dsh-prompt-enhancer

中文

A prompt-enhancement plugin for DeepSeek Harness (DSH): an "✨ Enhance" button in the composer that rewrites your draft prompt into a more detailed, clearer, and structured version using first-principles thinking — grounded in your session context, memory, and recent conversation. When key information is missing, it asks you through a clarification card before generating.

Features

• One-click enhance: an "✨ Enhance" button on the right end of the composer tool row
• First-principles methodology: Think (return to the essence) → Plan (break down the structure) → Execute (organize the expression) → Review (check coherence)
• Context-aware: automatically reads the current session's agent system prompt (persona, memory, skill injections) and the recent conversation history
• Clarification card: when the model judges key information is missing, it pops up a card (option buttons + free-text input + cancel) and continues after you answer
• Preview then apply: the enhanced prompt is shown for review first; clicking "Apply to input" writes it into the composer — it is never sent automatically
• DSH-native look & feel: blue button with white text, reusing DSH's send-button theme tokens (--dsw-alias-button-info-fill)

Installation

Method 1: Host profile plugin (recommended — survives refresh and restart)

1. Clone or copy this repository to any local path, e.g. ~/dsh-plugins/dsh-prompt-enhancer
2. Edit the DSH web profile (~/.dsh/profiles/web/):
   • Append to package.json → dependencies:

     "dsh-prompt-enhancer": "link:/absolute/path/to/dsh-prompt-enhancer"
     

   • Append to cordis.patch.yml:

     - insert:
         - name: dsh-prompt-enhancer
     

3. Run pnpm install in ~/.dsh/profiles/web/
4. Restart dsh web, then refresh the page — the button appears in the composer tool row

Method 2: Agent preset plugin

Add the plugin row to your agent preset's composition (~/.dsh/.agent-presets/<id>/agent.cordis.yml) and mount the package in that preset. Note: loading the browser UI via the host profile (Method 1) is the most stable.

Configuration

None required. The plugin automatically uses:

• Model route: the current session agent's model (agent.options), falling back to the agent-default-model selection; it calls the model with thinking disabled (reasoningEffort: off) for speed and stability
• Context: systemPrompt.assemble() (persona / memory / skills) + sessionQuery.readSurface() (last 30 messages)

Usage

1. Type a draft prompt in the composer, e.g. Write a weekly summary report
2. Click the "✨ Enhance" button
3. If the model decides key information is missing, a clarification card appears (e.g. "What did you complete this week?", "Who is the audience?") — select or type your answers, then click "Generate"
4. Preview the enhanced prompt, click "Apply to input" to write it into the composer, then edit and send as you like

The output language follows your input language.

Example

Input:

Write a weekly summary report

Enhanced output:

## Objective
Generate a submittable weekly report that covers this week's completed work,
achievements, issues, and next week's plan, so progress can be synced with
your manager or team.

## Background
You want a clearly structured, results-focused weekly report. Fill in the
placeholders below with your own details.

## Requirements
- Summarize completed tasks, milestones, and outcomes this week
- Note any blockers or issues and their current status
- List next week's planned work with priorities
- Keep the tone concise and professional; use bullet points and short sentences

## Output Format
Markdown with the sections: This Week's Work / Achievements / Issues / Next Week's Plan

## Acceptance Criteria
- All four sections are present and non-empty
- No placeholders remain
- The report can be submitted as-is

Dependencies

• DeepSeek Harness (DSH) web, with at least one working model route (e.g. DeepSeek official API)
• Node.js ^22.19.0 || >=24
• Browser-side packages (bundled with DSH web): @deepseek-ai/dsh-client-runtime, @deepseek-ai/dsh-client-connection, @deepseek-ai/dsh-client-ui-slots, @deepseek-ai/dsh-client-locale

How It Works

Component	Description
Host half (lib/index.js)	Serves POST /__dsh-enhance/api JSON RPC: prepare (decides whether clarification is needed and returns a structured question JSON), complete (generates the final enhanced prompt as plain text); assembles context and calls llm.stream
Client half (lib/client.js)	ModuleLoader browser plugin: the composer button (conversation.input.right) and the overlay cards (shell.overlay — busy / clarify / preview / error states)

License

MIT