dsh-media-serve

dsh-media-serve

Chinese version: README.md

⚠️ IMPORTANT: display only, not recognition

This plugin only displays files/images to the human in the chat via /media. It does not give the LLM vision — a text-only model cannot see or OCR the images. Do not ask the model to interpret image content through this plugin, and do not mistake “it displayed an image” for “it understood the image” — it never actually saw it, which can lead to wrong reasoning or wrong actions. Use a vision-capable model whenever real image understanding is required.

Author note: built by 大肥鱼 (Big-Fat-Whale), the affectionate nickname for DeepSeek, running on DeepSeek Harness 0.1.1-rc2. (A whale is big; DeepSeek is the whale, hence the playful nickname.)

What it does

A host-side (Node) DeepSeek Harness plugin that registers a /media route on the GUI's own web server and exposes workspace media files at:

http://<host>:<port>/media/<relative-path>

For example, an agent shows a screenshot straight into the conversation with:

![screenshot](http://127.0.0.1:3080/media/screenshots/your-file.png)

Because the route lives on the same server that serves the GUI, the browser can load it same-origin — no extra port, no extra process, and it survives restarts.

How the served roots are resolved (portable across machines / workspaces)

The plugin is not bound to one hard-coded folder. On each request it collects candidate roots and serves from the first that actually contains the requested file:

1. Explicit roots — config.root (a string or an array) or the DSH_MEDIA_ROOT env var (;-separated list). Relative paths resolve against dsh's working directory.
2. Automatic DSH workspace discovery — every workspace path from the DSH workspace registry, including the persisted $DSH_HOME/storages/workspace.json, tried in order.

So on another machine with a different workspace path, as long as the image lives in the workspace currently in use there, the plugin exposes it automatically — no config edits required. Explicit roots are only needed to force-expose a specific folder or one outside the registry. A stale config.root from another machine is simply skipped when the directory does not exist, so it never breaks the auto-discovery fallback.

Install (path-free first)

Restart dsh (the web profile) / refresh the GUI after installing.

• A. GitHub (recommended)

  dsh plugin --profile web add github:qfwycn/dsh-media-serve
  

• B. npm

  dsh plugin --profile web add dsh-media-serve
  # or a pinned version
  dsh plugin --profile web add dsh-media-serve@latest
  

• C. Local link (development / offline)

  dsh plugin --profile web add link:D:\path\to\dsh-media-serve
  

Optional — to force-expose a specific folder: set config.root in the profile patch, or the env var DSH_MEDIA_ROOT=C:/some/dir;D:/another.

Usage (near-zero effort for the user)

After install + restart, and with the agent instruction available (a user-global ~/.dsh/AGENTS.md), agents in new conversations already know they can display workspace images to you — no emphasis and no pasted URL needed. Just say “show this screenshot” or “open that image file”, and the agent will render it via /media automatically.

Fallback: if a particular conversation's agent is not aware yet (e.g. its machine has no user-global AGENTS.md), a single line is enough:

“You can display workspace images via http://127.0.0.1:3080/media/<relative-path>.”

Security

• GET/HEAD only; other methods → 405.
• The decoded relative path is validated and confined to the matched root (.., empty segments and NUL bytes are rejected).
• Only a whitelist of media/text extensions is served (png/jpg/jpeg/gif/webp/avif/svg/mp4/webm/txt/md/json/html); anything else → 404.
• No directory listing.
• If the web server binds 0.0.0.0, /media becomes LAN-visible — bind 127.0.0.1 or narrow config.root to a dedicated subfolder if that matters.

Uninstall

dsh plugin --profile web disable dsh-media-serve   # or remove

About this build

Built by 大肥鱼 (Big-Fat-Whale), the nickname for DeepSeek, running on DeepSeek Harness. The entire build cost less than ~US$1 (< 4 CNY) — a low-cost, high-efficiency example that “knowing how to ask + knowing how to drive Harness” beats reinventing the wheel. From “can you show an image?” to cross-machine / cross-workspace auto-adaptation, GitHub + npm publishing, and this bilingual documentation — nearly all of it automated, explained, and cleaned up as it went. Thanks to the human partner for the feedback loop.

Star the repo if it is useful, and open an issue for suggestions. 🐳