reuse. */}
+          <QrLoginPanel />
           <CookiePanel />
 
-          {/* Not a duplicate login flow: the crawler has no login-only mode, so a
-              QR login is a side effect of a real crawl -- which is exactly what
-              the 采集 page already does. This preselects it rather than
-              reimplementing it. */}
           <div className="pt-3 mt-1 border-t border-cyber-border-subtle space-y-2">
             <p className="text-[10px] font-mono text-cyber-text-muted leading-relaxed">
-              Cookie 不好使时，可以走一次**扫码登录**：二维码会显示在「采集」页的终端里。
+              在本机桌面运行时，也可以走「采集」页的扫码：二维码会打印在终端里。
-              扫码成功后浏览器 profile 会被更新，Cookie 的可靠性也会显著提升。
+              服务器没有显示器，那条路走不通，用上面的面板。
             </p>
             <Button
               variant="outline"
@@ -89,7 +92,7 @@ export function SettingsView({ onNavigate }: { onNavigate?: (view: AppView) => v
               }}
             >
               <QrCode className="w-3 h-3 mr-1" />
-              去扫码登录
+              去采集页扫码
             </Button>
           </div>
         </Section>

webui/src/hooks/useMonitor.ts

@@ -204,6 +204,46 @@ export function useClearCookie() {
   })
 }
 
+// --- QR login (over CDP) ---------------------------------------------------
+
+/** Polls only while a code is on screen; an idle session has nothing to watch. */
+export function useQrLoginStatus(polling: boolean) {
+  return useQuery({
+    queryKey: ['monitorQrLogin'],
+    queryFn: async () => (await monitorApi.getQrLogin()).data,
+    refetchInterval: polling ? 2000 : false,
+  })
+}
+
+export function useStartQrLogin() {
+  const queryClient = useQueryClient()
+  const platform = usePlatformParam()
+  return useMutation({
+    mutationFn: () => monitorApi.startQrLogin(platform),
+    onSuccess: (response) => {
+      queryClient.setQueryData(['monitorQrLogin'], response.data)
+    },
+    onError: (error: Error) => {
+      // The backend's reason -- "Chrome is not reachable on ...", "no QR on the
+      // page" -- is the entire value here, and axios would flatten it to
+      // "Request failed with status code 502".
+      const detail = (error as { response?: { data?: { detail?: string } } })?.response
+        ?.data?.detail
+      toast.error(`获取二维码失败：${detail ?? error.message}`)
+    },
+  })
+}
+
+export function useCancelQrLogin() {
+  const queryClient = useQueryClient()
+  return useMutation({
+    mutationFn: () => monitorApi.cancelQrLogin(),
+    onSuccess: (response) => {
+      queryClient.setQueryData(['monitorQrLogin'], response.data)
+    },
+  })
+}
+
 // --- Settings -------------------------------------------------------------
 
 export function useSettings() {

webui/src/lib/api.ts

@@ -11,6 +11,7 @@ import type {
   MonitorRun,
   MonitorTask,
   PlatformCapability,
+  QrLoginState,
   ReportResult,
   SettingsResponse,
   TaskCreatePayload,
@@ -256,6 +257,13 @@ export const monitorApi = {
     return api.get<ReportResult>(`/monitor/report?${params.toString()}`)
   },
 
+  // QR login over CDP. Reads the code out of the browser the crawler attaches
+  // to, so the operator can scan even when the host has no display.
+  startQrLogin: (platform?: string) =>
+    api.post<QrLoginState>('/monitor/login/qr', null, { params: { platform } }),
+  getQrLogin: () => api.get<QrLoginState>('/monitor/login/qr'),
+  cancelQrLogin: () => api.delete<QrLoginState>('/monitor/login/qr'),
+
   getWebhook: () => api.get<WebhookStatus>('/monitor/webhook'),
   setWebhook: (url: string) => api.post('/monitor/webhook', { url }),
   clearWebhook: () => api.delete('/monitor/webhook'),

webui/src/types/monitor.ts

@@ -161,6 +161,26 @@ export interface CookieStatus {
   last_ok_at: number | null
 }
 
+export type QrLoginStatus = 'idle' | 'waiting' | 'success' | 'expired' | 'error'
+
+/**
+ * A QR login driven over CDP, for servers with no display to show one on.
+ *
+ * `success` also covers "the browser was already signed in" — the scan and the
+ * existing session both just mean the profile the crawler attaches to is usable.
+ */
+export interface QrLoginState {
+  status: QrLoginStatus
+  platform: string | null
+  /** `data:image/...;base64,...`. Empty unless status is `waiting`. */
+  image: string
+  message: string
+  /** Seconds since the code was fetched. */
+  elapsed: number
+  /** Seconds left before the code is written off. */
+  expires_in: number
+}
+
 export interface MonitorOverview {
   tasks: number
   enabled_tasks: number