dsh-modlens
为纯文本 LLM 提供插件式视觉能力——ModLens 的 DeepSeek Harness (dsh) 分支，增加多引擎支持（Gemini、Volcengine Ark/Doubao、Claude）以及每次调用时的引擎选择
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:YZz-S/dsh-modlens#7aa254373166ec94212c835298df3f3430648f7cModlens 以 dsh-modlens 发布，当前版本为 3.16.6。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A fork of liustack/modlens — the vision plugin for DeepSeek Harness (dsh) — with multi-engine support and per-call engine selection.
这是 liustack/modlens 的分支——DeepSeek Harness（dsh）的视觉插件，支持多引擎与按调用选择引擎。
中文 · ← Back to DeepSeekHarnessPlugins
Plug-in vision for text-only LLMs. ModLens converts an image into structured JSON evidence (OCR, layout, semantics) so a text-only model can "see". This fork keeps all upstream behavior and adds:
provider argument on modlens_read_image — pin the vision engine per call (-p <provider> passthrough).doubao-seed-2.1-turbo via the Ark ark-coding-plan endpoint (OpenAI-compatible), alongside Gemini, Anthropic, and Claude Code.Upstream ModLens is under active development. This fork carries the provider passthrough and the doubao/ark recipe until they land upstream. Prefer upstream for everything else: https://github.com/liustack/modlens.
dsh plugin --profile web add github:YZz-S/dsh-modlens
Restart dsh, then look for the (modlens vision) model entries and the modlens_read_image tool.
Config lives in ~/.modlens/config.json (shared with upstream modlens). Run modlens doctor to see what is ready.
modlens config set gemini-api.apiKey <key> # https://aistudio.google.com
modlens config set provider gemini-api
modlens config set openai.baseUrl https://ark.cn-beijing.volces.com/api/coding/v3
modlens config set openai.apiKey <ARK_CODING_PLAN_API_KEY>
modlens config set openai.model doubao-seed-2.1-turbo
# optional: modlens config set provider openai
Claude Code (claude-cli), Anthropic (anthropic), Antigravity (antigravity-cli). Full recipes: skills/modlens/references/configure.md.
When the model needs to read an image and more than one engine is configured, it asks you which to use, then calls modlens_read_image with the provider argument:
provider: "gemini-api" | "openai" (doubao/Ark) | "anthropic" | "claude-cli" | "antigravity-cli"
Omit provider to use the configured default and its failover chain.
modlens doctor # local diagnostics, spends no quota
modlens -i image.png # end-to-end read (spends one read)
MIT. Original work © 2026 Leon Liu (liustack); fork changes © 2026 YZz-S. See LICENSE.
Upstream: https://github.com/liustack/modlens
English · ← 返回 DeepSeekHarnessPlugins
为纯文本 LLM 提供视觉能力。ModLens 把图片转换为结构化 JSON 证据（OCR、布局、语义），让纯文本模型也能"看见"。本分支保留全部上游行为，并新增：
modlens_read_image 增加 provider 参数——每次调用可指定视觉引擎（-p <provider> 透传）。ark-coding-plan 端点（OpenAI 兼容）使用 doubao-seed-2.1-turbo 的实测配置，与 Gemini、Anthropic、Claude Code 并列。上游 ModLens 正在积极开发中。本分支承载 provider 透传与豆包/Ark 配方，直到它们合入上游。其他一切请优先使用上游：https://github.com/liustack/modlens。
dsh plugin --profile web add github:YZz-S/dsh-modlens
重启 dsh，然后查找 (modlens vision) 模型条目与 modlens_read_image 工具。
配置位于 ~/.modlens/config.json（与上游 modlens 共用）。运行 modlens doctor 查看哪些已就绪。
modlens config set gemini-api.apiKey <key> # https://aistudio.google.com
modlens config set provider gemini-api
modlens config set openai.baseUrl https://ark.cn-beijing.volces.com/api/coding/v3
modlens config set openai.apiKey <ARK_CODING_PLAN_API_KEY>
modlens config set openai.model doubao-seed-2.1-turbo
# 可选：modlens config set provider openai
Claude Code（claude-cli）、Anthropic（anthropic）、Antigravity（antigravity-cli）。完整配方：skills/modlens/references/configure.md。
当模型需要读取图片且配置了多个引擎时，它会先询问你使用哪个，然后带 provider 参数调用 modlens_read_image：
provider: "gemini-api" | "openai"（豆包/Ark）| "anthropic" | "claude-cli" | "antigravity-cli"
省略 provider 则使用配置的默认引擎及其故障转移链。
modlens doctor # 本地诊断，不消耗额度
modlens -i image.png # 端到端读取（消耗一次读取）
MIT。原始作品 © 2026 Leon Liu（liustack）；分支改动 © 2026 YZz-S。见 LICENSE。
上游：https://github.com/liustack/modlens