Fishsb/dsh-prompt-enhancer

A prompt-enhancement plugin for DeepSeek Harness (DSH): type a rough prompt, click ✨, and an independent LLM call rewrites it into a stronger prompt — directly in the composer, fully undoable.

Features

• ✨ One-click enhance — independent LLM call replaces the draft in place
• ↩️ Undo anytime — restore the original with one click; manually editing the draft exits undo (undo also clears the previous memory pair)
• ⏹️ True cancel — click during enhancement aborts and restores the draft (AbortSignal propagated)
• 🛡️ Guards — empty input / slash commands / submitting states are handled; /cmd body optimizes only the body, keeping the prefix
• 🌐 i18n — follows the DSH interface language (中文 / English)
• 🎛️ 4 modes (v2.2.0) — Basic (direct, fastest) / Lite (local rule analysis) / Standard (rules + workspace & session retrieval injection) / Smart (LLM task-progress analysis + full retrieval)
• 🧠 Independent memory switch (v2.2.0) — available in every mode; when on, the previous optimization pair is injected into the next round (first run falls back to Lite automatically); when off, memory is never read or written; memory is written while on and cleared on undo
• 🔀 Combined stacking — mode context block + memory block can inject together (memory takes budget first, ≤1200 chars; the mode block uses the remainder)
• 🔄 Automatic config migration — v2.1 mode:'memory' / autoMemory migrate to mode:'lite' + memory:true; an explicit memory field takes precedence (including false)
• 🧪 Unit-tested — host pure-function tests (node:test, 38/38 passing) slice the PURE section, so tests run the very code that is shipped
• 🚀 Version check & one-click update (v2.4.0) — a new card at the top of the plugin-management page checks any public GitHub repo (default: this one) against the local version (remote tags are the source of truth); when a newer version exists, pull the 6 release files into a target directory with one click (default <workspace>/dsh-prompt-enhancer-<tag>/, configurable & persisted), then apply via the on-screen instructions (bundle dsh plugin update / dynamic re-define / script copy). A running plugin cannot replace itself — pull ≠ apply
• 🔤 Composer font parity (v2.4.0) — the ✨ button's idle / busy / undo text is explicitly anchored at 13px/500/20px, matching the composer model selector; the distribution copy has been upgraded (it previously lagged at v2.1.0)

Install

Option 1: bundle one-click install (recommended)

dsh plugin --profile web add github:Fishsb/dsh-prompt-enhancer

Restart DSH (dsh web) after installing — the ✨ button appears in the composer toolbar. Update with dsh plugin --profile web update dsh-prompt-enhancer, remove with dsh plugin --profile web remove dsh-prompt-enhancer.

Option 2: dynamic Cordis install

Dynamic Cordis plugin (host + client halves), installed inside a DSH session via the cordis toolchain:

1. Ask the agent in a DSH session to read plugin-host.js (host half) and plugin-client.js (client half) from this repo
2. Define the plugin with cordis_define: fill code.host with plugin-host.js and code.client with plugin-client.js (new plugin: plugin.kind: 'new'); it returns pluginId / packageId
3. Run it with cordis_run (mode: run)
4. The first client-half run requires browser approval
5. After approval, the ✨ button appears in the composer toolbar

Note: the dynamic client half is attached to the page connection active at activation time; a page refresh unloads it — just cordis_run again to restore.

Quick-install snippet (paste into any DSH session)

Install the dsh-prompt-enhancer plugin for me:
1. Read plugin-host.js and plugin-client.js from https://github.com/Fishsb/dsh-prompt-enhancer
2. Define it with cordis_define: code.host = plugin-host.js, code.client = plugin-client.js, plugin.kind = new
3. cordis_run the returned pluginId/packageId (mode: run)
4. Wait for me to approve in the browser

Bundle distribution (dsh plugin add one-click install) is supported — see Option 1 above.

Usage

1. Type any non-empty, non-slash-command text
2. Click the ✨ button (or press Enter while focused)
3. Wait for the independent LLM call; the draft is replaced with the enhanced version
4. Not satisfied? Click ✓ Optimized · Undo to restore the original

Configuration

Settings → "Models & plugins" → "Optimization" tab:

Setting	Description
Optimization mode	Basic (default, direct) / Lite / Standard / Smart; switching takes effect immediately and persists
Memory	On / Off; when on, the next round receives the previous optimization pair (first run falls back to Lite); when off, nothing is read or written
Context budget	0 / 2000 / 4000 / 8000 chars; 0 = no context injection (equivalent to Basic); the memory block is budget-constrained too
Timeout / Max tokens / Output limit	Request parameters
Template	Built-in / custom template text

The model chain lives in the "Models" tab: tried in order, reorderable, per-entry thinking toggle & level, inline connectivity test, restore defaults.

Privacy

• Mode context: on demand, injects "recent session messages + relevant workspace file snippets + related session fragments", bounded by the budget; sensitive files (.env / keys / credentials / logs, etc.) are hard-filtered and never injected
• Memory: only a boolean seen-marker lives in browser localStorage (dsh.enhance.seen.*, no content); the memory pair itself lives only in the current page's memory; turning the switch off stops all reads/writes
• The plugin itself records or reports nothing; diagnostics logs contain only metadata (mode, latency, etc.)
• Enhanced results come from an external LLM — verify before sending; after cancellation the underlying request may still run briefly on the provider side

Compatibility

• Depends on DSH runtime-injected APIs (llm / slots / harness / inputActions / sessionQuery / fs), which may change across DSH releases
• Use a recent DeepSeek Harness

License

MIT