Overview
dsh-eyes
On-demand vision for text-only LLMs in DeepSeek Harness: keep pasted/attached images in the backend, and let the model call view_image to look at them anytime via any OpenAI-compatible vision endpoint — as if the model were natively multimodal.
README / EN
Package documentation
Registry summary
On-demand vision for text-only LLMs in DeepSeek Harness: keep pasted/attached images in the backend, and let the model call view_image to look at them anytime via any OpenAI-compatible vision endpoint — as if the model were natively multimodal.
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
- **两个内部包装点**：因为 Harness 目前**没有公开的扩展点**用于「发送准入的图片能力判断」和「派发前剥离图片」，本插件直接包装了 `llm.resolveModelInfo` 与 `llm.streamWithRegistration` 两个方法。副作用是：纯文本主模型会在模型选择器里**显示为支持图片**（这是有意为之，才能放行带图消息）。 - **主模型自动适配**：任何纯文本主模型（不限 provider）都会被自动保护；本身原生支持图片的多模态主模型不受干预，图片直接原生通过。 - **视觉接口**要求 OpenAI 兼容（Chat Completions 或 Responses API）；Anthropic / Gemini 的原生接口不直接支持，需走它们的 OpenAI 兼容网关。 - **会话隔离与持久化**：图片引用索引按 `sessionId` 分片（一个会话只能查看自己的附件），并持久化到 `.dsh/attachments/v1/dsh-eyes-index.json`（启动时读回、新图落盘）。因此即使上下文被压缩、或进程重启，模型仍能通过 `attachment_id` 查看历史图片。 - API Key 请用环境变量 / 凭证服务管理，**不要硬编码进仓库**。