dsh-streamfold: Streaming Fold

Preface¶

DeepSeek Harness (DSH) uses a plugin-based architecture and allows feature extension at the client level. In long conversations, a large number of tool calls and thinking steps can easily drown out the main message text. The official Compact mode may fail to collapse correctly when the conversation history has not finished loading, making it unable to help at that point.

dsh-streamfold adds presentation enhancements on top of the official conversation flow. It does not depend on whether the history has been fully loaded, and it applies its own folding rules to ensure that process lines no longer overwhelm the main text.

Plugin Overview¶

dsh-streamfold is a client-side presentation enhancement plugin. Its core behavior is: while a turn is running, automatically expand the latest thinking pane and collapse the other process lines into a single summary; when a turn finishes, thinking and tool calls are collapsed automatically, while the main text inserted among them is preserved. The whole process advances frame by frame, like a stream of water.

• Maintainer: rezon-aki
• Category: Client
• License: MIT
• Core Mechanism: Pure client-side presentation enhancement; it only reads the conversation DOM, sends no network requests, does not touch credentials, and does not modify official source code.

Core Features¶

1. Expand the latest thinking pane automatically while running: the latest thinking pane for the current turn is expanded automatically, with an adjustable height limit; superseded older panes are collapsed back into a one-line summary after a grace period.
2. Collapse the remaining process lines into one summary: non-running turns are summarized into a single “N lines collapsed · click to expand” entry.
3. Auto-collapse after a turn finishes while preserving interleaved text: after a turn ends, thinking and tool calls are collapsed automatically, but content interleaved within the main text is preserved.
4. Pin the question at the top: whichever turn is currently in view, the question for that turn is pinned at the top of the conversation.
5. Follow-bottom scrolling: the view follows the progression of the conversation flow, and a “Back to bottom” button appears when the view is away from the bottom.
6. Pane sparks / body-text forging: the bottom of thinking panes and the streaming write head of the main text have a spark animation, with adjustable color, density, and speed.

Installation and Enablement¶

Run the following command in a terminal to install the plugin:

dsh plugin --profile web add dsh-streamfold

After installation, you need to restart the profile and refresh the browser page before the client-side code takes effect.

Prerequisite: the DeepSeek Harness version must be >= 0.2.0-rc.2.

Typical Usage¶

1. Open DSH settings.
2. Go to General Settings and find the Work Step Display option.
3. Select the Fold mode. This mode is provided by this plugin, which automatically hides the corresponding official folding option to avoid conflicts.
4. Adjust specific parameters on the dedicated Streaming Fold settings page, such as pane height, follow speed, and spark effects.

Runtime behavior:
- While running, only the latest thinking pane is expanded automatically.
- A pane you manually open will not be collapsed automatically, but an older pane superseded by a new thinking pane will collapse back into a one-line summary after the configured grace period (default 2 seconds).
- When a turn finishes, the view automatically and smoothly scrolls back to your question for that turn (if the relevant setting is enabled).

Use Cases and Notes¶

• Use cases: users handling long conversations who frequently need to inspect thinking steps but do not want to be overwhelmed by long text.
• Safety boundaries:
  • The plugin does not read or write files, does not register routes, and does not export Config; settings are stored only in the browser’s localStorage.
  • Changes to official modes are reversible (visibility only).
  • Uninstallation is clean: all injected styles, buttons, and hidden attributes are removed.

Resource Links:
- Plugin directory: https://www.skillhub.cn/plugins/rezon-aki/dsh-streamfold
- GitHub repository: https://github.com/rezon-aki/dsh-streamfold