SkillFoxx

DS Vision Skill

Gives text-only models vision: images go to GLM and Agnes, OCR to Baidu and Windows, PDFs to MinerU, the answer comes back as JSON

  • 158 ★ on the repository
  • Last commit
  • MIT
Skill

Extended access

Why this level

  • Images and documents are sent to third-party cloud APIs with your keys
  • The scripts store keys in the user configuration and launch local models
All reasons and checks

sorwcyra/ds-vision-skillAuthor:Sorwcyra

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/ds-vision-skill

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Extended accessThe tool runs code or makes network calls and only reads external data.

  • Images and documents are sent to third-party cloud APIs with your keys
  • The scripts store keys in the user configuration and launch local models

This is third-party code. Review the repository files before installing.

What it does

DS Vision Skill is a vision layer for agents on text-only models such as DeepSeek. When a user sends a picture, screenshot, photo, chart, scan, PDF or math problem, the skill detects the task type and picks a route through `scripts/vision-router.ps1`. Image understanding goes to a free GLM and Agnes pool where models race, then to three user slots and a local model. Plain text is read by Baidu OCR or Windows OCR, and documents are parsed by MinerU. Every channel returns one JSON envelope and the main model reasons over it; the skill does not replace it. Keys are configured with `scripts\setup.cmd -SetKey`, and your own OpenAI-compatible endpoints go into the `custom-1`, `custom-2` and `custom-3` slots. For local mode `local-select.ps1` picks a model by video memory (Ollama, LM Studio, llama.cpp). The author publishes speed measurements: version 0.5.0 cut median wall time by 44.82 percent versus 0.4.1. The scripts are written in PowerShell, with a `vlm-vision.sh` variant.

Who it is for. Users of DeepSeek and other text-only models in Claude Code and Codex who need to read screenshots, charts, scans and PDFs without switching the main model.

Good fit when

  • A text-only agent received a screenshot of an interface or an error and must understand it
  • Text has to be pulled from a scan or photo and cloud OCR is unavailable, so Windows OCR fits
  • A multi-page PDF or report must become Markdown through MinerU
  • Confidential images must stay on the machine and a local route is needed

Not a fit when

  • The main model is already multimodal: then a vision layer is redundant
  • You need a cross-platform install without PowerShell: scripts target Windows and Linux has only the `vlm-vision.sh` variant
  • Images must not go to third parties: cloud channels work with keys, and the only safe route is a local model

Example request

Read this error screenshot and explain what is wrong with it through ds-vision-skill

Limitations

The cloud channels GLM, Agnes, Baidu and MinerU receive your images and require their own API keys, and the terms of free pools are set by those providers. Confidential material is left with the local route through Windows OCR or Ollama. The scripts are written in PowerShell and primarily target Windows; the `cmd.exe` shell does not accept keys in angle brackets, so the examples use quotes.

Quality depends on the chosen model, and reference speed measurements come from a mock and from live channels. Local models need video memory: from 4 GB for small models up to 8 GB for 7B. Skill updates are installed with the separate `update-skill.ps1` script.

How to disable. Delete the `ds-vision-skill` folder from the agent skills directory. Keys are saved by `setup.cmd` into your user configuration; remove them with `scripts\setup.cmd -RemoveKey -Channel glm` before deleting the folder.

FAQ

Which routes does the skill choose?

Image understanding goes to the GLM and Agnes pool, then to three user slots and a local model. Text is read by Baidu OCR or Windows OCR and documents are parsed by MinerU.

Do I need keys?

For cloud channels yes: GLM, Agnes and Baidu need API keys and MinerU takes an optional token. Windows OCR and local models work without keys.

Does the skill replace the main model?

No, it only turns an image or document into text and JSON while the main model reasons and answers.

Does it work on Linux or macOS?

The scripts are mostly PowerShell and target Windows. For Linux there is `vlm-vision.sh`; test the other channels on your own system.

How do I add my own model?

Save an OpenAI-compatible endpoint, key and model in a slot with `scripts\setup.cmd -SetCustom -Slot 1 -BaseUrl "..." -Key "..." -Model "..." -Verify`. There are three slots.

Discussion

No questions or reviews yet.

Skills

Two skills for generative video and images: video writes Seedance, Kling and Veo prompts by dramaturgy rules, image picks the model and its syntax

SkillMinimal access658

A single-binary macOS tool, macvis: OCR, word coordinates, barcodes and face grouping on-device, with no cloud vision tokens.

CLIMinimal access63

Skills for an agent that makes video, images, speech and music through the PixVerse CLI using ready pipelines.

SkillExtended access65

Twelve skills that make an agent build single-file HTML animations: slides, protocol diagrams, architecture diagrams and demos for screen recording

SkillExtended access1.8K

Motion Video Skill

motion-video-skill

1080p beat-synced video: HyperFrames and GSAP, voice-over, karaoke captions, sound effects and music. Provider keys are kept in the CLI configuration

SkillBroad access142

Turns video, audio and screen recordings into searchable timestamped evidence and checks whether the intended result was achieved. MCP server, version 0.1.x

MCP serverExtended access576

README in short

The skill builds a chain for image work for text-only models: task detection, a race between fast cloud channels, fallback slots and a local model. The output is standard JSON. The author publishes methodology and raw measurements in `references/benchmarks.md`.

README badge

Original SKILL.md

---
name: ds-vision-skill
metadata:
  version: 0.5.1
  repository: https://github.com/Sorwcyra/ds-vision-skill
description: >
  为纯文本推理模型补充视觉能力。用户提供图片、截图、照片、图表、UI 截图、代码截图、数学题图片、
  扫描件、PDF 或文档，并要求描述、理解、推理、阅读、OCR、提取文字、解析图表或分析内容时使用。
  默认调用 scripts/vision-router.ps1 做自动路由：图片理解先走免费竞速池 GLM/Agnes，再走 custom-1/custom-2/custom-3/local，文档解析走 MinerU，
  纯文字识别走 Baidu OCR 或 Windows OCR。所有工具输出标准 JSON，再交给主模型推理和总结。
---

# DS Vision Skill

这个 skill 负责把视觉输入转换成文本或结构化 JSON。它不替代主模型，只负责识别任务、选择工具、执行视觉/OCR/文档解析，并把结果交给主模型继续推理。

## 首选入口

优先使用统一路由脚本：

```powershell
powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/vision-router.ps1 -Path "path/to/file.png" -Prompt "user request" -Intent auto -Json
```

如果 harness 默认使用 `cmd.exe`（例如 Zcode 或某些 Codex/Hermes 包装器），优先调用 `scripts/setup.cmd` 和 `scripts/vision-router.cmd`。不要把 PowerShell 专用语法或 `<KEY>` 占位符粘到 `cmd.exe` 中；`cmd.exe` 会把尖括号当成重定向符号，配置 key 时请使用 `"YOUR_KEY"` 这样的引号占位或真实引号值。

常用参数：

- `-Intent auto|reason|ocr|document`：默认 `auto`。图片默认走视觉理解免费竞速池；纯 OCR 请显式使用 `-Intent ocr`。
- `-Complex`：图表、数学、复杂 UI、代码截图、多步骤视觉推理时启用。
- `-AccurateOcr`：票据、扫描件、低清晰度文字识别时启用百度高精度 OCR。
- `-MaxTokens`：限制视觉模型输出长度，默认 `1024`；`-Complex` 未显式设置时使用 `2048`。
- `-TimeoutSec`：整场视觉竞速的最长等待时间，默认 `90` 秒。
- `-NoCache`：跳过缓存读取，也不写入本次结果。

只有在需要调试单个通道时，才直接调用底层脚本。

## 路由规则

1. PDF、论文、报告、长文档、多页扫描件：使用 `scripts/mineru-extract.ps1 -FilePath <file> -Mode flash -Json`。如果配置了 `MINERU_TOKEN` 且 flash 失败，再尝试 `-Mode extract`。
2. 图片且需要理解/推理：使用 `scripts/vision-router.ps1`，让 `agnes-2.5-flash`、`agnes-2.0-flash`、`glm`、`glm-thinking` 四个模型同时开始竞速；谁先成功返回就采用谁的结果。如果全部失败，再降级到 `custom-1`、`custom-2`、`custom-3` 和 `local`。
3. 图片默认进入视觉理解免费竞速池；需要纯文字识别时显式使用 `-Intent ocr`，优先 `scripts/baidu-ocr.ps1 -ImagePath <file> -Json`；未配置或失败时用 `scripts/windows-ocr.ps1 -ImagePath <file> -Json`。
4. 无法判断时：使用 `vision-router.ps1 -Intent auto -Complex -Json`。

## 降级链

Report a problem