Introduction

When using AI assistants to write code or look up information, replies often contain terms like RAG, LoRA, and MCP. If you don’t recognize them, you have to interrupt the current conversation: either copy the text out to search for it, or ask for clarification in the chat—both options break the flow; the former disrupts your rhythm, and the latter pollutes the context.

Below is a small plugin that automates this process: what-was-that. It places a small octopus desktop pet in the DSH web interface that automatically scans terms in AI replies and explains them in plain language.

What is this

what-was-that is a static client plugin for DSH developed by SH-9999 (running on the web platform), released under the MIT license, and published on npm (what-was-that@0.1.9, compressed package 212 KB, extracted 1.3 MB). Its positioning, in the author’s own words, is: a cute desktop pet plugin that automatically helps you “understand” terms and jargon in AI replies.

It has been rewritten from a prototype “paste JS” approach into a standardized TypeScript static plugin package: the source code is in src/, and resources (dictionary, pet SVG) are bundled with the package. It uses import.meta.url for relative resolution and does not rely on hardcoded absolute paths.

Core Features

Desktop Pet

The octopus has four states: Cute, Confused, Thinking, and Happy. It uses transparent SVG graphics, adapts to light and dark themes, and is draggable.

Automatic Scanning & Local Dictionary

The plugin automatically scans the body of AI replies. When a term in the local dictionary is hit, a red dot badge appears on the octopus and a bubble tooltip pops up. The local dictionary contains 255 terms, all explained in plain language, with zero consumption and no network connection required.

AI Deep Dive

When the local explanation isn’t enough, click “Explain a bit deeper” to let the model continue explaining. This is a multi-round progressive process: short explanation → alternative phrasing → long explanation → bottom line.

Text Selection in Chat Area

In the chat area, directly select an unknown word (within 10 characters). An explanation card in plain language pops up next to the selection; click anywhere outside the card to close it. This is a feature exclusive to static plugins that requires DOM access.

Settings

  • Explanation Model: By default, it follows the chat model. You can also pin a specific configured model. Settings persist after restart and only revert to the default if that model disappears.
  • “Help me explain” Prompt: Whether to show a prompt when a selected word is not found in the dictionary.

Privacy & Permission Design

This section deserves special mention:

  1. It does not register any model tools (Tool), only observes and displays.
  2. It does not read the model’s internal “thinking” process, only focuses on the final answer body.
  3. It never sends the full reply back to the model. AI deep dive only sends the selected word plus 80 characters before and after it, along with a cache.

Installation & Enablement

Requirements: pnpm must be installed on the machine (the dsh plugin command depends on it), and Node version >= 22 (from the package.json engines field). Execute the following in the DSH profile:

dsh plugin --profile web add what-was-that

After installation, restart dsh web and hard refresh (Ctrl+Shift+R). The appearance of the small octopus in the bottom right corner indicates success.

If you don’t have pnpm, you can manually copy the lib/ artifacts from the repository into the profile and mount them (as declared by cordis.patch.yml); no compilation is required.

Developer: Local Build

If you want to modify the source code yourself, first install the dependencies (esbuild and typescript are only in devDependencies), then execute the build script:

pnpm install
node build.mjs

node build.mjs will package lib/index.js and lib/client.js, and copy the assets. The repository is configured with GitHub Actions CI, which automatically runs typecheck + tests + build on every push/PR. The plugin has zero runtime dependencies, using only DSH’s built-in interfaces. It can be stopped or deleted at any time without leaving global side effects.

Use Cases & Notes

Suitable for users chatting in the DSH web interface, especially beginners who are often stuck on terminology. For terms not covered by the dictionary, you can use text selection combined with AI deep dive as a fallback.

A reminder before installation: the plugin runs with the permissions of the current dsh process. It is recommended to review the source code and license before installing. This repository is MIT licensed, and the source code, dictionary, and build artifacts are openly available on GitHub.

Conclusion

what-was-that solves a small but frequent problem: understanding unknown terms in AI replies without having to interrupt your current work to look them up. The dictionary explanations have zero cost, and even during deep dives, only the minimal context is sent, keeping the permissions very clean.

  • Community Directory: https://www.skillhub.cn/plugins/SH-9999/what-was-that
  • GitHub Repository: https://github.com/SH-9999/what-was-that

If you have questions, you can provide feedback via the repository’s Issues.