Skip to content
Back to search

envoy

MarketplaceWorks with Claude CodeGitHubIncomplete

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked yesterday
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/regrevia--envoy/badge.svg)](/listings/regrevia--envoy)

HTML

<a href="/listings/regrevia--envoy"><img src="/api/listings/regrevia--envoy/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

These steps run a script from the network that findsafeskills does not scan.

三个宿主各一行。安装脚本只要求 `curl`:uv 和 Python 3.11 它自己装,装好会打印 `envoy install: runtime … protocol … browser …`。

# Claude Code
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- claude-code

# Codex
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- codex

# DeepSeek Harness
curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- dsh

默认是完整档:runtime + Chromium + Whisper `base` 权重,≈1.0 GB(178 + 683 + 141 MB),装完页面上、字幕里、视频中要的东西都能离线取,第一次读不再等下载。Chromium 占其中 1–2 分钟,权重随后下载;两步都可以重跑补齐,幂等。磁盘吃紧或只在 CI/容器里跑,用 `--profile minimal`:只装 runtime(≈178 MB,读普通网页、PDF、图片),Chromium 与权重留给首次使用时的报错命令(`envoy browser setup`、`envoy asr install --model base`)。介于两者之间是 `--profile standard`(runtime + Chromium,≈860 MB,不求转写)。体积口径与平台差异见 [INSTALL.md](integrations/skills/envoy/INSTALL.md)。

不确定宿主是哪个:把最后一段换成 `auto`,脚本会识别已装的 Claude Code / Codex / DSH,都没有就写 MCP 条目(`mcp`、`skill` 也可以单独装;`all` 等同于 `auto`)。只想先看它要做什么,命令前加 `ENVOY_DRY_RUN=1`。从镜像或本地目录安装时,把下载根换成 `ENVOY_RELEASES_BASE`、清单地址换成 `ENVOY_MANIFEST_URL`(目录按 `v<版本>/` 放产物与 `SHA256SUMS`),默认值不变;整仓换到 fork 时设 `ENVOY_REPO`(默认 `regrevia/Envoy`)。Windows PowerShell 用 `irm https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.ps1 | iex`(默认 auto,指定宿主时下载后 `powershell -ExecutionPolicy Bypass -File install.ps1 dsh`)。

装完两条自检:

envoy doctor --json     # runtime.install.status 应为 pinned;完整档下 browser.ready 与 asr_model_available 也应为 true
envoy browser setup     # 最小档或换 Chromium 版本时用(约 683 MB),装完会自动清掉被替换的旧版本
envoy asr install --model base   # 最小档补齐 Whisper 权重(约 141 MB);有 whisper.cpp 就只下 GGML 权重,没有才装 torch
envoy browser prune     # 回收旧版本浏览器下载(升级几次后缓存会翻倍),--dry-run 只报不删

升级就是重跑同一行;卸载、企业镜像、各宿主插件目录见 [INSTALL.md](integrations/skills/envoy/INSTALL.md)。

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 15 lines
  1. {
  2. "name": "envoy",
  3. "owner": {
  4. "name": "regrevia",
  5. "url": "https://github.com/regrevia"
  6. },
  7. "plugins": [
  8. {
  9. "name": "envoy",
  10. "source": "./integrations/claude-code/envoy",
  11. "description": "蒸馏一切链接:抖音/B 站视频、图文笔记、公众号文章、小红书、PDF、播客,一个链接蒸馏出摘要、Markdown、Wiki 或知识库。视频按时序 + 关键帧 + ASR 转写读,材料带来源定位可引用。/ Distill any link: Douyin and Bilibili video, image posts, WeChat articles, Xiaohongshu notes, PDFs and podcasts become located, citable material for a summary, Markdown, wiki or knowledge base; video is read on its own timeline with keyframes and ASR transcripts.",
  12. "version": "0.1.63"
  13. }
  14. ]
  15. }
README.md · README · 322 lines · 8 findings marked
  1. <a id="top"></a>
  2. <p align="center">
  3. <picture>
  4. <source media="(prefers-color-scheme: dark)" srcset="frontend/public/brand/envoy-bird-paper.png">
  5. <img src="frontend/public/brand/envoy-bird-ink.png" alt="Envoy" width="104">
  6. </picture>
  7. </p>
  8. <h1 align="center">蒸馏一切链接</h1>
  9. <p align="center"><strong>把链接分享给你的 AI</strong></p>
  10. <p align="center">
  11. 抖音 / B 站视频 / 图文、公众号文章、小红书、PDF / 视频 —— 一个链接,蒸馏出你需要的一切产物:摘要、Markdown、Wiki、知识库。
  12. </p>
  13. <p align="center">
  14. <a href="#安装">安装 Envoy</a> ·
  15. <a href="https://github.com/regrevia/Envoy/releases">下载发行版</a> ·
  16. <a href="https://github.com/regrevia/Envoy/issues/new/choose">反馈问题</a> ·
  17. <a href="#english">English</a>
  18. </p>
  19. 也可以什么都不写:刷到一条干货视频、一篇笔记,丢个链接给 AI,让它直接当素材和参考。按「时序 + 关键帧 + ASR 转写」读,每一段都定到 `[mm:ss]`,AI 少花 token 就把你要的全部关键内容理解好,而不是先啃标题、推荐流和弹幕。
  20. 已验证适配 **Claude Code**、**Codex**、**DeepSeek Harness**。
  21. > 同一条链接、同一个模型,只差装不装 Envoy:60 轮里能正常读完、并把没读到的说清楚的,22% → 92%;跑到超时也没结论的,68% → 0%。这是疑似完成的统计,没有人工判定对错,口径与原始记录见[评测](#评测)。
  22. 发布出去的产品只有两样:命令行 `envoy`,和 Claude Code / Codex / DeepSeek Harness 的插件与 skills,都装在本机。仓库根目录 `docker-compose.yml` 那套 Web 界面(FastAPI + Next.js + MySQL)是遗留选项,装了 CLI 也不需要它。Python 包目前只随 GitHub Release 发 `envoy_cli-<版本>-py3-none-any.whl`,PyPI 上还没有发布。
  23. ## 演示
  24. 60 秒演示:一条真实链接 → 直接拿到 `[mm:ss]` 字幕片段 → 没读到的范围显式列为待覆盖(读 TED 演讲 0–30 秒)。
  25. ![60 秒演示:链接 → [mm:ss] 片段 → 待覆盖](docs/demo/envoy-demo-60s.gif)
  26. 录制与复现命令见 [docs/demo/README.md](docs/demo/README.md)。
  27. 各站点实际能读到什么(YouTube / B 站 / Vimeo / Dailymotion / Instagram 的真实实测,含网络不通时自动走可信代理)见 [站点可读性说明](docs/acceptance/video-platforms-2026-11-06.md)。
  28. ## 安装
  29. 三个宿主各一行。安装脚本只要求 `curl`:uv 和 Python 3.11 它自己装,装好会打印 `envoy install: runtime … protocol … browser …`。
  30. ```sh
  31. # Claude Code
  32. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- claude-code
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 48 on GitHub
  33. # Codex
  34. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- codex
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 51 on GitHub
  35. # DeepSeek Harness
  36. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- dsh
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 54 on GitHub
  37. ```
  38. 默认是完整档:runtime + Chromium + Whisper `base` 权重,≈1.0 GB(178 + 683 + 141 MB),装完页面上、字幕里、视频中要的东西都能离线取,第一次读不再等下载。Chromium 占其中 1–2 分钟,权重随后下载;两步都可以重跑补齐,幂等。磁盘吃紧或只在 CI/容器里跑,用 `--profile minimal`:只装 runtime(≈178 MB,读普通网页、PDF、图片),Chromium 与权重留给首次使用时的报错命令(`envoy browser setup`、`envoy asr install --model base`)。介于两者之间是 `--profile standard`(runtime + Chromium,≈860 MB,不求转写)。体积口径与平台差异见 [INSTALL.md](integrations/skills/envoy/INSTALL.md)。
  39. 不确定宿主是哪个:把最后一段换成 `auto`,脚本会识别已装的 Claude Code / Codex / DSH,都没有就写 MCP 条目(`mcp`、`skill` 也可以单独装;`all` 等同于 `auto`)。只想先看它要做什么,命令前加 `ENVOY_DRY_RUN=1`。从镜像或本地目录安装时,把下载根换成 `ENVOY_RELEASES_BASE`、清单地址换成 `ENVOY_MANIFEST_URL`(目录按 `v<版本>/` 放产物与 `SHA256SUMS`),默认值不变;整仓换到 fork 时设 `ENVOY_REPO`(默认 `regrevia/Envoy`)。Windows PowerShell 用 `irm https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.ps1 | iex`(默认 auto,指定宿主时下载后 `powershell -ExecutionPolicy Bypass -File install.ps1 dsh`)。
  40. 装完两条自检:
  41. ```sh
  42. envoy doctor --json # runtime.install.status 应为 pinned;完整档下 browser.ready 与 asr_model_available 也应为 true
  43. envoy browser setup # 最小档或换 Chromium 版本时用(约 683 MB),装完会自动清掉被替换的旧版本
  44. envoy asr install --model base # 最小档补齐 Whisper 权重(约 141 MB);有 whisper.cpp 就只下 GGML 权重,没有才装 torch
  45. envoy browser prune # 回收旧版本浏览器下载(升级几次后缓存会翻倍),--dry-run 只报不删
  46. ```
  47. 升级就是重跑同一行;卸载、企业镜像、各宿主插件目录见 [INSTALL.md](integrations/skills/envoy/INSTALL.md)。
  48. ## 能做什么
  49. **关键帧** — 视频里的 PPT、图表、演示画面按时间点抽帧交给 AI 看,不用来回拖进度条。
  50. **视频理解** — 优先读公开字幕;没有字幕时对指定时间段跑本地 Whisper,返回带 `[mm:ss]` 的转写片段。
  51. **图文** — 网页、公众号文章、PDF、图片帖,变成干净正文、图片和来源。
  52. **知识库 / Wiki** — 长视频整段读完,按主题分节,每条结论带时间戳,直接放进 Obsidian、Notion。
  53. ## 怎么用
  54. 对 AI 说:
  55. > 把这条视频整理成知识库:`https://www.bilibili.com/video/BV...`
  56. >
  57. > 先读前 10 分钟,按主题分节;每条结论后标注 `[mm:ss]`;没覆盖到的部分单列“待补”,不要猜。
  58. 材料到手就已经分好段、带定位,所以产出可以直接引用:
  59. ```markdown
  60. ## 2. 为什么选择固定轮
  61. - 固定轮把连续动作拆成可复用的阶段,便于逐段调试。[04:18]
  62. - 作者对比了三种方案,并给出取舍条件。[07:02]
  63. - 这一段没有展示实际测试数据;待补。[09:31]
  64. ```
  65. ## 工作原理
  66. 1. **取材料** — 打开链接,拿到标题、正文、字幕或转写、PDF 文本、图片。
  67. 2. **带定位** — 切成片段并附 `video_time` 或来源坐标,给一个 material id。
  68. 3. **按需续读** — `envoy_search` 定位关键词,`envoy_expand` 补齐没读的部分,不重拉整页。
  69. 4. **缺口照实报** — 没字幕、要登录、模态缺失、下载预算不够,标记为未覆盖,不写成结论。
  70. 浏览器在后台运行,遇到登录或验证才请你接管。接入方式是 Agent Skill + MCP,Claude Code、Codex、DeepSeek Harness 各有插件(Codex 0.156 起需执行一次 `envoy hook install --host codex`)。
  71. ## 边界
  72. 不绕过登录、验证码、地区限制和 DRM;不替代联网搜索。Envoy 只对实际读取并返回的时间段负责。抖音主页集合、小红书图文等能力按[平台验收矩阵](docs/acceptance/platform-coverage-matrix-2026-10-21.md)逐项验证。
  73. ## 评测
  74. 同一条链接、同一个模型,只差装不装 Envoy:20 个任务包 × 3 轮提问,每臂 60 轮,题目与来源冻结在 `benchmarks/fixed-set-v1`。
  75. | 发生了什么 | 没装 Envoy | 装了 Envoy |
  76. | --- | --- | --- |
  77. | 进程正常结束、没有自报读不到(疑似完成,未核实) | 22%(13/60) | **92%(55/60)** |
  78. | 跑到超时也没结论 | 68%(41/60) | **0%** |
  79. | 答案至少命中一条参考事实关键词 | 23%(14/60) | **82%(49/60)** |
  80. | 参考事实关键词覆盖率 | 9.1%(47/516) | **38.2%(197/516)** |
  81. | 第一次拿到证据(中位) | 88 秒 | **48 秒**(少 45%) |
  82. | 回答长度(中位) | 3,125 字 | **8,142 字**(2.6 倍) |
  83. | 空回答(另一组 60 对) | 31 处 | **0 处** |
  84. 百分比是各臂组内占比。两臂时间上限不同(900 秒 vs 2400 秒),完成率差距会被放大;`verified` 两臂都是 0,即本轮没有人工判定答案对错,`suspect_complete` 只表示「正常结束且没说读不到」。
  85. 判定口径、偏差与复核记录见[评测方法](benchmarks/README.md)、[口径与偏差](benchmarks/fixed-set-v1/AB-PROTOCOL.md)与[复核记录](docs/acceptance/envoy-read-failure-2026-10-25.md)。
  86. ## 链接兼容性
  87. 123 条真实链接(新闻、公众号、B站、播客、PDF、论坛、代码文档……)跑了一遍,看三件事:
  88. | 看什么 | 结果 |
  89. | --- | --- |
  90. | 100 条链接里,能拿到可用材料 | **80 条(80.5%)** |
  91. | 能连上的 100 条里,能读出正文 | **93 条(92.5%)** |
  92. | 打不开的链接要干等多久 | **26 秒**(原来 77 秒,少了 67%) |
  93. 「可用」= 能当素材:整篇可引用,或读到一部分并写明缺了什么;读不出来会说清原因(要登录、被人机校验拦住、网络不通),不猜。
  94. 剩下没拿到的 24 条里,16 条根本连不上(全是境外站点,卡在本机代理出口),8 条卡在登录墙或人机校验——这两种在本轮实验的条件下装不装 Envoy 都读不到,Envoy 会直说,让 Agent 走别的路。
  95. 装不装 Envoy 的对照见[评测](#评测)。复算:`scripts/acceptance/link_coverage.py`(报告与口径见[兼容性实验](docs/experiments/link-compatibility.md))。
  96. ## 本地 Web 界面(Docker)
  97. **插件与 skills 不依赖 Docker。** Claude Code / Codex / DSH 三个装法都是本机跑:用 `install.sh` / `install.ps1` 安装脚本,或从 GitHub Release 下 `envoy_cli-<版本>-py3-none-any.whl` 交给 `uv tool install`,再加一个 skills 目录或插件包(PyPI 上还没有发布包),不需要容器、不需要这个 compose。Docker 只服务仓库里那套旧的 Web 界面(FastAPI + Next.js + MySQL),用来自本机点着看任务和产物;不用这个界面就完全不用管它。
  98. 需要 Docker Desktop,仓库根目录:
  99. ```bash
  100. cp .env.example .env
  101. docker compose up -d --build
  102. ```
  103. 前端 `http://127.0.0.1:8200`,后端健康检查 `http://127.0.0.1:8140/api/v1/health` 返回 `{"status":"ok"}`。`docker compose ps` 里 `mysql`、`redis`、`backend`、`frontend` 都 `healthy` 即可用;MySQL 首次启动会自动执行 `sql/` 下的建表脚本,上传与导出都写在 `envoy_runtime` 卷里,默认不需要对象存储。首次构建会拉 MySQL 8.4、Redis,并在后端镜像里装 FFmpeg、Tesseract、Chromium,耗时几分钟。`docker compose down` 保留数据卷;只有 `down -v` 会删库和本地文件。端口、密码、模型配置都在 `.env`,构建加速与排错见 [Docker 一键启动](docs/quickstart-docker.md)。
  104. ## 文档
  105. - [安装、更新与 MCP 配置](integrations/skills/envoy/INSTALL.md)
  106. - [Claude Code 插件](integrations/claude-code/envoy/README.md)
  107. - [Codex 插件](integrations/codex/envoy/README.md)
  108. - [DeepSeek Harness 插件](integrations/dsh/README.md)
  109. - [产品与技术状态](docs/architecture/agent-link-reader-status.md)
  110. - [平台真实验收矩阵](docs/acceptance/platform-coverage-matrix-2026-10-21.md)
  111. - [链接覆盖基线与复测](docs/product/link-coverage-benchmark.md)
  112. - [登录扫码交接](docs/product/qr-signin-handoff.md)
  113. - [版本记录](CHANGELOG.md)
  114. - [安全与隐私说明](NOTICE)
  115. <a id="english"></a>
  116. ## English
  117. <p align="center"><a href="#top">中文</a></p>
  118. <h3 align="center">Distill any link</h3>
  119. <p align="center"><strong>Share a link with your AI</strong></p>
  120. <p align="center">
  121. Douyin / Bilibili video / image posts, WeChat articles, Xiaohongshu notes, PDF / video — one link, every artifact you wanted out of it: a summary, Markdown, a wiki, a knowledge base.
  122. </p>
  123. Or write nothing at all: when a video or a note is worth keeping, hand the link to the AI and let it work from the material itself as source and reference. It is read as ordered segments with keyframes and ASR transcripts locatable to `[mm:ss]`, so the AI grasps everything that matters on far fewer tokens instead of chewing through the title, the recommendation rail and the danmaku.
  124. Verified on **Claude Code**, **Codex**, and **DeepSeek Harness**.
  125. > Same link, same model, with and without Envoy: over 60 turns per arm, ends normally with the gaps named 22% -> 92%, runs out of time with no conclusion 68% -> 0%. Completion is suspected rather than human-graded; method and raw records: [evaluation](#evaluation).
  126. ### Install
  127. One line per host. The installer only needs `curl`: it installs uv and Python 3.11 itself and prints `envoy install: runtime … protocol … browser …` when done.
  128. ```sh
  129. # Claude Code
  130. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- claude-code
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 199 on GitHub
  131. # Codex
  132. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- codex
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 202 on GitHub
  133. # DeepSeek Harness
  134. curl -fsSL https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.sh | sh -s -- dsh
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 205 on GitHub
  135. ```
  136. The default is the full profile: runtime + Chromium + the Whisper `base` weights, about 1.0 GB (178 + 683 + 141 MB), so every capability works offline after install and the first read never waits on a download. Chromium accounts for 1-2 minutes of it, the weights follow; both steps are idempotent and re-running the installer fills in anything missing. On a small disk, or in CI and containers, pass `--profile minimal`: the runtime alone (about 178 MB - plain pages, PDFs, images), where the first page that needs rendering reports `envoy browser setup` (a one-time ~683 MB Chromium download) and the first transcript reports `envoy asr install --model base` (~141 MB of GGML weights when whisper.cpp is present). `--profile standard` (runtime + Chromium, about 860 MB) sits between them. [INSTALL.md](integrations/skills/envoy/INSTALL.md) has the size accounting and platform differences.
  137. Not sure which host you have: use `auto` instead — it detects an installed Claude Code / Codex / DSH and falls back to an MCP entry (`mcp` and `skill` also install on their own; `all` is the same as `auto`). Add `ENVOY_DRY_RUN=1` to see what it would do first. To install from a mirror or a local directory, point `ENVOY_RELEASES_BASE` at the download root and `ENVOY_MANIFEST_URL` at the manifest (layout: `v<version>/` with the artifacts and `SHA256SUMS`); the defaults are unchanged; to point at a fork set `ENVOY_REPO` (default `regrevia/Envoy`). On Windows PowerShell: `irm https://raw.githubusercontent.com/regrevia/Envoy/main/integrations/install.ps1 | iex` (auto by default; for one host download it and run `powershell -ExecutionPolicy Bypass -File install.ps1 dsh`).
  138. Two checks after installing:
  139. ```sh
  140. envoy doctor --json # runtime.install.status should be pinned; the minimal profile starts with browser.ready and asr_model_available false, and both turn true once fetched
  141. envoy browser setup # the minimal install has no Chromium; this fetches it (~683 MB), and repairs or changes versions, pruned stale copies included
  142. envoy asr install --model base # the minimal install has no Whisper weights; this fetches them (~140 MB)
  143. envoy browser prune # reclaim browser downloads left behind by upgrades; --dry-run only reports
  144. ```
  145. Upgrading means running the same line again. Uninstall, mirrors, and each host's plugin directory: [INSTALL.md](integrations/skills/envoy/INSTALL.md).
  146. ### What it does
  147. **Keyframes** — pull slides, charts, and demo frames at chosen timestamps and hand them to the AI, no scrubbing.
  148. **Video** — captions first; when there are none, run local Whisper over the requested range and return transcript segments with `[mm:ss]`.
  149. **Text and images** — web pages, articles, PDFs, image posts become clean text, images, and their source.
  150. **Knowledge base / wiki** — read a long video end to end, group it by theme, put a timestamp on every claim, and drop the result into Obsidian or Notion.
  151. ### How to use it
  152. Ask your AI:
  153. > Turn this video into a knowledge base: `https://www.bilibili.com/video/BV...`
  154. >
  155. > Read the first 10 minutes, group sections by theme, and add `[mm:ss]` after each claim. List anything not covered under "To fill"; do not guess.
  156. The material arrives already segmented and located, so the output can cite directly:
  157. ```markdown
  158. ## 2. Why use a fixed wheel
  159. - A fixed wheel splits a continuous action into reusable stages for per-stage debugging. [04:18]
  160. - The author compares three approaches and states the trade-offs. [07:02]
  161. - This section shows no measured data; to fill. [09:31]
  162. ```
  163. ### How it works
  164. 1. **Fetch** — open the link and take the title, body, captions or transcript, PDF text, and images.
  165. 2. **Locate** — split it into segments carrying `video_time` or source coordinates, and return a material id.
  166. 3. **Continue on demand** — `envoy_search` finds a term, `envoy_expand` fills the unread range; no second full fetch.
  167. 4. **Report gaps** — no captions, login required, missing modality, exhausted download budget: marked as uncovered rather than written up as a conclusion.
  168. The browser runs headlessly and asks you to take over only for login or verification. Access is an Agent Skill plus MCP server; Claude Code, Codex, and DeepSeek Harness each have a plugin (Codex 0.156+ needs one `envoy hook install --host codex`).
  169. ### Scope
  170. No bypassing login, CAPTCHA, regional restrictions, or DRM, and no replacement for web search. Envoy is responsible only for the ranges it actually read and returned. Douyin creator collections and Xiaohongshu image posts follow the [platform acceptance matrix](docs/acceptance/platform-coverage-matrix-2026-10-21.md).
  171. ### Evaluation
  172. Same link, same model, the only difference is whether Envoy is installed: 20 task packages × 3 questions per arm (60 turns each) on a frozen package (`benchmarks/fixed-set-v1`).
  173. | What happens | Without Envoy | With Envoy |
  174. | --- | --- | --- |
  175. | Ends normally without saying it could not read (suspected complete, unverified) | 22% (13/60) | **92% (55/60)** |
  176. | Times out with no answer | 68% (41/60) | **0%** |
  177. | Answer hits at least one reference-fact keyword | 23% (14/60) | **82% (49/60)** |
  178. | Share of reference-fact keywords hit | 9.1% (47/516) | **38.2% (197/516)** |
  179. | Time to the first piece of evidence (median) | 88 s | **48 s** (45% less) |
  180. | Answer length (median) | 3,125 chars | **8,142 chars** (2.6×) |
  181. | Empty answers (a separate 60-pair run) | 31 | **0** |
  182. Percentages are within-arm shares. The arms ran under different caps (900 s vs 2400 s), which inflates the completion gap; `verified` is 0 for both, meaning no human correctness review, and `suspect_complete` only means "exited normally without saying it could not read".
  183. Scoring definitions, deviations and the review record are in the [evaluation method](benchmarks/README.md), the [protocol and deviations](benchmarks/fixed-set-v1/AB-PROTOCOL.md) and the [review record](docs/acceptance/envoy-read-failure-2026-10-25.md).
  184. ### Link coverage
  185. 123 real links (news, WeChat posts, Bilibili, podcasts, PDFs, forums, code docs) — three questions:
  186. | Question | Answer |
  187. | --- | --- |
  188. | Out of 100 links, how many give usable material | **80 (80.5%)** |
  189. | Out of the 100 that answered at all, how many yielded body text | **93 (92.5%)** |
  190. | How long an unreachable link makes you wait | **26 s** (was 77 s, 67% less) |
  191. "Usable" means quotable in full, or partially read with the gap named; a failure states why (login, bot check, network) instead of guessing.
  192. Of the 24 that were not usable, 16 hosts were unreachable — all overseas, blocked at the local proxy exit — and 8 sat behind a login or bot-check wall. Under the conditions of this run neither case reads with or without Envoy installed; Envoy says so and the agent takes another route.
  193. With vs without Envoy: see [Evaluation](#evaluation). Rerun with `scripts/acceptance/link_coverage.py`; reports and definitions are in the [link-compatibility experiment](docs/experiments/link-compatibility.md).
  194. ### Local web interface (Docker)
  195. **The plugin and skills do not need Docker.** All three installs (Claude Code, Codex, DSH) run locally: use the `install.sh` / `install.ps1` installer, or download `envoy_cli-<版本>-py3-none-any.whl` from the GitHub Release and hand it to `uv tool install`, plus a skills directory or plugin package (PyPI has no published package yet) — no container, no compose. Docker only serves the legacy web interface in this repo (FastAPI + Next.js + MySQL) for browsing tasks and artifacts on your own machine; skip it entirely if you do not want that UI.
  196. With Docker Desktop running, from the repository root:
  197. ```bash
  198. cp .env.example .env
  199. docker compose up -d --build
  200. ```
  201. The interface is at `http://127.0.0.1:8200`; `http://127.0.0.1:8140/api/v1/health` answers `{"status":"ok"}`. The stack is ready when `mysql`, `redis`, `backend` and `frontend` all report `healthy`; MySQL bootstraps the schema from `sql/` on first start, and uploads plus exports stay on the `envoy_runtime` volume, so no object storage is required. The first build pulls MySQL 8.4 and Redis and installs FFmpeg, Tesseract and Chromium into the backend image, so it takes a few minutes. `docker compose down` keeps the volumes; only `down -v` drops the database and local files. Ports, passwords and model settings live in `.env`; build acceleration and troubleshooting are in [Docker quickstart](docs/quickstart-docker.md).
  202. ### Documentation
  203. - [Installation, updates, and MCP setup](integrations/skills/envoy/INSTALL.md)
  204. - [Claude Code plugin](integrations/claude-code/envoy/README.md)
  205. - [Codex plugin](integrations/codex/envoy/README.md)
  206. - [DeepSeek Harness plugin](integrations/dsh/README.md)
  207. - [Product and engineering status](docs/architecture/agent-link-reader-status.md)
  208. - [Platform acceptance matrix](docs/acceptance/platform-coverage-matrix-2026-10-21.md)
  209. - [Link coverage baseline and re-run](docs/product/link-coverage-benchmark.md)
  210. - [Sign-in QR handoff](docs/product/qr-signin-handoff.md)
  211. - [Changelog](CHANGELOG.md)
  212. - [Security and privacy](NOTICE)

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.