dsh-vision-router 1.0.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 dsh-vision-router contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,268 @@
+<p align="center">
+  <img src="assets/hero.svg" width="100%" alt="DSH Vision Router — eyes for text-only DeepSeek Harness agents" />
+</p>
+
+<h1 align="center">dsh-vision-router</h1>
+
+<p align="center"><strong>Paste an image and it just works — eyes for text-only agents on DeepSeek Harness. Free out of the box, no key, no Python, one command.</strong></p>
+
+<p align="center">DeepSeek keeps thinking; the built-in free vision chain and ten pixel-level tools do the seeing. Image turns behave like ordinary tool-calling turns — grounded, measurable, repeatable.</p>
+
+<p align="center">
+  <a href="https://awesome-dsh-plugin.com"><img src="https://awesome-dsh-plugin.com/badge.svg" alt="awesome · DSH plugin" /></a>
+  <a href="https://github.com/ysr666/dsh-vision-router/releases/tag/v1.0.0"><img src="https://img.shields.io/badge/release-v1.0.0-5B4CF0?style=flat-square" alt="Release v1.0.0" /></a>
+  <a href="tests"><img src="https://img.shields.io/badge/verified-86%20tests-2EA44F?style=flat-square" alt="Verified: 86 tests" /></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-2EA44F?style=flat-square" alt="License: MIT" /></a>
+  <a href="package.json"><img src="https://img.shields.io/badge/Node.js-%3E%3D22-339933?style=flat-square&amp;logo=nodedotjs&amp;logoColor=white" alt="Node.js >=22" /></a>
+  <img src="https://img.shields.io/badge/runtime-no%20Python-8A2BE2?style=flat-square" alt="No Python" />
+  <a href="cordis.patch.yml"><img src="https://img.shields.io/badge/DSH-Web%20profile-5B4CF0?style=flat-square" alt="DSH Web profile" /></a>
+</p>
+
+<p align="center">English · <a href="README.zh.md">中文</a></p>
+
+<p align="center">
+  <img src="assets/vision-demo.gif" width="640" alt="Demo: paste an image, the agent locates the send button with vision_ground / vision_crop / vision_pixel_diff and answers with coordinates" />
+</p>
+
+## Why this exists
+
+Most DSH vision plugins bridge images to DeepSeek as *text descriptions* — lossy, one-shot, and blind to pixels. This plugin keeps the **original pixels on the vision model's side** and DeepSeek on the reasoning side, and makes looking at an image an **ordinary tool call**:
+
+- **One command install.** The package ships its own composition patch (`dsh.bundle.patch`): `dsh plugin add` wires the row, the admission wrapper, the stealth takeover and the attachment limits automatically — zero manual file edits.
+- **Free by default.** The vision chain starts with a built-in OVHcloud anonymous endpoint (`Qwen2.5-VL-72B-Instruct`, no account, no key, 2 req/min per IP). Paid chains (OpenRouter, Pi-AI providers, direct OpenAI-compatible endpoints) are optional upgrades.
+- **No Python.** The whole pipeline — downscale, grounding, crop, pixel diff, palette, OCR, SVG trace, cutout, HTML screenshot — runs on sharp / potrace / tesseract / system Chrome.
+- **Continuous multi-step image work.** An image turn is a text turn that calls tools: `vision_ground` → `vision_crop` → `vision_describe` → `vision_pixel_diff` → fix → screenshot again. The agent keeps iterating until the work is done.
+- **DeepSeek stays the brain.** Text turns are untouched in model, cost and context. The vision model is only the eyes, called on demand; answers are cached by image content.
+- **Transparent to the user.** Uploaded images keep rendering as images in the conversation UI; the rewrite that points the model at the vision tools happens only inside the model call, never in the session log.
+
+## How it compares
+
+The closest alternative is [@anionex/dsh-vision-toolkit](https://github.com/Anionex/dsh-vision-toolkit) (Anionex), a native DSH bundle of the well-known `agent-vision-toolkit` lineage. Both packages ship a `vision-tools` skill and a family of pixel-level tools; they differ in philosophy — **zero-config paste-and-go** versus **agent-driven visual engineering**:
+
+| | dsh-vision-router | @anionex/dsh-vision-toolkit |
+|---|---|---|
+| Image Q&A out of the box | ✅ Built-in free chain (anonymous OVHcloud endpoint) — no account, no key | Requires your own vision API key (local pixel tools work without one) |
+| Runtime | ✅ Node only — no Python | Python 3.11+ managed runtime |
+| Getting an image in | ✅ Paste it — the turn auto-routes to the vision chain and auto-mounts the tools | Workspace path + `/vision-tools` command, then explicit tool calls |
+| Turn routing | ✅ Image turns switch to vision, text turns switch back to DeepSeek — stealth takeover, the model picker looks stock | Tool-driven; no whole-turn auto-routing |
+| Profiles | Web | Web + Headless |
+| Playbooks | The pixel loop: ground → crop → diff → fix → screenshot again | Richer case library (long-screenshot OCR, UI restoration, GUI automation) |
+| Tests | 86 | 162 |
+| Install | One command | One command (npm) |
+
+Both are MIT-licensed and one command away. Pick this plugin when you want images to *just work* with zero setup; pick theirs when you need headless profiles or the extended playbook library. (Feature comparison reflects their README as of 2026-08.)
+
+## Quick start
+
+```sh
+dsh plugin --profile web add github:ysr666/dsh-vision-router
+```
+
+Restart `dsh web` — done. Zero configuration:
+
+- the plugin's bundle patch mounts the row, takes over the official DeepSeek route (stealth mode — the model picker looks exactly like stock), and relaxes attachment limits to 20 MB / 100 MP;
+- the default vision chain is the built-in free endpoint;
+- every setting is editable live in **Settings → Plugins → Plugin config → 视觉路由（自动识图）**.
+
+Then just paste an image into a conversation. The agent mounts the vision tools automatically and looks at it through `vision_describe` (and friends) — multi-step if needed.
+
+### See it in action
+
+*Left: an image turn — the user sends a picture, the agent calls `vision_describe` through the free chain and answers. Right: the finished structured answer.*
+
+<p align="center">
+  <img src="assets/dsh-conversation-image-qa.png" width="49%" alt="A conversation turn in which the agent looks at an uploaded image through vision_describe." />
+  <img src="assets/dsh-conversation-image-qa-result.png" width="49%" alt="The agent's structured answer describing the image content." />
+</p>
+
+## Highlights
+
+- **Original pixels, real answers.** The vision chain reads the image at original resolution (auto-downscaled only to protect latency/quota); the agent's question travels with the image, so answers are about *your* question, not a generic description.
+- **Automatic failover with classified errors.** Region blocks, ToS filtering, 402 quota, 429 rate limits (with Retry-After backoff), context overflow, network failures — the chain walks providers one by one and only reports after all of them failed, with actionable advice.
+- **Image memory.** Vision answers are cached by attachment content hash; later text turns substitute the recorded description (marked as untrusted evidence), so DeepSeek genuinely remembers earlier images without re-spending vision calls.
+- **A verifiable pixel loop.** Reference → `vision_html_screenshot` → `vision_pixel_diff` (ratio + red heatmap + worst-region ranking) → fix → repeat until the mismatch converges. UI restoration becomes measurable instead of eyeballed.
+- **Progressive schema exposure.** Only a zero-arg `vision_activate` bootstrap is always visible; image turns auto-mount all ten deep tools with a one-time usage note, and a `vision-tools` skill is registered for text-only turns.
+- **Selective proxy.** Only the configured vision provider hosts go through your local proxy; DeepSeek stays direct.
+
+### Pixel loop in practice
+
+<p align="center">
+  <img src="assets/pixel-loop.png" width="100%" alt="Reference design and final agent rebuild, verified with vision_pixel_diff at 2.54% final difference." />
+</p>
+
+The agent rebuilt the UI from the reference image, then verified the final result with `vision_pixel_diff`: **2.54% final diff** (32,939 / 1,296,000 differing pixels, threshold 16/channel).
+
+## How it works
+
+<p align="center">
+  <img src="assets/how-it-works.svg" width="100%" alt="How DSH Vision Router keeps DeepSeek as the brain and vision tools as the eyes." />
+</p>
+
+The vision model is **only the eyes**; DeepSeek is **always the brain**. An image turn is never hijacked by a one-shot vision answer — the agent drives the tools itself and can keep operating on the image across as many steps as the task needs.
+
+## Tools
+
+All ten deep tools mount automatically on image turns (`autoActivateOnImage`); text turns can mount them via `vision_activate` or the `/vision-tools` skill. Built on sharp / potrace / tesseract / system Chrome — no Python:
+
+<p align="center">
+  <img src="assets/vision-tools.svg" width="100%" alt="Nine vision tools available in DSH Vision Router." />
+</p>
+
+| Tool | What it does | Artifact |
+|---|---|---|
+| `vision_describe` | Image Q&A / multi-image compare / structured-evidence JSON mode (summary + layout regions + entity inventory + verbatim transcription) | — |
+| `vision_ground` | Locate a target → **original-pixel box x1/y1/x2/y2** | annotated PNG (optional) |
+| `vision_detect` | Numbered inventory of every element of a kind (buttons/inputs/links…) with original-pixel boxes | annotated PNG with numbered boxes |
+| `vision_crop` | Crop and zoom into a pixel box | PNG |
+| `vision_pixel_diff` | Per-pixel comparison: diff ratio + worst 8×8-grid regions | red heatmap PNG + JSON report |
+| `vision_colors` | Dominant colors (hex + share) | — |
+| `vision_ocr` | Text transcription: local tesseract (chi_sim+eng) first, vision model fallback | — |
+| `vision_trace` | SVG vectorization (potrace posterization; icons/logos) | SVG |
+| `vision_extract_foreground` | Cutout via border flood fill (uniform backgrounds) | transparent PNG |
+| `vision_html_screenshot` | Screenshot a local HTML file (headless system Chrome) | PNG |
+
+Formats are sniffed from magic bytes, so extensionless content-addressed attachment files work everywhere (no `.png` renaming needed).
+
+**Common workflows**
+
+```text
+vision_ground image="ref.png" target="the send button"
+vision_detect image="page.png" target="input fields"
+vision_crop   image="ref.png" region="1067,841,1108,881"
+vision_describe paths=["ref.png","impl.png"] question="list the differences" json=true
+vision_pixel_diff original="ref.png" rebuilt="screenshot.png"
+vision_ocr image="screenshot.png"
+vision_colors image="ref.png" top=8
+vision_trace image="icon.png" steps=4
+vision_extract_foreground image="logo.png"
+vision_html_screenshot source="page.html" width=1200 height=720
+```
+
+## Provider fallback chain
+
+The vision chain walks providers in order and only surfaces an error after every one failed:
+
+1. the **built-in free endpoint** (`vision-http` → `ovh/Qwen2.5-VL-72B-Instruct`) — no key, best-effort, 2 req/min per IP;
+2. configured `httpProviders` (direct OpenAI-compatible endpoints with optional `apiKeyEnv`);
+3. configured `providers` / `provider` + `fallbacks` (any adapter-backed provider, e.g. a Pi-AI profile like OpenRouter or Zhipu).
+
+> In the legacy `routing: true` mode, the whole-turn chain walks only `provider + fallbacks` — `httpProviders` (including the free fallback) do not participate there. The default `routing: false` (tools-first) tries everything.
+
+Failures are classified (region / tos / quota / rate-limit / context / network) and the final error carries advice; `429` responses honor `Retry-After` once with a capped backoff. Oversized uploads are downscaled before the call (default budget 4 MP) to keep tool calls fast.
+
+## Stealth mode
+
+A default install takes over the official `deepseek-official` route: the model picker looks exactly like stock (same DeepSeek group, same model names), but each entry is the auto-vision wrapper that declares image input and delegates text turns to a rebuilt native DeepSeek adapter (same `llm-deepseek` settings section and credentials). Old sessions keep working through the hidden `deepseek-vision` alias.
+
+To keep the stock row instead, override it in your profile patch layer (`~/.dsh/profiles/<profile>/cordis.patch.yml`):
+
+```yaml
+- id: llm-deepseek
+  name: '@deepseek-ai/dsh-llm-deepseek'
+```
+
+With the stock row present, the plugin falls back to the visible "DeepSeek + 自动识图" wrapper entry — pick it in the model picker for image turns. Recovery from a broken install is the same one-line override.
+
+## Web settings
+
+The Web profile registers a **视觉路由（自动识图）** card under **Settings → Plugins → Plugin config**, styled like the built-in cards. It live-edits:
+
+- switches: whole-turn legacy routing, vision tools, image-block rewriting, stealth;
+- vision request timeout, wrapper/chain route names;
+- the **vision chain** (one `provider/model` per line, top-down fallback) and the text model;
+- every field shows an "overridden" badge with a one-click reset to the composition default, plus discard/save.
+
+<p align="center">
+  <img src="assets/vision-settings.png" width="72%" alt="The vision-router card in Settings > Plugins > Plugin config." />
+</p>
+
+> PR [#8](https://github.com/ysr666/dsh-vision-router/pull/8) upgrades the panel with catalog-driven model dropdowns, add/remove fallback rows, and proxy settings.
+
+## Configuration
+
+Everything is optional; defaults work out of the box. Edit via the Web card or a profile patch:
+
+| Field | Default | Meaning |
+|---|---|---|
+| `provider` / `model` | `vision-http` / `ovh/Qwen2.5-VL-72B-Instruct` | shorthand chain (adapter-backed provider + model) |
+| `fallbacks` | `[]` | backup models for the shorthand provider |
+| `providers` | `[]` | multi-provider chain `{ provider, model, fallbacks[] }`, tried in order; wins over the shorthand |
+| `httpProviders` | built-in OVH entry | direct OpenAI-compatible endpoints `{ name, baseURL, model, apiKeyEnv, maxTokens }` |
+| `routing` | `false` | legacy whole-turn chain routing (one-shot answer). `false` = tools-first flow (recommended) |
+| `reverseRouting` | `true` | with `routing: true`, route text turns back to `textProvider` |
+| `wrapperRoute` / `chainRoute` | `deepseek-vision` / `vision-chain` | admission wrapper route name / fallback chain route name (empty disables) |
+| `stealth` | `true` | take over the official `deepseek-official` route |
+| `textProvider` | `deepseek-official` / `deepseek-v4-pro` | the model that reasons (your daily model) |
+| `tool` / `progressiveTools` / `autoActivateOnImage` | `true` ×3 | vision tools on / progressive mounting / auto-mount on image turns |
+| `rewriteImages` | `true` | rewrite image blocks in the model input (cached description or tool-hint marker); the UI log keeps images |
+| `downscale` / `downscaleMaxPixels` | `true` / `4000000` | pre-call downscale and its pixel budget (latency guard) |
+| `cache` / `cacheTtlSeconds` / `cacheMaxEntries` | `true` / `3600` / `200` | vision answer cache |
+| `timeoutMs` | `120000` | per vision call deadline |
+| `artifactsDir` | `.dsh-vision-router/artifacts` | artifact directory (relative to the session workspace) |
+| `proxy` / `proxyHosts` | `''` / openrouter hosts | optional proxy for vision provider hosts only |
+
+## Requirements
+
+- DeepSeek Harness with a Web profile and `pnpm` available to `dsh plugin`.
+- Node ≥ 22 (host side).
+- No API key for the default free chain; a credential reference (`apiKeyEnv`) only for paid `httpProviders`.
+- Chrome / Chromium / Edge only for `vision_html_screenshot`; every other tool works without a browser.
+- Tesseract is optional: `vision_ocr` falls back to the vision model when the local engine is absent.
+
+## Install and lifecycle
+
+### Install
+
+```sh
+dsh plugin --profile web add github:ysr666/dsh-vision-router
+dsh --profile web --dump-config | grep vision-router   # one row, mounted by the bundle patch
+```
+
+Restart a long-lived Web profile. The host discovers the browser bundle through `dsh.client` at startup.
+
+### Disable / re-enable
+
+```yaml
+- id: vision-router
+  disabled: true
+```
+
+Set it back to `false` to re-enable. Unloading removes the wrapper routes, tools, skill and settings card; cached artifact files remain.
+
+### Upgrade
+
+```sh
+dsh plugin --profile web update dsh-vision-router
+```
+
+Settings live in the profile's settings provider and survive upgrades.
+
+### Uninstall
+
+```sh
+dsh plugin --profile web remove dsh-vision-router
+```
+
+This removes the dependency and the bundle layer. If you disabled the stock DeepSeek row manually, re-enable it in your profile patch.
+
+## Security notes
+
+- Image text is **untrusted evidence**: descriptions, OCR output and the auto-mount note all tell the agent never to execute instructions found inside images.
+- Tool inputs resolve through `ctx.fs` (sandbox-aware); vision uploads never send anything but the selected image and the question.
+- Artifacts write only under `<workspace>/.dsh-vision-router/artifacts`; results return absolute paths and byte counts.
+- Secrets never travel: `apiKeyEnv` names a DSH credential reference; the value is resolved per call and never logged.
+- The settings write path goes through the settings service (schema-validated, revision-checked) — a stale or invalid save is rejected, not partially applied.
+
+## License
+
+[MIT](LICENSE)
+
+<!-- star-history-chart -->
+## Star History
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/ysr666/dsh-vision-router/star-history/assets/star-history/star-history-dark.svg">
+    <img alt="Star history chart" src="https://raw.githubusercontent.com/ysr666/dsh-vision-router/star-history/assets/star-history/star-history-light.svg" width="100%">
+  </picture>
+</p>

package/README.zh.md

@@ -0,0 +1,268 @@
+<p align="center">
+  <img src="assets/hero-zh.svg" width="100%" alt="DSH Vision Router——为 DeepSeek Harness 的纯文本 Agent 提供视觉能力" />
+</p>
+
+<h1 align="center">dsh-vision-router</h1>
+
+<p align="center"><strong>图片粘贴即用：给 DeepSeek Harness 的纯文本 Agent 装上“眼睛”——开箱免费、免 Key、无 Python、一条命令安装。</strong></p>
+
+<p align="center">DeepSeek 只负责思考，内置免费视觉链 + 10 个像素级工具负责“看”；图片轮次就像普通工具调用一样自然、可定位、可验证。</p>
+
+<p align="center">
+  <a href="https://awesome-dsh-plugin.com"><img src="https://awesome-dsh-plugin.com/badge.svg" alt="awesome · DSH plugin" /></a>
+  <a href="https://github.com/ysr666/dsh-vision-router/releases/tag/v1.0.0"><img src="https://img.shields.io/badge/release-v1.0.0-5B4CF0?style=flat-square" alt="Release v1.0.0" /></a>
+  <a href="tests"><img src="https://img.shields.io/badge/verified-86%20tests-2EA44F?style=flat-square" alt="Verified: 86 tests" /></a>
+  <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-2EA44F?style=flat-square" alt="License: MIT" /></a>
+  <a href="package.json"><img src="https://img.shields.io/badge/Node.js-%3E%3D22-339933?style=flat-square&amp;logo=nodedotjs&amp;logoColor=white" alt="Node.js >=22" /></a>
+  <img src="https://img.shields.io/badge/runtime-no%20Python-8A2BE2?style=flat-square" alt="No Python" />
+  <a href="cordis.patch.yml"><img src="https://img.shields.io/badge/DSH-Web%20profile-5B4CF0?style=flat-square" alt="DSH Web profile" /></a>
+</p>
+
+<p align="center"><a href="README.md">English</a> · 中文</p>
+
+<p align="center">
+  <img src="assets/vision-demo.gif" width="640" alt="演示：粘贴图片，Agent 用 vision_ground / vision_crop / vision_pixel_diff 定位发送按钮并给出坐标" />
+</p>
+
+## 为什么做这个
+
+大多数 DSH 视觉插件把图片“翻译”成一段文字描述再喂给 DeepSeek——有损、一次性、看不见像素。本插件把**原图像素留在视觉模型侧**、把推理留在 DeepSeek 侧，并把“看图”变成一次**普通的工具调用**：
+
+- **一条命令安装。** 包自带组合补丁（`dsh.bundle.patch`）：`dsh plugin add` 自动完成插件行挂载、准入包装、隐身接管与附件限制放宽——不用手改任何文件。
+- **默认免费。** 视觉链内置 OVHcloud 匿名端点（`Qwen2.5-VL-72B-Instruct`，免注册、免 Key，每 IP 2 次/分钟）。付费链路（OpenRouter、Pi-AI 供应商、任意 OpenAI 兼容直连端点）是可选升级。
+- **无 Python。** 整条管线——缩放、定位、裁剪、像素对比、取色、OCR、SVG 矢量化、抠图、HTML 截图——全部基于 sharp / potrace / tesseract / 系统 Chrome。
+- **可连续多步看图。** 图片轮 = 调用工具的文本轮：`vision_ground` → `vision_crop` → `vision_describe` → `vision_pixel_diff` → 修复 → 再截图，Agent 可以一直迭代到任务完成。
+- **DeepSeek 始终是大脑。** 文字轮在模型、成本、上下文上完全不动；视觉模型只当“眼睛”、按需调用，答案按图片内容缓存。
+- **界面无感。** 上传的图片在会话界面里照常显示为图片；指向视觉工具的改写只发生在模型输入层，从不写入会话日志。
+
+## 对比同类插件
+
+最接近的同类是 [@anionex/dsh-vision-toolkit](https://github.com/Anionex/dsh-vision-toolkit)（Anionex），它是知名 `agent-vision-toolkit` 系列的 DSH 原生版。两者都提供 `vision-tools` 技能和一组像素级工具，区别在理念：**零配置粘贴即用** vs **Agent 主导的视觉工程**：
+
+| | dsh-vision-router | @anionex/dsh-vision-toolkit |
+|---|---|---|
+| 开箱图片问答 | ✅ 内置免费视觉链（OVHcloud 匿名端点），免注册免 Key | 远程工具需自备视觉 API Key（本地像素工具免 Key） |
+| 运行时 | ✅ 纯 Node，无需 Python | 需要 Python 3.11+ 受管运行时 |
+| 图片怎么进来 | ✅ 直接粘贴——轮次自动切视觉链并自动挂载工具 | 工作区路径 + `/vision-tools` 命令，再显式调用工具 |
+| 轮次路由 | ✅ 图片轮切视觉、文本轮切回 DeepSeek——隐身接管，模型选择器与官方一致 | 工具驱动，无整轮自动路由 |
+| 支持 profile | Web | Web + Headless |
+| 玩法库 | 像素循环：定位 → 裁剪 → 对比 → 修复 → 再截图 | 更丰富的案例库（长截图 OCR、UI 还原、GUI 自动化） |
+| 测试 | 86 | 162 |
+| 安装 | 一条命令 | 一条命令（npm） |
+
+两者都是 MIT 许可、一条命令安装。想要图片**粘贴即用**、零配置就选本插件；需要 Headless 部署或更丰富的案例库，可以看 @anionex/dsh-vision-toolkit。（功能对比以其 README 2026-08 状态为准。）
+
+## 快速开始
+
+```sh
+dsh plugin --profile web add github:ysr666/dsh-vision-router
+```
+
+重启 `dsh web`——完成，零配置：
+
+- 插件的 bundle 补丁自动挂载插件行、接管官方 DeepSeek 路由（隐身模式——模型选择器看起来和原版一模一样）、放宽附件限制到 20MB / 1 亿像素；
+- 默认视觉链就是内置免费端点；
+- 全部配置可在 **设置 → 插件 → 插件配置 → 视觉路由（自动识图）** 实时修改。
+
+然后直接往对话里贴一张图。Agent 自动挂载视觉工具，通过 `vision_describe`（以及其余 8 个工具）看图，需要时连续多步。
+
+### 实际效果
+
+*左：一次图片轮——用户发图，Agent 通过免费链路调用 `vision_describe` 并作答。右：最终的结构化解读。*
+
+<p align="center">
+  <img src="assets/dsh-conversation-image-qa.png" width="49%" alt="Agent 通过 vision_describe 查看上传图片的对话轮次。" />
+  <img src="assets/dsh-conversation-image-qa-result.png" width="49%" alt="Agent 对图片内容的结构化解读。" />
+</p>
+
+## 亮点
+
+- **原图像素，真实答案。** 视觉链按原始分辨率读图（仅为保护延迟/额度自动缩放）；你的问题随图一起发送，答案围绕*你的问题*，而不是一段泛泛的描述。
+- **自动降级 + 分类报错。** 地区限制、ToS 风控、402 额度、429 限流（尊重 Retry-After 退避重试）、上下文超长、网络故障——链路逐供应商尝试，全部失败才报错并给出可操作的建议。
+- **图片记忆。** 视觉答案按附件内容哈希缓存；后续文字轮用记录的描述替换历史图片（标注为不可信证据），DeepSeek 真正“记得”之前发过的图，且不重复消耗视觉调用。
+- **可验证的像素闭环。** 参照图 → `vision_html_screenshot` → `vision_pixel_diff`（差异率 + 红色热力图 + 最差区域排行）→ 修复 → 再对比，直到差异收敛。UI 还原从“目测”变成“实测”。
+- **渐进式 schema 暴露。** 平时只有一个零参引导工具 `vision_activate`；图片轮自动挂载全部 10 个深看工具（附一次性使用提示），并为纯文字轮注册 `vision-tools` 技能。
+- **选择性代理。** 只有配置的视觉供应商域名走本地代理；DeepSeek 保持直连。
+
+### 像素闭环实测
+
+<p align="center">
+  <img src="assets/pixel-loop-zh.png" width="100%" alt="参考设计与 Agent 最终复刻，通过 vision_pixel_diff 实测最终差异为 2.54%。" />
+</p>
+
+Agent 仅根据参考图复刻 UI，再用 `vision_pixel_diff` 验证最终结果：**最终差异 2.54%**（32,939 / 1,296,000 个差异像素，threshold 16/channel）。
+
+## 工作原理
+
+<p align="center">
+  <img src="assets/how-it-works-zh.svg" width="100%" alt="DSH Vision Router 的工作原理：DeepSeek 作为大脑，视觉工具作为眼睛。" />
+</p>
+
+视觉模型**只当眼睛**，DeepSeek **始终是大脑**。图片轮永远不会被一次性视觉答案“劫持”——Agent 自己驱动工具，可以跨多个步骤持续对同一张图操作。
+
+## 工具
+
+10 个深看工具在图片轮自动挂载（`autoActivateOnImage`）；文字轮可通过 `vision_activate` 或 `/vision-tools` 技能挂载。全部基于 sharp / potrace / tesseract / 系统 Chrome——无 Python：
+
+<p align="center">
+  <img src="assets/vision-tools-zh.svg" width="100%" alt="DSH Vision Router 的 10 个视觉工具。" />
+</p>
+
+| 工具 | 作用 | 产物 |
+|---|---|---|
+| `vision_describe` | 看图问答 / 多图对比 / 结构化证据 JSON 模式（摘要 + 布局区域 + 实体清单 + 原文转写） | — |
+| `vision_ground` | 定位目标 → **原图像素框 x1/y1/x2/y2** | 标注 PNG（可选） |
+| `vision_detect` | 盘点某类元素（按钮/输入框/链接…）→ 编号清单 + 原图像素框 | 编号标注 PNG |
+| `vision_crop` | 按像素框裁剪放大 | PNG |
+| `vision_pixel_diff` | 逐像素对比：差异率 + 最差 8×8 网格区域 | 红色热力图 PNG + JSON 报告 |
+| `vision_colors` | 主色提取（十六进制 + 占比） | — |
+| `vision_ocr` | 文字转写：本地 tesseract（中英）优先，视觉模型兜底 | — |
+| `vision_trace` | SVG 矢量化（potrace 分色；图标/logo） | SVG |
+| `vision_extract_foreground` | 边界洪泛抠图（纯色背景） | 透明 PNG |
+| `vision_html_screenshot` | 给本地 HTML 文件截图（无头系统 Chrome） | PNG |
+
+图片格式按**魔数识别**，无扩展名的内容寻址附件文件也能直接用（不用再复制成 `.png`）。
+
+**常用流程**
+
+```text
+vision_ground image="ref.png" target="发送按钮"
+vision_detect image="page.png" target="输入框"
+vision_crop   image="ref.png" region="1067,841,1108,881"
+vision_describe paths=["ref.png","impl.png"] question="列出两图的差异" json=true
+vision_pixel_diff original="ref.png" rebuilt="screenshot.png"
+vision_ocr image="screenshot.png"
+vision_colors image="ref.png" top=8
+vision_trace image="icon.png" steps=4
+vision_extract_foreground image="logo.png"
+vision_html_screenshot source="page.html" width=1200 height=720
+```
+
+## 供应商降级链
+
+视觉链按顺序逐个尝试，全部失败才报错：
+
+1. **内置免费端点**（`vision-http` → `ovh/Qwen2.5-VL-72B-Instruct`）——免 Key、尽力而为、每 IP 2 次/分钟；
+2. 配置的 `httpProviders`（OpenAI 兼容直连端点，可选 `apiKeyEnv`）；
+3. 配置的 `providers` / `provider` + `fallbacks`（任何有适配器的供应商，例如 Pi-AI 配置的 OpenRouter 或智谱）。
+
+> 在旧版 `routing: true` 模式下，整轮链只走 `provider + fallbacks`——`httpProviders`（含免费兜底）不参与。默认的 `routing: false`（工具优先）会尝试全部。
+
+失败会分类（地区 / 风控 / 额度 / 限流 / 上下文 / 网络），最终报错附带建议；`429` 会尊重 `Retry-After` 做一次有上限的退避重试。超大上传图在调用前自动压缩（默认预算 400 万像素），保证工具调用不卡。
+
+## 隐身模式
+
+默认安装即接管官方 `deepseek-official` 路由：模型选择器看起来和原版完全一样（同一个 DeepSeek 组、同样的模型名），但每个条目背后都是声明了图片输入的自动识图包装；文字轮交给插件重建的原生 DeepSeek 适配器（读取同一个 `llm-deepseek` 设置段与凭据）。老会话通过隐藏的 `deepseek-vision` 别名继续工作。
+
+想保留官方行，在你的 profile 补丁层（`~/.dsh/profiles/<profile>/cordis.patch.yml`）覆写即可：
+
+```yaml
+- id: llm-deepseek
+  name: '@deepseek-ai/dsh-llm-deepseek'
+```
+
+官方行在场时，插件回退为选择器里可见的「DeepSeek + 自动识图」包装入口——发图前选它即可。安装出问题时同样用这一行恢复。
+
+## Web 设置
+
+Web 配置页在 **设置 → 插件 → 插件配置** 下注册「视觉路由（自动识图）」卡片，样式与内置卡片一致，可实时修改：
+
+- 开关：整轮自动路由（旧模式）、识图工具、图片块改写、隐身模式；
+- 视觉请求超时、包装/链路由名；
+- **视觉模型链**（每行一个 `provider/model`，自上而下降级）与文本模型；
+- 每个字段都有「已覆盖」徽标与一键恢复组合默认，以及放弃/保存。
+
+<p align="center">
+  <img src="assets/vision-settings.png" width="72%" alt="设置 → 插件 → 插件配置 里的视觉路由卡片。" />
+</p>
+
+> PR [#8](https://github.com/ysr666/dsh-vision-router/pull/8) 会把面板升级为目录驱动的模型下拉框、可增删的备用模型行与代理设置。
+
+## 配置项
+
+全部可选，默认即可用。通过 Web 卡片或 profile 补丁修改：
+
+| 字段 | 默认值 | 含义 |
+|---|---|---|
+| `provider` / `model` | `vision-http` / `ovh/Qwen2.5-VL-72B-Instruct` | 简写链路（有适配器的供应商 + 模型） |
+| `fallbacks` | `[]` | 简写供应商的备用模型 |
+| `providers` | `[]` | 多供应商链路 `{ provider, model, fallbacks[] }`，按序尝试；优先于简写形式 |
+| `httpProviders` | 内置 OVH 条目 | OpenAI 兼容直连端点 `{ name, baseURL, model, apiKeyEnv, maxTokens }` |
+| `routing` | `false` | 旧版整轮链路由（一次性整轮回答）。`false` = 工具优先流程（推荐） |
+| `reverseRouting` | `true` | 开启 `routing` 时，文字轮路由回 `textProvider` |
+| `wrapperRoute` / `chainRoute` | `deepseek-vision` / `vision-chain` | 准入包装路由名 / 降级链路由名（置空关闭） |
+| `stealth` | `true` | 接管官方 `deepseek-official` 路由 |
+| `textProvider` | `deepseek-official` / `deepseek-v4-pro` | 负责思考的模型（你的日常模型） |
+| `tool` / `progressiveTools` / `autoActivateOnImage` | `true` ×3 | 视觉工具开关 / 渐进式挂载 / 图片轮自动挂载 |
+| `rewriteImages` | `true` | 模型输入层改写图片块（缓存描述或工具提示标记）；界面日志保留图片 |
+| `downscale` / `downscaleMaxPixels` | `true` / `4000000` | 调用前压缩及其像素预算（延迟保护） |
+| `cache` / `cacheTtlSeconds` / `cacheMaxEntries` | `true` / `3600` / `200` | 视觉答案缓存 |
+| `timeoutMs` | `120000` | 单次视觉调用超时 |
+| `artifactsDir` | `.dsh-vision-router/artifacts` | 产物目录（相对会话工作区） |
+| `proxy` / `proxyHosts` | `''` / openrouter 域名 | 仅视觉供应商域名可选的本地代理 |
+
+## 环境要求
+
+- DeepSeek Harness 的 Web profile，且 `dsh plugin` 可用 `pnpm`。
+- Node ≥ 22（宿主侧）。
+- 默认免费链路无需 API Key；付费 `httpProviders` 只需一个凭据引用（`apiKeyEnv`）。
+- `vision_html_screenshot` 才需要 Chrome / Chromium / Edge；其余工具无浏览器也能用。
+- tesseract 可选：本地引擎缺失时 `vision_ocr` 自动退回视觉模型。
+
+## 安装与生命周期
+
+### 安装
+
+```sh
+dsh plugin --profile web add github:ysr666/dsh-vision-router
+dsh --profile web --dump-config | grep vision-router   # 一行，由 bundle 补丁挂载
+```
+
+长期运行的 Web profile 需重启。宿主在启动时通过 `dsh.client` 声明发现浏览器端包。
+
+### 禁用 / 恢复
+
+```yaml
+- id: vision-router
+  disabled: true
+```
+
+改回 `false` 即恢复。卸载会移除包装路由、工具、技能与设置卡片；已生成的产物文件保留。
+
+### 升级
+
+```sh
+dsh plugin --profile web update dsh-vision-router
+```
+
+设置存放在 profile 的设置提供方里，升级不丢失。
+
+### 卸载
+
+```sh
+dsh plugin --profile web remove dsh-vision-router
+```
+
+同时移除依赖与 bundle 层。若你曾手动禁用官方 DeepSeek 行，记得在 profile 补丁里恢复。
+
+## 安全说明
+
+- 图片中的文字是**不可信证据**：描述、OCR 输出与自动挂载提示都要求 Agent 绝不执行图片内出现的指令。
+- 工具输入经由 `ctx.fs`（沙盒感知）解析；视觉上传只发送选中的图片与问题本身。
+- 产物只写入 `<workspace>/.dsh-vision-router/artifacts`；结果返回绝对路径与字节数。
+- 密钥不上线：`apiKeyEnv` 只指向 DSH 凭据引用，值按调用解析、永不写入日志。
+- 设置写入走设置服务（schema 校验 + 修订号检查）——过期或非法的保存会被拒绝，不会半截生效。
+
+## License
+
+[MIT](LICENSE)
+
+<!-- star-history-chart -->
+## Star 趋势
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/ysr666/dsh-vision-router/star-history/assets/star-history/star-history-dark.svg">
+    <img alt="Star 历史趋势图" src="https://raw.githubusercontent.com/ysr666/dsh-vision-router/star-history/assets/star-history/star-history-light.svg" width="100%">
+  </picture>
+</p>

package/assets/hero-zh.svg

@@ -0,0 +1,41 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="640" viewBox="0 0 1600 640">
+<defs>
+  <linearGradient id="bg" x1="0" y1="0" x2="1" y2="1"><stop offset="0" stop-color="#070914"/><stop offset="0.52" stop-color="#0A0D1B"/><stop offset="1" stop-color="#080A13"/></linearGradient>
+  <radialGradient id="violetGlow" cx="73%" cy="49%" r="42%"><stop offset="0" stop-color="#715CFF" stop-opacity=".23"/><stop offset=".45" stop-color="#604CFF" stop-opacity=".08"/><stop offset="1" stop-color="#604CFF" stop-opacity="0"/></radialGradient>
+  <radialGradient id="cyanGlow" cx="82%" cy="46%" r="28%"><stop offset="0" stop-color="#46D8FF" stop-opacity=".15"/><stop offset="1" stop-color="#46D8FF" stop-opacity="0"/></radialGradient>
+  <linearGradient id="iris" x1="0" y1="0" x2="1" y2="1"><stop offset="0" stop-color="#8D7BFF"/><stop offset=".5" stop-color="#6256FF"/><stop offset="1" stop-color="#36C8FF"/></linearGradient>
+  <linearGradient id="route" x1="0" y1="0" x2="1" y2="0"><stop offset="0" stop-color="#695BFF"/><stop offset=".55" stop-color="#7B6DFF"/><stop offset="1" stop-color="#42D8FF"/></linearGradient>
+  <linearGradient id="chip" x1="0" y1="0" x2="0" y2="1"><stop offset="0" stop-color="#171B2C" stop-opacity=".95"/><stop offset="1" stop-color="#0D1020" stop-opacity=".95"/></linearGradient>
+  <filter id="glow"><feGaussianBlur stdDeviation="5" result="b"/><feMerge><feMergeNode in="b"/><feMergeNode in="SourceGraphic"/></feMerge></filter>
+  <filter id="softGlow"><feGaussianBlur stdDeviation="12" result="b"/><feMerge><feMergeNode in="b"/><feMergeNode in="SourceGraphic"/></feMerge></filter>
+  <pattern id="dots" width="34" height="34" patternUnits="userSpaceOnUse"><circle cx="1.4" cy="1.4" r="1.35" fill="#65708A" opacity=".12"/></pattern>
+</defs>
+<rect width="1600" height="640" rx="32" fill="url(#bg)"/><rect width="1600" height="640" rx="32" fill="url(#dots)" opacity=".45"/><rect width="1600" height="640" rx="32" fill="url(#violetGlow)"/><rect width="1600" height="640" rx="32" fill="url(#cyanGlow)"/>
+<path d="M790 70 C980 15 1240 42 1515 180" fill="none" stroke="#8D7BFF" stroke-opacity=".08"/><path d="M780 575 C1020 620 1300 585 1510 425" fill="none" stroke="#47CFFF" stroke-opacity=".08"/><path d="M925 36 C1120 160 1380 132 1565 250" fill="none" stroke="#FFFFFF" stroke-opacity=".035"/>
+<g transform="translate(92,86)">
+  <g><rect x="0" y="0" width="224" height="34" rx="17" fill="#111528" stroke="#7568FF" stroke-opacity=".42"/><circle cx="19" cy="17" r="4" fill="#6A5BFF" filter="url(#glow)"/><text x="33" y="22" fill="#BFB8FF" font-family="Inter, Arial, sans-serif" font-size="14" font-weight="700" letter-spacing="1.7">DEEPSEEK HARNESS</text></g>
+  <text x="0" y="101" fill="#F7F8FF" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="68" font-weight="800" letter-spacing="-2.2">DSH 视觉</text>
+  <text x="0" y="176" fill="url(#iris)" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="70" font-weight="800" letter-spacing="-2.2">路由器</text>
+  <text x="2" y="228" fill="#C7CCDB" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="24">给 DeepSeek 一双眼，而不是换一个大脑。</text>
+  <text x="2" y="277" fill="#7F889F" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="16.5">内置免费视觉链 · 自动故障切换 · 像素级工具</text>
+  <text x="2" y="306" fill="#7F889F" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="16.5">一条命令安装 · 无需注册 · 无需 API Key · 无需 Python</text>
+  <g transform="translate(0,355)" fill="#CFD5E6" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="14" font-weight="650">
+    <g><rect width="132" height="39" rx="12" fill="#111526" stroke="#252B42"/><circle cx="18" cy="19.5" r="4" fill="#50D890"/><text x="31" y="25">免费链路</text></g>
+    <g transform="translate(145,0)"><rect width="132" height="39" rx="12" fill="#111526" stroke="#252B42"/><circle cx="18" cy="19.5" r="4" fill="#7B6DFF"/><text x="31" y="25">自动路由</text></g>
+    <g transform="translate(290,0)"><rect width="132" height="39" rx="12" fill="#111526" stroke="#252B42"/><circle cx="18" cy="19.5" r="4" fill="#4CD5FF"/><text x="31" y="25">像素循环</text></g>
+  </g>
+  <g transform="translate(2,438)"><path d="M0 0 H422" stroke="#252A3C"/><text x="0" y="38" fill="#666F85" font-family="ui-monospace, SFMono-Regular, Menlo, monospace" font-size="14">github.com/ysr666/dsh-vision-router</text></g>
+</g>
+<g>
+  <rect x="760" y="64" width="760" height="510" rx="34" fill="#0C0F1C" fill-opacity=".45" stroke="#252B40" stroke-opacity=".72"/><rect x="780" y="84" width="720" height="470" rx="26" fill="none" stroke="#FFFFFF" stroke-opacity=".035"/>
+  <g transform="translate(815,248)"><rect width="116" height="142" rx="20" fill="url(#chip)" stroke="#2B3148"/><rect x="16" y="17" width="84" height="82" rx="13" fill="#0A0D17" stroke="#2D3550"/><rect x="28" y="31" width="23" height="23" rx="5" fill="#5C51E8"/><rect x="55" y="31" width="33" height="23" rx="5" fill="#183C58"/><rect x="28" y="58" width="38" height="28" rx="5" fill="#154A63"/><rect x="70" y="58" width="18" height="28" rx="5" fill="#7F6FFF"/><text x="58" y="121" text-anchor="middle" fill="#AEB6CC" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="13" font-weight="700">图像</text></g>
+  <path d="M931 319 C978 319 995 319 1032 319" fill="none" stroke="url(#route)" stroke-width="2.7" stroke-linecap="round" filter="url(#glow)"/><circle cx="974" cy="319" r="4" fill="#7668FF"/><circle cx="1004" cy="319" r="3" fill="#7668FF" opacity=".65"/>
+  <g transform="translate(1111,319)"><circle r="102" fill="#0D1020" stroke="#2B314B" stroke-width="1.4"/><circle r="87" fill="none" stroke="#6659FF" stroke-opacity=".18"/><circle r="68" fill="none" stroke="#46D7FF" stroke-opacity=".10"/><path d="M-64 0 C-35 -38 35 -38 64 0 C35 38 -35 38 -64 0Z" fill="#11162A" stroke="url(#iris)" stroke-width="2.5" filter="url(#glow)"/><circle r="28" fill="url(#iris)" opacity=".92" filter="url(#softGlow)"/><circle r="12" fill="#080B14"/><circle cx="7" cy="-8" r="4.5" fill="#DDF8FF"/><circle cx="-91" r="4" fill="#7568FF"/><circle cx="91" cy="-49" r="4" fill="#49D5FF"/><circle cx="101" r="4" fill="#49D5FF"/><circle cx="91" cy="49" r="4" fill="#49D5FF"/><text x="0" y="132" text-anchor="middle" fill="#7A839A" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="13" font-weight="700">视觉路由核心</text></g>
+  <path d="M1212 270 C1258 270 1260 183 1308 183" fill="none" stroke="url(#route)" stroke-width="2.2"/><path d="M1214 319 C1264 319 1266 319 1308 319" fill="none" stroke="url(#route)" stroke-width="2.2"/><path d="M1212 368 C1258 368 1260 455 1308 455" fill="none" stroke="url(#route)" stroke-width="2.2"/><circle cx="1268" cy="213" r="4" fill="#65D9FF"/><circle cx="1262" cy="319" r="4" fill="#7B6DFF"/><circle cx="1268" cy="425" r="4" fill="#65D9FF"/>
+  <g transform="translate(1308,137)"><rect width="164" height="92" rx="17" fill="url(#chip)" stroke="#2C334B"/><circle cx="24" cy="26" r="8" fill="#2ED887" fill-opacity=".14" stroke="#4BDB96"/><path d="M20 26 l3 3 l6 -7" fill="none" stroke="#60E6A5" stroke-width="2"/><text x="43" y="31" fill="#E0E5F2" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="14" font-weight="700">免费视觉</text><text x="22" y="61" fill="#737D94" font-family="monospace" font-size="11.5">Qwen2.5-VL</text><text x="22" y="78" fill="#535D74" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="10.5">内置 · 无需 Key</text></g>
+  <g transform="translate(1308,273)"><rect width="164" height="92" rx="17" fill="url(#chip)" stroke="#403772"/><circle cx="24" cy="26" r="8" fill="#7A6BFF" fill-opacity=".15" stroke="#7A6BFF"/><path d="M20 26 h8 M24 22 v8" stroke="#A79DFF" stroke-width="1.7"/><text x="43" y="31" fill="#E7E4FF" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="14" font-weight="700">故障切换</text><text x="22" y="61" fill="#8E86C9" font-family="monospace" font-size="11.5">provider chain</text><text x="22" y="78" fill="#615B83" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="10.5">配额 · 区域 · 429</text></g>
+  <g transform="translate(1308,409)"><rect width="164" height="92" rx="17" fill="url(#chip)" stroke="#263D52"/><circle cx="24" cy="26" r="8" fill="#43D6FF" fill-opacity=".13" stroke="#44D6FF"/><path d="M19 26 h10 M24 21 v10" stroke="#7AE3FF" stroke-width="1.7"/><text x="43" y="31" fill="#DDF7FF" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="14" font-weight="700">像素工具</text><text x="22" y="61" fill="#7DB5C6" font-family="monospace" font-size="11.5">crop · diff · OCR</text><text x="22" y="78" fill="#52707D" font-family="monospace" font-size="10.5">ground · trace · color</text></g>
+  <g transform="translate(982,106)"><rect width="316" height="40" rx="20" fill="#101526" stroke="#2D3450"/><circle cx="21" cy="20" r="4" fill="#A092FF"/><text x="36" y="25" fill="#BAC1D2" font-family="PingFang SC, Noto Sans CJK SC, Microsoft YaHei, Inter, Arial, sans-serif" font-size="13" font-weight="650">DEEPSEEK 仍是大脑</text><path d="M239 20 h46" stroke="#343C59" stroke-width="1.5"/><circle cx="291" cy="20" r="4" fill="#4BD5FF"/></g>
+</g>
+<rect x=".75" y=".75" width="1598.5" height="638.5" rx="31.25" fill="none" stroke="#FFFFFF" stroke-opacity=".08" stroke-width="1.5"/>
+</svg>