dsh-smartprompt
A DSH Web plugin that adds a Smart Prompt button
under the composer. Click it and your rough draft is rewritten into a structured, executable prompt
grounded in your workspace conventions — then filled straight back into the composer.
┌─ composer (draft: "make the login form safer") ─┐
└─────────────────────────────────────────────────┘
✨ Smart Prompt ← appears only while the draft is non-empty
- Empty draft → the button is not rendered.
- Click → the button shows Optimizing…, the result replaces the draft, the button turns into ✓ Optimized.
- Failure → one line of explanation next to the button (timeout / no model / model error); the draft is untouched.
- Each click leaves one
command/run + command/done pair in the session log (that is how DSH's command
channel works). The plugin renders that as a single-line summary in the conversation
(✨ Smart Prompt · N chars · Fill composer), never as the whole prompt.
中文摘要见文末。
Install
Three routes, in order of availability today:
| Route | Command | Status |
|---|
| Git (works now) | plugin_manager(action="install_bundle", target="github:shellpower96/dsh-prompt-optimizer#v0.1.0") | ✅ published |
| DSH Discover (market index) | add https://raw.githubusercontent.com/shellpower96/dsh-prompt-optimizer/main/market/index.json as an index source | ✅ published |
| npm | npm i -g dsh-smartprompt / install_bundle dsh-smartprompt | ⏳ name reserved, publish pending |
The package name is dsh-smartprompt (the npm name dsh-prompt-optimizer already belongs to a
different project). The repository keeps its original name, so every git spec stays valid.
DSH's plugin catalog is the package manager itself — there is no separate submission portal.
Install this bundle from its git repository:
plugin_manager(action="install_bundle", target="github:shellpower96/dsh-prompt-optimizer#v0.1.0")
...or from a local checkout while developing:
plugin_manager(action="install_bundle", target="/absolute/path/to/dsh-prompt-optimizer")
Installing needs pnpm on PATH (the plugin manager runs pnpm add); without it you get
spawn pnpm ENOENT. Remove with:
plugin_manager(action="remove_bundle", target="dsh-prompt-optimizer")
The Host half registers the command; the Client half ships as a browser module via
dsh.client. Installing a new bundle can arrive over HMR — if the button does not appear,
refresh the page once.
Minimal example
An empty config is enough:
- insert:
- id: prompt-optimizer
name: 'dsh-prompt-optimizer'
config: {}
To point it at your own conventions (this is the only integration surface — the plugin assumes
no project layout):
- insert:
- id: prompt-optimizer
name: 'dsh-prompt-optimizer'
config:
docs: ['AGENTS.md', 'README.md', 'docs/architecture.md']
contextChars: 6000
systemPromptLocale: auto
Configuration
Every value is read from the row's config; the plugin reads no environment variables and no
settings module of its own. Invalid values throw at activation with the field name.
| Field | Default | Meaning |
|---|
enabled | true | Master switch; false registers nothing |
maxDraftChars | 4000 | Reject longer drafts instead of sending them to the model |
contextChars | 8000 | Total character budget for the project-facts block |
maxOutputChars | 400 | Character cap for the optimized prompt, defined in Chinese characters; the English ceiling scales to 1.8x. The meta prompt targets 75% of it, so the hard cap only trims a runaway answer |
timeoutMs | 120000 | Per-call timeout |
docs | ["AGENTS.md","README.md"] | Workspace-relative documents read as facts, in order; missing files are skipped, absolute paths and .. are rejected |
includeGoal | true | Include the session's current goal in the facts block |
systemPromptLocale | "auto" | Language of the meta prompt: auto follows the draft's script, zh/en force one. The produced prompt always follows the draft's language |
How the prompt is produced
The meta prompt pins five properties (src/prompt.js):
- Faithful — clarify, complete, structure; never widen or narrow scope. User decisions become
TBD: ….
- Grounded — modules, files, commands, conventions and test approaches may only be named if they
actually appear in the facts block; otherwise the model writes needs locating instead of inventing them.
- Structured — Goal / Current state and landing spot / Constraints and conventions / Task breakdown /
Acceptance criteria; empty sections are dropped.
- Executable — acceptance criteria must be verifiable (a command, an observable behavior, a concrete file).
- Concise — one length budget per locale, expressed in that locale's native unit (Chinese characters /
English words), plus a bullet discipline: at most 3 bullets per section and 10 in total, no background
preamble, no restating the draft, no duplicated constraints.
- Body only — no preamble, no reasoning, no code fence;
finalizeOptimizedText() strips a fence or
preamble the model adds anyway.
Length budget
With the default maxOutputChars: 400:
| Locale | Target written into the meta prompt | Hard ceiling on the result | Token budget |
|---|
| zh | 300 字 | 400 characters | 334 |
| en | 94 words | 720 characters | 600 |
The target sits below the ceiling on purpose, so a compliant answer is never truncated. If a model
overruns anyway, finalizeOptimizedText() cuts at the last complete line (falling back to a sentence
end, then a word break, then a hard cut) and appends a single … — it never slices mid-sentence and
never writes explanatory text into the prompt body.
Layout
| File | Role |
|---|
index.js | Host entry: validate config, register /optimize-prompt |
src/config.js | Defaults + hand-written validation (fails loudly) |
src/facts.js | Fact collection: read-only, bounded, never throws; branch comes from .git/HEAD, no subprocess |
src/prompt.js | Pure: draft parsing, meta-prompt assembly (zh/en), output cleanup |
src/optimize.js | Orchestration: parse → collect → resolve model route → stream → clean |
client.js | Browser half: composer button, command-row renderer, copy |
Model route: the session Agent's own provider/model wins, then the process default model
selection (ctx.agentDefaultModel). The call goes through ctx.llm.stream(), so llm/stream
middleware still applies.
Tests
node --test
Three groups: public-API unit tests, no-project-context instantiation (no fs, no cwd,
no goals, no default model — assertions are "explicit outcome, never an uncaught throw"), and a
project-context regression that drives the plugin with a caller-declared docs list.
Known limits
- Only renders once a session exists. The button lives in
conversation.composer.dock
(scope session), so a brand-new conversation's first screen has no button; it appears after the
first message.
- The command name is fixed (
optimize-prompt): both halves build the command line from a
shared constant, and DSH rejects two registrations of the same name in one scope. Renaming it
means editing index.js and client.js together.
- Not a tool: the plugin changes no system prompt and adds no agent tool.
- Length is a soft promise plus one hard cut: models are asked to stay inside the budget; the cap is
only a guard rail. Raising
maxOutputChars raises both the target and the ceiling. Every default in
src/prompt.js derives from this one config value, so they cannot drift apart.
- Verified against DSH 0.2.0-rc.2. No
peerDependencies are declared on purpose: DSH blocks
installation on an incompatible peer, which would be worse than a documented version note.
License
MIT — see LICENSE.
中文摘要
在 DSH Web 输入框下方加一个「✨ 智能提示词」按钮：有草稿时才出现，点击后结合会话工作目录里
你自己声明的项目文档（默认 AGENTS.md / README.md）、git 分支与会话目标，把草稿改写成
「目标 / 现状与落点 / 约束与规范 / 任务拆解 / 验收标准」五段式提示词并回填输入框。
配置全部由调用方注入，插件不读任何环境变量或项目设置；校验失败在激活时直接抛错。
安装：plugin_manager install_bundle 指向本仓的 git 地址。已知边界：全新会话（还没有 session）
时按钮不出现；命令名 optimize-prompt 不可配。
