Overview
dsh-websift
Token-efficient web reading for DeepSeek Harness, powered by a real Chrome over CDP. One call fetches several pages, gates them by relevance and returns a single digest (80-98% fewer context tokens). Real search URLs, login flow, screenshots, auto proxy.
README / EN
Package documentation
Registry summary
Token-efficient web reading for DeepSeek Harness, powered by a real Chrome over CDP. One call fetches several pages, gates them by relevance and returns a single digest (80-98% fewer context tokens). Real search URLs, login flow, screenshots, auto proxy.
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
| 限制 | 说明 / 绕法 | |---|---| | ~~自动化特征会被站点识别~~ | **0.9.8 已修**：启动参数默认带 `--disable-blink-features=AutomationControlled`，`navigator.webdriver` 不再是 `true`，GitHub 文件编辑器这类页面能正常渲染 | | 内核启动 3–6 秒 | 首次调用慢；之后复用同一实例 | | 单页正文上限 | `budget_chars` 最大 20000；再大也不会更准（见上面的基准） | | 抽取式摘要 | 句子可能不连贯；纯改写式总结需要再过模型 | | 只处理公开可访问的页面 | 需要短信/扫码的强验证站点要人工介入一次 | | Playwright 不参与 | 插件不依赖 Playwright/Puppeteer，只用 CDP + Node 内置 `WebSocket`，零运行时依赖 | ---