dsh-prompt-polish

English | 中文

Input-bar prompt optimizer for DeepSeek Harness (DSH): an ✨ Optimize button that rewrites your draft into a clearer, more specific, more executable prompt — optionally grounded in the current session's context.

• 6 strategies × 3 languages plus custom instructions
• Optional context: goal, todo list, compaction summary, tool results, conversation history
• Confirm dialog (Adopt / Keep draft / Later) — the draft is never silently overwritten
• Last-5 optimization history (review / refill / delete)
• Composer freeze while optimizing, with dual timeouts
• Three-tier settings persistence (browser L1 → settings row L2 → workspace file L3)

Install

dsh plugin --profile web add github:1321928757/dsh-prompt-polish#v0.2.0

Or from npm (once published):

dsh plugin --profile web add dsh-prompt-polish

PowerShell one-liner (bootstraps pnpm, pins versions, self-checks):

irm https://raw.githubusercontent.com/1321928757/dsh-prompt-polish/v0.2.0/scripts/install.ps1 | iex

Restart dsh web and the ✨ Optimize button appears on the left of the composer. Verify:

dsh --profile web --dump-config | findstr dsh-prompt-polish

Uninstall: dsh plugin --profile web remove dsh-prompt-polish

Requirements: DeepSeek Harness (Node ≥ 20). The package ships prebuilt JavaScript, so git installs need no build approval. Every model call goes through the model you already configured in DSH; nothing is sent to third parties.

Usage

Entry	What it does
Composer ✨ Optimize	Rewrites the draft with the current strategy; includes session context when 📖 is on
Composer ⚙	Settings panel: 📖 context toggle, 🔧 tool-results toggle, mode, language, custom instructions, optimization history, carry stats
Settings page row	L2 settings row, linked with the panel toggles

Modes: Default (disambiguate and fix) / Concise / Structured (目标·背景·要求·期望输出) / Creative / Translate / Code request. Languages: follow draft / Chinese / English (target language in Translate mode).

Output guarantee: the result is always the user's first-person request to an AI — it never opens with "可以" / "好的" / "我来" and never answers the draft's questions. This is a fixed, byte-identical prompt red line.

Screenshots

Composer with the ✨ Optimize button and the ⚙ settings toggle:

Optimization result confirm dialog (Adopt / Keep draft / Later):

Settings panel (context toggles, mode, language, custom instructions, optimization history):

Settings persistence (three tiers)

1. L1 browser: localStorage key ptopt.settings.v1, effective immediately;
2. L2 settings page: a row under DSH Settings → General, linked with the panel;
3. L3 workspace file: <workspace>/.dsh/prompt-optimizer/settings.json, synced across browsers/machines; writes are bound by the DSH sandbox policy and silently degrade to L1 when denied.

Compatibility

• Required host capabilities (all provided by the default web profile via @deepseek-ai/dsh-base): llm, sessions, fs, sandboxPolicy, agentDefaultModel, timer, typert-loader;
• The browser half talks to the host over Typert RPC (remote.promptPolish.*) — no extra network endpoints;
• All UI consumes --dsw-* theme tokens and follows the global light/dark theme.

Development

node --test test/          # unit tests (strategy assembly, truncation, classification)
node --check lib/index.js  # syntax check

Local install loop:

dsh plugin --profile demo add ./path/to/dsh-prompt-polish
dsh --profile demo --dump-config

License

MIT. Not affiliated with DeepSeek. Installing runs third-party code on your machine — review the source first.