vision-subagent

vision-subagent is a skill for Claude Code, Codex from OrinVoss/dsh-math-team. It costs 67 tokens per session (1,546 once invoked), scanned A, original, MIT.

A method for delegating image inspection to a separate AI agent when the main agent cannot read images.

In plain words
What is it for?
Use it to describe images, verify visual details, or structure findings from one or more pictures.
Why use it?
It provides a way to review pictures or check visual quality despite the main agent's image-input limits.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/orinvoss/dsh-math-team/vision-subagent
Any agent
npx skills add OrinVoss/dsh-math-team --skill vision-subagent
Clone the repo
git clone --depth 1 https://github.com/OrinVoss/dsh-math-team

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for vision-subagent

README.md
[![agentmods](https://agentmods.dev/badge/skills/orinvoss/dsh-math-team/vision-subagent.svg)](https://agentmods.dev/skills/orinvoss/dsh-math-team/vision-subagent)
Your own site
<a href="https://agentmods.dev/skills/orinvoss/dsh-math-team/vision-subagent"><img src="https://agentmods.dev/badge/skills/orinvoss/dsh-math-team/vision-subagent.svg" alt="Measured on agentmods" height="20"></a>
Per session 67 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,546 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00067 $0.01546
Opus 5 $0.00034 $0.00773
Sonnet 5 $0.00013 $0.00309
Haiku 4.5 $0.00007 $0.00155

Measured 5d ago against content hash 89a6d73fa5b0, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

vision-subagent scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/vision-subagent/SKILL.md · 79 lines

How it starts

The opening of the file, as written. The whole thing — 79 lines — stays where its author put it; the contents beside it link to each section on GitHub.

视觉子代理(Vision Sub-agent)

主模型不支持图像输入read_image 会拒读,如纯文本模型)时,用 workflow 派生一个指定视觉模型的子代理来"看"图——它可以用 read_image 读取图片,再根据你的要求描述内容或审查质量并返回结论。

本 skill 是全局可用的看图方案,任何预设/会话/项目都可加载,不限定业务场景。

触发条件

  • 主模型 read_image 报错拒读(如 deepseek-v4-flash 不声明图像输入)
  • 需要查看某张图片的实际内容、核对图片质量、或基于图片内容做判断,但主模型看不了

原理

DeepSeek Harness 的 workflow 工具允许在 agent() 上独立指定 providermodel。只要选一个当前部署里真正支持图像输入的模型(其 inputModalitiesimage),派生的子代理就能用 read_image 读取并处理图片。

不同部署注册的视觉模型名可能不同,不要硬编码模型名,应先探测。

如何找可用的视觉模型

llm 服务遍历各 provider 的模型,过滤出 inputModalitiesimage 的:

// 思路:遍历 provider -> resolveModelInfo -> 过滤 inputModalities 含 'image',得到 { provider, model }
const vision = /* inputModalities 含 'image' 的 {provider, model} */
await agent(prompt, { provider: vision.provider, model: vision.model })

成本优先(重要):识图/看图是高频、轻量任务,优先选择便宜的经济型视觉模型,不要使用贵的大模型。具体原则:

  • 首选opencode-go/mimo-v2.5kimi-coding/kimi-for-coding(K2.7 Code)、kimi-coding/kimi-for-coding-highspeed 等经济型视觉模型(均 text+image)。
  • 避免kimi-coding/k3k3-256k 这类旗舰/大上下文模型——识图不需要它们的强推理与长上下文,成本不划算。除非便宜模型全部不可用且任务确需更强能力,才考虑它们。
  • 换部署后先探测:按"便宜视觉模型优先 → 经济型 → 旗舰兜底"的顺序选。
  • 同一批图片审查应尽量一次性派发(一次 workflow 调用审多张),减少调用次数与成本。

已验证候选(示例):opencode-go/mimo-v2.5kimi-coding/kimi-for-codingkimi-coding/kimi-for-coding-highspeed(均 text+image,经济型)。

用法模板

用 workflow 派发看图子代理(provider/model 替换为探测到的可用视觉模型):

await agent(
  '用 read_image 工具读取 <图片绝对路径>,然后完成任务:<你的具体诉求,如描述内容 / 审查质量 / 判断是否达标>,并结构化返回结论。',
  { provider: '<探测到的provider>', model: '<探测到的视觉 model>' }  // 指定视觉模型
)

把"任务诉求"填进 prompt,子代理会先看图,再按你的要求输出。例如:描述图里有什么、检查是否有空白/遮挡/文字重叠、核验图与某条结论是否一致等。

输出约定(建议)

按需要求子代理给出结构化结果,例如:

  • 内容描述:图中关键元素、布局、文字/标题等
  • 质量核验:是否空白、模糊、遮挡/重叠、要素缺失
  • 目标核对:图是否支撑它要支撑的结论或任务
  • 判定PASS / FAIL,FAIL 时给原因与改进建议

复验闭环(FAIL → 修改 → 重审,强制)

图片核验不通过后,必须进入"修改-重审"循环,直到 PASS,才算完成。 不得"审一次出个 FAIL 就当作已核验"。

规则:

  1. FAIL 即回退修改:子代理返回 FAIL 时,按它给出的缺陷与改进建议,把图片退回给制作/修改图片的一方进行修改。
  2. 修改后必须重审:每次修改后,重新派发识图子代理对修改后的图完整复审一遍,不能只让修改方自称"改好了"。
  3. 循环直至 PASS:复审 FAIL 则继续"修改 → 重审",直到 PASS 为止;每次重审都记录结论(哪一版图、审了什么、结果)。
  4. 有界兜底(可选):可约定最大重审轮次(如 ≤3 轮);超限仍不 PASS 时,如实标记"多次修改未通过核验,存在持续缺陷",不得降级成"视为通过"。
  5. 记录留痕:把每次"FAIL 原因 → 修改动作 → 重审结果"写进审查/核验记录,供追踪与验收。

Read the full file on GitHub · 79 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 5d ago First seen · 79 lines · 67 tokens per session scan A 89a6d73fa5b0

Subscribe to this mod's changes

vision-subagent is a skill published in the GitHub repository OrinVoss/dsh-math-team (6 stars, last pushed 10d ago), licensed MIT. It adds 67 tokens to every session and 1,546 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other skills, from other repositories

project-blueprint

为新项目快速建立完整 AI 编程规范体系(AGENTS.md、文档目录、CI/CD、Git规范、测试制度)。自主发现引擎:扫描项目→分类文件→推断技术栈,覆盖 7 语言 15 框架 70+ 组件,未知栈三层递进联网回退。Establish AI coding conventions for new projects — autonomous discovery engine with heuristic dep classification, 7 languages 15 frameworks 70+ components, web search fallback for unknowns.

shuguang1994/project-blueprint · 113 tokens

design-taste-frontend

前端设计品味:反模板化的落地页/作品集/改版设计——先读需求推断调性(Design Read),再给设计方向与规范,避免 AI 味模板。当用户想要一个不像模板、有调性的页面时使用。Anti-slop frontend design taste for landing pages, portfolios, redesigns.

xiehuan123/coding-coach · 79 tokens

product-manager

产品经理:从需求发现到上线的产品全生命周期——澄清需求、写 PRD、机会评估、RICE 优先级、路线图、发布计划与度量。当用户想做某个产品/功能但需求模糊,或需要产品规划、需求文档时使用。.

xiehuan123/coding-coach · 69 tokens

coach-playbook

产品编排:面向非开发人员的「从 0 到 1」固定编排模式。当用户说「我想做一个网站 / 应用 / 小程序 / 看板 / 落地页」等想做产品的诉求时,用这个技能选择对应的固定流水线,按「澄清 → 产品体验与用户路径 → 配色与调性 → UI → 原型验证 → 实现 → 验收 → 上线」的顺序一步步带用户做出体验好的产品。Orchestration playbook for building a product end-to-end from zero.

xiehuan123/coding-coach · 130 tokens

deploy-to-vercel

Deploy applications and websites to Vercel. Use when the user requests deployment actions like "deploy my app", "deploy and give me the link", "push this live", or "create a preview deployment".

xiehuan123/coding-coach · 47 tokens

vercel-react-view-transitions

Guide for implementing smooth, native-feeling animations using React's View Transition API ( component, addTransitionType, and CSS view transition pseudo-elements). Use this skill whenever the user wants to add page transitions, animate route changes, create shared element animations, animate enter/exit of components…

xiehuan123/coding-coach · 127 tokens