Skip to content

v8.2.0

Choose a tag to compare

View all tags
@knoop7 knoop7 released this 16 May 07:11
v8.2.0
6174c74

Acknowledgements

Thanks to the open-source project Scrapling (by D4Vinci). This release incorporates its core concepts of browser fingerprint impersonation and Cloudflare automatic bypass.

Smarter Web Search and Fetching

Previously, some websites could detect non-human access and would block or return blank pages. Now Claw disguises itself as a real browser:

  • Browse like a real person: each request carries a real browser's identity (TLS fingerprint + browser feature headers), making it much harder for websites to distinguish AI from human.
  • Automatic disguise rotation: if blocked, automatically retries with a different browser identity (Chrome -> Edge -> Safari) — three sets in rotation.
  • Automatic Cloudflare challenge solving: when encountering Cloudflare's "Verify you are human" page, Claw automatically completes the challenge without getting stuck.
  • More stable search results: search engine requests are fully upgraded with impersonation, significantly reducing the risk of being blocked.

Request flow

search/fetch request
  ├─ curl_cffi (chrome TLS) → 403?
  ├─ curl_cffi (edge TLS)   → 403?
  ├─ curl_cffi (safari TLS) → all failed?
  └─ aiohttp (fallback, 3 UA rotation retries)
      └─ HTML received → lxml Selector parsing
          └─ js_shell? → patchright rendering (optional)

Smoother Dashboard Operations

  • Auto-navigate to target view: after the AI adds or modifies cards on a different dashboard view, the browser automatically switches to the corresponding dashboard page, allowing you to see results immediately instead of waiting on the wrong page.

Memory System Improvements

  • No duplicate memories: ConversationMemory (simple preference memory) and MemoryGraph (knowledge graph) are now mutually exclusive — the same information will not be stored twice.
  • Automatic graph association: newly stored knowledge nodes automatically search for related nodes and create association edges, linking information automatically.
  • One-click cleanup: MemoryGraph adds a cleanup action that can delete duplicate nodes and garbage data in one click, and automatically rebuild associations.

Frontend Optimizations

  • Scroll listener refactored: the auto-scroll logic of the chat window is reorganized, eliminating duplicate binding issues.
  • File attachment display optimized: attachment tag handling is unified into a global hook, avoiding race conditions from multiple patching.
  • Input security hardened: custom font injection interface now includes input filtering to prevent CSS injection.
  • Polling energy saving: JS execution queue polling reduced from 500ms to 2000ms, and polling automatically stops when the page is not visible.

Technical Optimizations

  • HTML parser migration: replaced BeautifulSoup with a custom lightweight lxml adapter, zero new dependencies, faster startup.
  • Fixed startup warnings: resolved the "Detected blocking call to listdir" warning in the logs.

致谢

感谢开源项目 Scrapling（作者 D4Vinci），本次更新引入了其浏览器指纹伪装和 Cloudflare 自动绕过的核心思路。

更聪明的网页搜索和抓取

以前搜索和打开网页时，有些网站会识别到不是"真人"在访问，直接拒绝或返回空白页。现在 Claw 会伪装成真实浏览器来访问：

  • 像真人一样上网：每次请求都携带真实浏览器的"身份证"（TLS 指纹 + 浏览器特征头），网站很难分辨是 AI 还是真人。
  • 被拦了会自动换马甲：第一次被拦，自动换一个浏览器身份重试（Chrome → Edge → Safari），共三套轮换。
  • Cloudflare 验证页自动通过：遇到 Cloudflare 的"请验证您是人类"页面，Claw 会自动完成验证，不再卡住返回空白。
  • 搜索结果更稳定：搜索引擎请求也全面升级了伪装，降低被封概率。

请求链路示意

搜索/抓取请求
  ├─ curl_cffi (chrome TLS) → 403?
  ├─ curl_cffi (edge TLS)   → 403?
  ├─ curl_cffi (safari TLS) → 全失败?
  └─ aiohttp (原逻辑兜底，3次UA轮换重试)
      └─ 拿到HTML → lxml Selector 解析
          └─ js_shell? → patchright 渲染（可选）

仪表盘操作更丝滑

  • 自动跳转到目标页面：AI 在其他页面添加或修改卡片后，浏览器会自动切到对应的仪表盘页面，直接看到结果，不再停在原来的页面干等。

记忆系统改进

  • 记忆不再重复：ConversationMemory（简单偏好记忆）和 MemoryGraph（知识图谱）现在互斥——同一条信息不会被同时存两份。
  • 图谱自动关联：新存入的知识节点会自动搜索相关节点并建立关联边，信息自动串联。
  • 一键清理：MemoryGraph 新增 cleanup 动作，可以一键删除重复节点、垃圾数据，并自动重建关联。

前端优化

  • 滚动监听重构：聊天窗口的自动滚动逻辑重新整理，消除了重复绑定的问题。
  • 文件附件显示优化：附件标签处理统一到全局钩子，避免了多次 patch 的竞争问题。
  • 输入安全加固：自定义字体注入接口增加了输入过滤，防止 CSS 注入。
  • 轮询节能：JS 执行队列的轮询从 500ms 降到 2000ms，页面不可见时自动停止轮询。

技术优化

  • HTML 解析引擎：用自研的轻量 lxml 适配器替换了 BeautifulSoup，零新依赖、启动更快。
  • 修复启动警告：解决了日志里的 Detected blocking call to listdir 警告。

Full Changelog: v8.1.0...v8.2.0

Assets 2
Loading