zzy6-a/dsh-prompt-enhance

One-click prompt enhancement for the DSH composer. Type a rough sentence, hit ✨ next to the model selector, and your wording comes back as a task brief with a clear goal and the missing details filled in — with one-click undo.

Prompts for coding agents usually fail on what was not said: what to change, what to leave alone, what the deliverable is, how "done" is judged. This plugin turns that into a single click, using the model you are already talking to and nothing else.

Features

✨ Enhance	Reads the draft → fills in goal / scope / deliverable / constraints / acceptance → replaces the draft
↩ Undo	The undo button stays available until you edit the draft yourself
🎛 Five styles	Fill in details (default) · Concise · Detailed · Add constraints · Structured
🧭 Model routing	Follow the session model (default) or pin a cheap model for rewriting
🧩 Chip aware	With @file references, /commands or skill chips in the draft the button greys out and explains why
🪄 Empty draft	Grey out, or "derive a prompt from the recent conversation"
🌏 Bilingual	UI copy follows the DSH language (zh / en)

Usage

1. Write whatever you have, e.g. make the log page filter a dropdown;
2. Click ✨ left of the model selector;
3. The button turns into ···, then the draft is replaced by the rewritten prompt;
4. Not happy? Click ↩ to restore the original text.

If the draft contains @ file references, / commands or skill chips, ✨ stays disabled — on purpose: a whole-draft replace would downgrade those structured references to plain text.

Settings

Settings → Prompt enhance (its own section in the sidebar, it does not take over another plugin's card).

Setting	Options	Default
Enable the enhance button	on / off	on
Rewrite style	Fill in details / Concise / Detailed / Add constraints / Structured	Fill in details
Model	Follow the session / Fixed provider + model	Follow the session
Empty draft	Disable the button / Derive a prompt from the recent conversation	Disable the button
Custom system template	Empty = built-in style template; filled = overrides it completely	empty

Values live in the host settings namespace prompt-enhance and apply immediately.

How it works

composer draft ──POST /api/dsh-prompt-enhance/enhance──▶ host half
                                                            │
                      system = built-in style / custom template ◀┘
                                                            │
                                            ctx.llm.stream (one-shot call)
                                                            │
      rewritten prompt ◀──────── JSON ───────────────────────┘
                │
   inputActions.setDraft(text)  ← whole-draft replace (undoable)

• One-shot model call: a hand-built request that never enters the session log or the agent context — no conversation pollution, no prompt-prefix cache churn;
• Route priority: the fixed pair from settings → the session's most recent main-model call (captured off the llm/stream waterfall, auxiliary calls excluded) → the host default model;
• The host half does three things: compose the system prompt, call the model, strip a stray code fence. It stores no drafts.

HTTP API

POST /api/dsh-prompt-enhance/enhance

// request
{ "sessionId": "session-…", "draft": "make the log page filter a dropdown" }

// success
{ "ok": true, "text": "…", "route": { "provider": "opencode-go", "model": "deepseek-flash" },
  "style": "default", "fromContext": false }

// failure
{ "ok": false, "error": "草稿为空" }

Status	Meaning
200	Rewritten
400	Empty draft (and context mode off)
403	Disabled in settings, or an untrusted origin
413	Draft longer than 12000 characters
422	Empty draft + context mode, but no usable conversation context
502	Model call failed / returned nothing
503	No usable model route

Known limits

• References are not preserved: drafts with chips can only be greyed out. The current plugin-facing input surface exposes setDraft only (plain text, reference placeholders stripped); preserving references needs an upstream chip-aware write action on InputActions.
• No button in a brand-new session: conversation.input.right is a session-scoped seat and does not render without a session id.
• It rewrites, it does not append: the whole draft is replaced.
• A model may ignore the "no code fence" rule; the host strips one wrapping fence as a fallback.

Install

GitHub Release (current channel)

Grab dsh-prompt-enhance-<version>.tgz from the latest release:

cd ~/.dsh/profiles/web
npm i /path/to/dsh-prompt-enhance-0.1.0.tgz
# then add "dsh-prompt-enhance" to dsh.profile.bundles

dshmarket

Search and install from the market once listed (see awesome-dsh-plugin).

From source

git clone https://github.com/zzy6-a/dsh-prompt-enhance.git
cd dsh-prompt-enhance
npm install
npm run build
# mount it in your profile your own way (bundle patch: cordis.patch.yml)

Requirements

• Node.js ≥ 22.19
• DSH ≥ 0.1.5-rc.1 (needs the settings, sessions, llm and webServer host services)
• A usable model route (session model or pinned in settings)

Development

npm install        # four dev dependencies only (schemastery / @types/node / tsdown / typescript)
npm run build      # host: tsc → lib/; client: tsdown → lib/client.js
npm run typecheck
npm test           # node:test, six host-side unit tests
npm pack           # produces dsh-prompt-enhance-<version>.tgz

Security & boundaries

• The route only serves same-origin / loopback callers (everything else: 403);
• 64 KB request cap, 12000-character draft cap, 90 s end-to-end deadline;
• No draft ever touches disk, nothing is written to the session log;
• The only network activity is sending your draft to the model route you configured.

License

MIT © 2026 zzy6-a