dsh-suggest-ghost: DeepSeek Harness Web Input Prediction and Ghost Completion Plugin

Introduction¶

When developing or interacting with DeepSeek Harness Web, the lack of intelligent suggestions in the input box is a common experience gap. Unlike tools such as Claude Code, the DSH Web interface does not natively provide input prediction. The dsh-suggest-ghost plugin uses a hook mechanism to automatically predict the next prompt after a turn ends, or to perform prefix completion from history while typing, simulating an autocomplete experience similar to IDEs or chat tools.

What It Is¶

dsh-suggest-ghost is a pure plugin hook component that does not modify DSH core code. By listening to turn events and using a client-side DOM overlay, it implements two autocomplete modes:

1. LLM Next Suggestion: When the draft is empty, it uses an auxiliary LLM call to predict the user’s next input. It inherits the main request routing by default and requires no additional configuration.
2. History Prefix Completion: When the draft is non-empty, similar to zsh autosuggestions, it searches session history for prefix matches and completes them.

Ghost text is rendered as an approximate DOM overlay, following the input box font and scrolling. It does not rely on the official setGhost capability and does not intercept keypresses during Chinese input method composition.

Core Features¶

• Ghost Text Display: When the draft is empty, suggestion content appears as ghost text in the input box; while typing, it finds prefix completions from history.
• Quick Acceptance: Press Tab to accept the entire suggestion with one action, or → to accept it token by token.
• Bounded Throughout: Upper limits are set for input bytes, output tokens, and timeout duration to prevent resource exhaustion.
• Mode Switching: Automatically switches between LLM suggestions and history prefix completion based on the input box draft state.

Installation and Enablement¶

Installing this plugin requires DeepSeek Harness core version 0.2.0-rc.2 or earlier. The installation command is:

dsh plugin --profile web add dsh-suggest-ghost

After installation, restart DSH Web and refresh the page to use it.

Configuration¶

All settings are located in the Suggest ghost card within the Settings -> Plugins panel.

UI Settings¶

• LLM Next Suggestion: enable toggle, output token limit, suggestion character limit, number of reference turns, transcript character budget, timeout (ms), acceptance shortcut key, provider/model routing (leave empty to inherit the main request).
• History Prefix Completion: enable toggle, cross-session search, max history entries, minimum input characters, token-by-token acceptance.
• Frequency Management: manages a cross-session list of high-frequency phrases, supporting pinning (no eviction), deletion, manual addition, and clearing.

Code-Level Configuration¶

Initial values can also be overridden via cordis.patch.yml:

- id: suggest-ghost
  config:
    maxInputBytes: 4096
    maxOutputTokens: 512
    timeoutMs: 60000
    maxRecentTurns: 3
    maxTranscriptChars: 12000
    contextSummaryEnabled: true
    contextSummaryChars: 1500
    maxSuggestionChars: 240
    acceptKey: Tab
    llmEnabled: true

Compatibility and Dependencies¶

This plugin requires DSH core 0.2.0-rc.2 or earlier. Starting from 0.2.0, peer dependency checks become hard rejections at load time, and plugin loading is skipped directly if the version requirement is not met.

In version 0.1.7, DSH removed the settings.yaml system, and the plugin’s data channel was migrated accordingly to the new Config structure. The client reads configuration through remote.settings.describe(), while the host pushes the _push field through ctx.settings.update().

Usage Examples¶

• History Prefix Completion: Type “list do” in the input box, and the plugin automatically completes the remaining part of that historical message.
• LLM Next Suggestion: After a turn ends, if the input box is empty, the plugin predicts the next input based on context and displays the suggestion in gray.
• Accept Suggestion: Press Tab to accept the entire suggestion, or → to accept it token by token.

Notes¶

• Version Requirement: You must use DSH core version 0.2.0-rc.2 or earlier.
• Requires Restart: It is available immediately after restarting DSH Web; history is stored in session logs.
• Input Method Interference: During Chinese input method composition, the plugin does not intercept keypresses so as not to interrupt typing.
• License: Defaults to the MIT license.

Conclusion¶

dsh-suggest-ghost provides practical input assistance for DeepSeek Harness Web users through a lightweight DOM overlay and smart completion strategies. Whether it predicts inputs using an LLM or quickly completes them from history, it can significantly improve interaction efficiency.

Plugin directory: https://www.skillhub.cn/plugins/dsh-plugin-collection/dsh-suggest-ghost
GitHub repository: https://github.com/dsh-plugin-collection/dsh-suggest-ghost