When processing long conversations in the DeepSeek Harness (DSH) Web GUI, dealing with large amounts of text, quickly locating loaded turns, viewing summaries, and reviewing historical content usually requires frequent page scrolling. dsh-ui-turn-rail is a plugin designed for the DSH Web GUI. It adds a persistent progress indicator beside the conversation scrolling area, helping users intuitively manage conversation progress.

Plugin Overview

Maintained by HenryHwong, this plugin is a client-side component in the DSH plugin ecosystem. It reads data from conversation snapshots, performs derivation in the browser, and displays turn progress as dots in the interface.

Core Features

  1. Persistent Progress Bar: The progress bar is fixed beside the conversation scrolling area. Even if the conversation content is long, the progress bar remains visible and does not disappear when the page scrolls.
  2. Dots and Placeholders: Each loaded conversation turn corresponds to a solid dot. For turns that have not yet been loaded (the question is not within the current window), a dashed placeholder is displayed. AI responses and reasoning processes are not counted as loaded prompt points.
  3. Summary Preview: Hovering the mouse over a dot displays a one-line summary of the first question in that turn. Summaries are generated for both user input and AI option selection.
  4. Deep History Loading: Clicking an unloaded dashed placeholder causes the plugin to automatically paginate and load all older batches until the question for the target turn appears within the window, then it automatically jumps to the line containing that question.
  5. Click to Jump: Clicking a loaded solid dot directly scrolls the conversation log to the question position of that turn.
  6. Scrollable Window: The progress bar displays approximately ten dots and supports internal scrolling. When the reader moves their reading position, the dot for the current turn automatically follows into view.

Installation and Activation

Installing and enabling this plugin requires two steps: installing the plugin package through the DSH plugin manager, and applying host instrumentation (a source-level patch).

1. Install the Plugin

Run the following command in the command line to add the plugin to the Web configuration file:

dsh plugin --profile web add @huanghanheng/dsh-ui-turn-rail

2. Apply Host Instrumentation

This plugin requires registering a read-only slot in the DSH Web client. Because this involves source-level modifications, the dsh plugin command cannot complete it automatically. You must apply a patch in the DSH code checkout directory and then rebuild.

Run the following script to apply the patch and build:

scripts/apply-instrumentation.sh /path/to/harness-checkout
(cd /path/to/harness-checkout && pnpm run build)

3. Manual Installation Support

If you need to install it manually, you can also install the package directly using npm/pnpm and declare the bundle in the profile’s package.json. After installation, you still need to run the above Step 2 to apply instrumentation and build.

Security and Permissions

This plugin is designed as a purely presentational component:
* No Network Requests: It does not send any requests to the server, does not write session logs, and does not touch credentials.
* No New Permissions: It only adds a read-only slot to the main chat view and introduces no new RPCs or permissions.
* Local Computation: All data derivations (turn ordering, node mapping, and summaries) run on the client-side ChatSnapshot; the plugin itself does not hold session state.

Ecosystem Context

DeepSeek Harness follows the “everything is a plugin” philosophy. This plugin catalog is located on a community-hosted platform (SkillHub), has no direct affiliation with official DeepSeek, and is provided for reference by the technical community.

Summary

By introducing a visual progress bar into the Web GUI, dsh-ui-turn-rail addresses navigation difficulties in long conversations. It provides features such as hovering to view summaries, loading history with one click, and clicking to jump. All operations are performed locally and do not affect model requests or caches. For developers who frequently process long-context conversations, this is a practical auxiliary tool.