dsh-prompt-enhance-cn

dsh-prompt-enhance-cn

Prompt enhancement for the DeepSeek Harness web GUI — one click turns a rough composer draft into a well-structured prompt: explicit role and goal, executable steps, output format, acceptance criteria, and edge cases. Your intent is never changed, nothing is fabricated, and the original draft is always preserved.

English | 简体中文

Why

Good agent prompts state *who the model should be*, *what to deliver*, *in what format*, and *how success is judged*. Most drafts don't. This plugin adds a WorkBuddy-style "enhance" affordance to the dsh composer: it sends your draft through a low-temperature rewrite pass (via the harness's own LLM service — including the base tier, which asks only for a light tidy-up), shows the result next to the original, and lets you apply, copy, or discard it. Undo is one click away, and every failure path leaves your draft exactly as you typed it.

Features

· ✨ Composer button · A small button in the input box's tool row (next to the send button), always at hand 🎚️ Four tiers · A capsule left of ✨ shows the active tier (base / standard / professional / system) and switches it in one click; the rules come from the same core/rules.md the Windows hotkey layer reads, and the base tier only does a light tidy-up 🧷 Context toggle · The dropdown has its own row for "use conversation context"; switching tiers seeds that tier's preset, which you can then override ⚡ One-click fill · The base / standard / professional tiers write the rewrite straight into the composer — no dialog. The system tier still previews first, because an eight-section system prompt rewrites a whole persona 🔀 Preview panel · Original vs. enhanced side by side, with model name and elapsed time. Opens for the system tier, and whenever your draft moved while the request was in flight ↩️ Undo / `Ctrl+Z` · After a fill, a quiet bar above the composer restores the original draft — or just press Ctrl+Z. Once you edit the filled text, both stand down and the browser's own character undo takes over ⌨️ Shortcut · Default Ctrl+Shift+Alt+A, fully configurable, acts on the composer you are working in. The Windows hotkey program shipped in the same repository is a separate component with its own default, Ctrl+Shift+Alt+E — deliberately different so installing both never fires twice 💬 `/enhance` command · Rewrite any text from the slash-command plane; the result never enters model history 🧠 Model routing · Settings pair → current session model → harness default model, in that order 🔑 Zero credential setup · Calls ride the harness LLM service; keys come from the harness credential store ⚙️ Live settings · Every knob (model override, temperature, budgets, system prompt, shortcut) hot-applies from Settings → 插件配置 🛡️ Draft safety · Empty, over-length, images-only, and command-chip inputs are rejected locally; upstream failures are mapped to readable messages; the draft is never mutated on failure

How it works

flowchart LR
    A[Composer draft] --> B{Local guards<br/>empty / length / chips / busy}
    B -- pass --> C["POST /prompt-enhance/enhance<br/>(loopback-fenced host route)"]
    B -- reject --> P["Inline notice above<br/>the composer, draft untouched"]
    C --> D["ctx.llm.stream<br/>rewrite with system strategy"]
    D --> E[Normalize:<br/>strip fences, trim, refuse empty]
    E --> Q{system tier, or<br/>draft changed meanwhile?}
    Q -- no --> G["setDraft(enhanced)<br/>original pushed to undo stack"]
    Q -- yes --> F[Preview panel:<br/>original / enhanced]
    F -- Apply --> G
    F -- Cancel / Copy --> H[Draft untouched]
    G --> U[Undo bar: one click restores]

The plugin is one npm package with two halves, following the dsh plugin conventions:

• Host half (exports ".", Node): registers the prompt-enhance settings section (schemastery — rendered automatically by the built-in plugin config page), the POST /prompt-enhance/enhance route on the shared webserver (loopback-fenced, body-capped), and the /enhance slash command. The model call goes through ctx.llm.stream with a normalized output pass — the same auxiliary-call discipline the harness applies to session titles: deadline + caller cancellation rechecked during and after the stream, terminal-finish validation, tool-call rejection.
• Browser half (exports "./client"): registers the enhance button into the conversation.input.right slot and the undo bar into conversation.input.dock, binds a live mirror of the settings namespace, and installs the global shortcut. All copy is localized (zh/en) through the harness locale system.

Requirements

• dsh >= 0.1.1-rc.2 < 0.3.0 — the range states what has actually been exercised, not an open-ended promise.
• Boot-verified on 0.1.1-rc.2, 0.1.2-alpha.3, 0.1.7-rc.2, and 0.2.0-rc.2 (layer mounts, enhance route answers, client bundle builds; on alpha the plugin takes the newer ctx.settings.installSection registration path, and on 0.2.0-rc.2 it also passes the runtime's plugin-compatibility gate). The 0.2.0-rc.2 checks include the dsh desktop application — the environment this plugin is normally used in. On a 0.2.x runtime expect some *other* plugins to disappear from the UI: the gate skips any bundle whose peerDependencies pin an older @deepseek-ai/dsh-* range (observed: dshmarket, dsh-find-plugin, dsh-better-sidebar). The harness has broken two APIs the plugin touches: @deepseek-ai/dsh-settings changed its registration entry between the two early lines, and from 0.1.5-rc.3 the settings service became SettingsForms (no synchronous get) while dsh-agent-default-model moved the harness default model behind its own ctx.agentDefaultModel.currentSelection() service. The plugin probes all three shapes at runtime and adapts, no configuration needed.
• The ceiling covers the whole 0.2.x line (including its prereleases); a 0.3.x is untested and therefore outside the range. dsh.engines is advisory metadata read by the loader and plugin manager — it states support, it does not enforce it.
• Node ^22.19.0 || >=24.0.0 (for building from source)
• Browser peers react / react-dom (^18.2.0) come from the web shell's module registry. The host-half module @deepseek-ai/dsh-llm is a runtime module supplied by the dsh host and is not installed from npm (the lib/ bundles keep every package import external), so it is declared as an *optional* peer: an install that does not provide it produces no warning, and the plugin reports a normal error at enhance time rather than failing to load.

· dsh · >= 0.1.1-rc.2 =24.0.0 Plugin · 0.3.x

Install

From npm (recommended):

dsh plugin --profile web add dsh-prompt-enhance-cn
# restart dsh web

From the release tarball (the built lib/ is inside, so no local build happens):

dsh plugin --profile web add https://github.com/qwe3120695033/prompt-enhancer/releases/download/vX.Y.Z/dsh-prompt-enhance-cn-X.Y.Z.tgz
# restart dsh web

The plugin sits in the dsh-plugin/ subdirectory of that repository — the repo root is the Windows hotkey program, so dsh plugin add github:qwe3120695033/prompt-enhancer will not find it. Use the tarball URL above (or the npm package name).

From a local checkout (for development — changes rebuild + restart take effect):

git clone https://github.com/qwe3120695033/prompt-enhancer.git
cd prompt-enhancer/dsh-plugin && npm install && npm run build
dsh plugin --profile web add link:C:\path\to\prompt-enhancer\dsh-plugin

Uninstall:

dsh plugin --profile web remove dsh-prompt-enhance-cn
# check the profile's package.json `dsh.profile.bundles` array for a leftover
# "dsh-prompt-enhance-cn" row and remove it if present, then restart dsh web

Usage

Composer button / shortcut — type (or leave) a draft, hit ✨ or Ctrl+Shift+Alt+A:

1. Local guards run first: empty drafts, over-length drafts (never auto-truncated — that would change your meaning), images-only drafts, and drafts containing command or file-reference chips are refused with a clear message. Chips are rejected because filling back would destroy them.
2. Base / standard / professional: the ✨ becomes a real spinner and a dock row above the composer reads Enhancing… + elapsed seconds + cancel. The first tokens land in the draft about a second in, typed character by character; when the run settles, that row becomes the one-line "replaced / undo" prompt. No dialog, no interruption. The ticking counter is the proof the call is alive, and ✕ aborts at any point — an abort is not a failure and never leaves half-written text behind. The base tier also calls the model now, but asks only for a light tidy-up and rejects anything longer than 1.5× the draft.
3. System tier: the preview panel opens with both texts side by side. Apply fills the enhanced text back and raises the undo bar; Copy puts it on the clipboard; Cancel (or Esc, or clicking the overlay) discards everything. A direct fill that raced with your typing promotes itself to this same panel instead of overwriting what you wrote. Panel mode shows no "Enhancing…" row — the panel is its own progress indicator.
4. Undo: the Undo link or `Ctrl+Z` restores the original. Keep typing and both retire themselves so stale text can never overwrite newer edits — the composer's own character-level undo is left alone. Undo entries live only in the current page's memory (up to 3 per session) — a page reload, a web restart, or a session switch clears them.

`/enhance ` — rewrite any text from the slash menu. The result renders in the command plane (copyable) and never enters the conversation history or the model's context. To enhance the *composer draft itself*, use the button or shortcut — the draft lives in the browser. Cancelling a running /enhance follows the harness command plane; if the client offers no cancel affordance, the call simply runs to completion or times out.

Language consistency: the strategy instructs the model to mirror the input language (Chinese in → Chinese out). This is a best-effort instruction, not a hard guarantee.

Configuration

Everything lives in the prompt-enhance settings namespace, edited from the web GUI's Settings → 插件配置 page. Changes apply to the very next call — no restart. Every enhancement is one billable LLM call: maxOutputTokens bounds its cost, and the host-side concurrency/rate caps bound how often calls can be made.

Field · Default · Description enabled · true · Master switch; off hides the button and disables every trigger provider + model · empty · Explicit route override; must be filled as a pair (or both empty to follow the current session model) temperature · 0.3 · Low temperature keeps the rewrite faithful to the original maxOutputTokens · 2048 · Output token budget of one enhancement call maxInputChars · 12000 · Input character cap (counted in Unicode code points — an emoji is one character); over-limit drafts are rejected, never truncated timeoutMs · 60000 · End-to-end deadline of one call defaultTier · standard · Default tier: base only tidies up, standard emits one continuous paragraph, professional emits the five-section structured prompt, system emits an eight-section system prompt. This is what the tier capsule next to the composer writes; it hot-applies and new sessions follow it systemPrompt · built-in strategy · Custom strategy text; how it combines with the built-in strategy is set by strategyMode strategyMode · replace-default · How a custom strategy combines with the built-in one: replace-default swaps it out entirely (backward compatible, but the built-in hard rules — preserve intent, never fabricate, body-only output, language mirroring — are not retained and must be carried into your own text); extend-default appends your text after the built-in strategy, keeping those rules in force shortcut · ctrl+shift+alt+a · Global shortcut spec (at least one modifier + one alphanumeric/function key — bare keys are ignored so normal typing can never be swallowed); empty disables it maxConcurrent · 2 · Concurrency cap within a single host process; extra requests answer 429 (concurrency-limit). The browser UI admits exactly one in-flight request (a single preview panel), so this cap mainly protects the /enhance command plane and multi-client callers rateLimitPerMinute · 10 · Sliding-window rate cap per minute, within a single host process, counting successful calls only — failures (timeout / upstream error / cancellation) never consume the window, so a run of failures cannot rate-limit you out; extra requests answer 429 (rate-limit) with a Retry-After in seconds streaming · true · Show the model's output as it streams instead of waiting for the whole rewrite — display-only, the final text is still the normalized full result, and the client transparently falls back to the one-shot JSON route when the host or network does not support text/event-stream contextAware · true · Read the current conversation's recent history and use it to ground the rewrite (resolve pronouns, fill in stated constraints, mirror the established stack and terminology). With no session, no history, or this switch off, the call degrades to the original single-prompt enhancement — context is an optimization, never a prerequisite. Privacy: the assembled history snippet is sent to the configured LLM alongside the draft — if the conversation contains secrets, tokens, or other sensitive content, switch contextAware off (or scope the conversation before enhancing). contextMaxMessages · 8 · Context window breadth — how many recent user/assistant turns may ground the rewrite; 0 admits none contextMaxChars · 4000 · Context window depth — character budget of the assembled history snippet, spent newest-first so the most recent turns win when the budget is tight; 0 admits none provider + model values · — · Match the harness settings: each key under llm-pi-ai.providers (e.g. zhipu, muyuu) is a provider and each models[].id under it (e.g. glm-5.3-flash) is a model. Example pair: provider: zhipu + model: glm-5.3-flash

Model routing precedence: explicit settings pair → the route recorded in the current session's request header → the harness-wide default model (agent-default-model). If none of them names a route (e.g. a fresh session with no default model), the plugin fails with an actionable message instead of guessing.

Four tiers and the single rule source

This fork replaces the upstream single built-in strategy with four tiers, whose wording lives in one place: core/rules.md at the repo root, the same file the Windows hotkey layer reads.

Tier · Output shape base · Light tidy-up through the model: smooths word order and connectives, completes missing punctuation/spacing/paragraph breaks, and fixes typos. No expansion, no added goals/steps/acceptance criteria, no new headings or bullets; a length guard rejects output beyond 1.5× the draft standard (default) · One continuous paragraph, 80–220 Chinese characters, no headings, numbering, bullets, tables, or code fences professional · Five fixed sections: 任务目标 / 背景与范围 / 执行要求 / 约束条件 / 交付要求, 250–500 Chinese characters (soft ceiling 700) system · Eight-section system prompt: role, core objective, working principles, rules & constraints, capability boundaries, input handling, output spec, error handling & self-check

Every tier includes the shared [COMMON] block, which carries the fidelity floor (placeholders such as __PO_SKILL_0001__ are preserved verbatim), the no-fabrication rule, and body-only output — so switching tiers can never drop the safety baseline.

The rules file is the only source of truth: npm run gen:tiers slices it into src/generated/tier-rules.ts, and npm run build runs generate-tiers.mjs --check first, failing when the generated file is stale. Edit the wording in core/rules.md, regenerate, rebuild.

Context toggle. Each tier presets a context window (base never uses history; standard 4 messages / 3000 chars; professional and system 12 / 6000), applied when you switch tiers. The row in the dropdown overrides it for the current session only — the persisted default is untouched. The preset and the configured contextMaxMessages / contextMaxChars are intersected, so tightening the global caps still works.

The built-in enhancement strategy

The default system prompt instructs the model to be a prompt-rewriting expert and to apply what the draft actually needs:

1. Role and goal — state who the assistant acts as and what the deliverable is.
2. Context and constraints — add only what the draft implies; never invent facts, data, names, or requirements.
3. Steps — break vague or multi-part requests into numbered, executable steps.
4. Output format — specify structure, language, length, and style where implied.
5. Acceptance criteria — state how to recognize a correct result.
6. Boundary conditions — list edge cases and what to do when information is missing.

Hard rules: preserve intent exactly (never remove, alter, or contradict user information); never fabricate — insert an explicit placeholder like (待补充：…) / (TBD: …) for unknown details; keep the scope unchanged; output only the rewritten body (no explanations, fences, or pleasantries); mirror the input language; stay within roughly 1–3× the original length; lightly polish already-well-formed prompts instead of inflating them.

Set systemPrompt in the settings to use your own strategy; strategyMode decides how it combines. The default replace-default is a complete replacement — the built-in safety constraints (preserve intent, never fabricate, body-only output, mirror the input language, treat the framed draft as data) are not retained automatically, so your own strategy must include them. Switching to extend-default appends your text after the built-in strategy instead, keeping those rules in force.

Error handling & edge cases

Case · Behavior Empty / whitespace / invisible-only draft · Local refusal: "input box is empty" Draft over maxInputChars · Rejected locally and at the route with exact counts; no auto-truncation Images attached but no text · Refused: text-only feature Command or file-reference chips in the draft · Refused: fill-back would destroy the chips Submitting / busy phase or a request already in flight · Refused with "try again in a moment" Per-minute cap exceeded · 429 (rate-limit) with a Retry-After in seconds; the host message names the exact wait. The window counts successful calls only — failures do not consume it Concurrency cap full · 429 (concurrency-limit), no Retry-After — a slot frees whenever an in-flight call settles Upstream model failure · Stable codes mapped to readable hints: AUTH → check API key, RATE_LIMIT → retry later, QUOTA_EXCEEDED → check balance, CONTEXT_WINDOW_EXCEEDED → shorten input, NO_ADAPTER/unconfigured → configure a model Output reaches maxOutputTokens · Refused with a hint to raise the cap or shorten the draft Model returns empty / fence-wrapped / tool-call output · Normalized (fences stripped) or refused; retryable Timeout · 504-mapped message with the configured seconds; retryable Browser tab closed mid-flight · The host route detects the disconnect and aborts the model call Session switch · Panel state, undo stack, and shortcut targeting are per-session; switching closes the panel and clears its undo entries

Every failure surfaces where the run would have landed — the plugin's own panel for the system tier, or the inline notice row for a direct fill. The composer draft is never modified by a failed call, so manual input continues undisturbed.

Error codes & localization

The host route answers structured errors of the shape { code, message?, params? }: the browser renders its localized primary line from code + params using the current language dictionary, and message — when present — is an optional diagnostic detail (the provider's raw failure text, a config error) shown verbatim beneath the primary line. The /enhance command plane has no locale dictionary; a host-side renderer of the same errors produces the Chinese text directly.

Error code · HTTP status · Browser primary copy (dictionary key) · Params rejected · 403 / 413 / 415 / 422 · generic error.rejected; with { count, max } it reuses error.tooLong · over-length input: { count, max } rate-limit · 429 · error.rateLimit · { limit, retryAfterSeconds } concurrency-limit · 429 · error.concurrencyLimit · { max } timeout · 504 · error.timeout · { seconds } unconfigured · 409 · error.unconfigured · — upstream · 502 · generic error.upstream; with reason it uses error.upstream.{reason} (auth / invalidCredential / rateLimit / quota / empty / contextWindow / toolCall / maxTokens) · { reason } internal · 500 / 502 · error.internal · —

Troubleshooting

Button missing / shortcut dead Settings → 插件配置 → prompt-enhance section: is enabled true? Is the plugin installed (dsh plugin --profile web list) and dsh web restarted? Any apply error in the browser console?

"No model resolved for the enhancement" Routing follows settings pair → session model → harness default; if all three are empty there is nothing to call. Pair provider/model in the settings, or send a message in the current session first so it carries a model route. Check the agent-default-model settings section.

Authentication failures The message detail line carries the concrete cause (e.g. 401). Check the provider's API key in the harness credential store; quota/balance problems surface as the quota hint.

Rate-limited (429) right after a run of failures Should not happen — the window counts successes only. If it still does, check whether several profiles/processes are mounted (each counts independently and stacks up) or whether rateLimitPerMinute is set too low.

Poor rewrites (fabrication, dropped requirements, broken formatting) With strategyMode set to replace-default, a custom systemPrompt replaces the built-in strategy entirely — its hard rules (never fabricate, body-only output, mirror the input language) are not retained automatically; carry them into your own text or switch to extend-default.

Behavior changed after upgrading Since 0.1.6 the host no longer sends Chinese primary-line copy: the browser localizes errors by code (message is now only a diagnostic detail), and the rate-limit window counts successful calls only. No configuration change needed.

Security model

The enhance route is served by your own ds