dsh-prompt-workbench
适用于 DeepSeek Harness 的提示词重写工作台——支持四种重写模式、可取消的流式输出，以及令牌级的前后差异对比。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:fufuyin/dsh-prompt-workbench#73f3c08ce2716648f43ff6577149dc1138c888daPrompt Workbench 以 dsh-prompt-workbench 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →A prompt rewrite workbench for DeepSeek Harness — four rewrite modes, streaming output you can stop, and a real before/after diff.
一个给 DeepSeek Harness（dsh）用的提示词重写工作台：四种增强模式、可中断的流式输出、以及真正的增强前后逐词差异对比。
Writing a good prompt is the highest-leverage thing you do in an agent harness, and it is also the thing everyone skips. You type a vague sentence, the agent guesses, and three turns later you are explaining what you actually meant.
dsh-prompt-workbench sits above the composer and turns that vague sentence into a
prompt an agent can act on — without leaving the box you are already typing in.
It is deliberately not just an "enhance" button. Three things make it a workbench rather than a one-shot rewriter:
| Four rewrite modes | 细化需求 (refine) · 补充约束 (constraints) · 结构化重写 (structure) · 精简表达 (concise) |
| Bilingual by design | Chinese and English drafts both work; 自动 mode detects the draft's script and answers in kind |
| Token-level diff | LCS diff with additions highlighted and deletions struck through; degrades to plain text past its token cap instead of freezing |
| Streaming + cancel | Incremental delivery with a live character counter and elapsed time; stop a run at any point |
| Two surfaces, one state | Composer tool-row pill and a floating panel above the composer — both reading one store, so they never disagree |
| Theme-native | Every colour is a --dsw-alias-* variable; light and dark both work, and prefers-reduced-motion is respected |
| Prompt hygiene | Preserves @path references, {variables}, <slots>, TODOs, URLs and fenced code blocks verbatim; never answers the prompt it is rewriting |
dsh plugin --profile web add github:fufuyin/dsh-prompt-workbench
Then restart the profile and reload the web GUI. Most plugins go live after a page refresh; if the pill does not appear in the composer tool row, restart the profile.
Install from a local checkout while developing:
git clone https://github.com/fufuyin/dsh-prompt-workbench.git
cd dsh-prompt-workbench
npm install
npm run build
dsh plugin --profile web add .
⭯ to pull the current composer draft in.)Ctrl/Cmd + Enter). The right pane streams the rewrite.| Mode | What it does |
|---|---|
| 细化需求 refine | Turns a vague wish into concrete, checkable requirements: the real goal, the expected outcome, the deliverables. |
| 补充约束 constraints | Makes the implicit boundaries explicit — in scope / out of scope, platform limits, style rules, non-goals, acceptance criteria. |
| 结构化重写 structure | Reorganizes the whole prompt into a scannable hierarchy (目标 / 背景 / 输入 / 具体要求 / 输出格式 / 验收标准) without changing content. |
| 精简表达 concise | Removes filler, hedging and repetition, merging overlaps while keeping every distinct fact. |
There is no settings page and no config file. The plugin follows your session's default model route — change the model in the composer's model selector and the next rewrite uses it. The active route is shown as a read-only chip in the panel header.
Deliberate limits, all enforced host-side:
| Limit | Value | Behaviour past it |
|---|---|---|
| Draft length | 24,000 characters | Refused with an explicit message; no silent truncation |
| Live diff size | 700 tokens per side | Diff view degrades to plain text with a note |
| Concurrent runs | 8 tracked tasks | Finished tasks are reaped first |
An everything-is-a-plugin harness means this is two halves and a patch layer:
cordis.patch.yml the profile patch that installs the plugin
src/index.ts HOST — plugin entry; registers the plugin's HTTP routes
src/runs.ts HOST — the run table and the one model call per run
src/prompt.ts HOST — modes + system instruction (pure, testable)
src/http.ts HOST — JSON / same-origin / body helpers
src/client/index.ts CLIENT — inject declaration, style injection, mounting
src/client/panel.ts CLIENT — the two composer surfaces
src/client/api.ts CLIENT — the fetch carrier
src/client/store.ts CLIENT — the one store they share
src/client/diff.ts CLIENT — token-level LCS diff
src/client/styles.ts CLIENT — the whole stylesheet, as one string
scripts/preflight.mjs the post-build assertions `prepack` runs
The host half owns the model call and the run table; the browser half polls it on a fixed cadence and only ever receives the delta since its last cursor, so a long rewrite costs a bounded amount of traffic rather than re-sending the whole buffer.
Through the plugin's own HTTP routes on ctx.webServer.register(), not through
@Remote / ctx.remote. That is not a stylistic choice: the host gateway does
discover @Remote markers from a third-party package, but the client's $mount
path rejects any descriptor that is not a generated strict codec, and the
assembly that mounts remote namespaces is a closed in-tree list. Shipping
generated codecs would require the harness's own codegen toolchain.
Because plugin-owned routes are not covered by the connection's Host/Origin
checks, the two mutating routes gate themselves with a same-origin check. The
polling route deliberately does not — browsers omit Origin on same-origin GET,
so enforcing it there would reject the plugin's own traffic.
Targets the DeepSeek Harness developer preview. The harness itself ships compatibility-breaking changes, so pin the version you build against.
MIT © 2026 fufuyin
这是一个给 DeepSeek Harness 用的提示词重写工作台。
在 agent harness 里，写好提示词是收益最高、也最容易被跳过的一步。你打一句模糊的话， agent 只能猜，三轮之后你才开始解释自己到底想要什么。
这个插件就停在输入框上方，把那句模糊的话变成 agent 能准确执行的任务说明——不用离开你 正在打字的那个框。
它刻意不只是一个「增强」按钮，有三点让它更像个工作台：
| 模式 | 作用 |
|---|---|
| 细化需求 | 把模糊诉求拆成明确、可执行、可验收的具体要求 |
| 补充约束 | 补齐边界、非目标、技术/风格约束与验收标准 |
| 结构化重写 | 重排为 目标 / 背景 / 输入 / 具体要求 / 输出格式 / 验收标准 的清晰层级 |
| 精简表达 | 去冗余与重复，合并重叠表述，关键信息零丢失 |
dsh plugin --profile web add github:fufuyin/dsh-prompt-workbench
装完重启 profile 并刷新 Web GUI。若输入框工具行没出现胶囊按钮，重启 profile。
在输入框打好草稿 → 点工具行的 ✦ 提示词增强 → 选模式 → 开始增强（或 Ctrl/Cmd + ↵）
→ 右栏流式出结果 → 点 差异高亮 看逐词改动 → 替换输入框 写回。
没有设置页，也没有配置文件。插件跟随你当前会话的默认模型路由——在输入框的模型选择器 里换模型，下一次增强就用新模型。面板标题栏会显示当前路由。
| 限制 | 阈值 | 超限行为 |
|---|---|---|
| 草稿长度 | 24,000 字符 | 明确报错，不静默截断 |
| 差异对比规模 | 每侧 700 token | 降级为纯文本并注明原因，避免卡住界面 |
| 并发任务 | 8 条 | 优先回收已完成任务 |
MIT © 2026 fufuyin