dsh-prompt-enhance

A ✨ button in the DeepSeek Harness composer that rewrites your rough draft into a clear, well-structured prompt — then puts it straight back into the input box.

Type a half-formed thought, click once, send a well-specified request. No copy-paste, no second window, no leaving the conversation.

Same idea as the "enhance prompt" button in WorkBuddy and similar chat clients, built as a native DSH plugin.

---

Contents

• What it looks like
• Real example
• Why not just retype it
• Requirements
• Install
• Configuration
• Usage
• How it works
• Design decisions
• Troubleshooting
• Known limitations
• Development notes
• Project layout
• Uninstall
• FAQ
• License

---

What it looks like

A sparkle button appears in the composer toolbar — boxed in red above, between the permission selector and the model picker.

It registers into conversation.input.right, a list slot immediately left of the model selector. Being a list slot matters: it can coexist with other composer buttons, and it never competes with the voice-input button that owns the neighbouring single-valued conversation.input.activity slot.

Button states

State	Look	Meaning
Idle	Grey ✨ at 75% opacity, brightens on hover	Ready
Busy	Small spinning ring	Waiting on the model
Error	✨ turns red; hover shows the reason	See Troubleshooting
Disabled	45% opacity, no hover response	No session, or a submission is in flight

---

Real example

Input:

关于 iPhone 多评测的微博，折痕到底是有没有折痕

After clicking ✨:

请基于目前公开的 iPhone 折叠屏评测内容，说明屏幕折痕的实际程度：
折痕是否明显、在哪些角度或光线下可见、与同类产品相比表现如何。
请区分客观测试结论与主观手感描述，并说明信息来源。

Input:

带哦上官金虹感觉看电视呢

After clicking ✨:

请说明上官金虹与看电视这一说法之间的关联，我的表述可能有误，
请先澄清可能的原意，再给出对应解释。

The rewrite keeps your intent and language, fixes garbled text, makes the goal explicit, and adds structure only when it helps. It never invents requirements you did not imply, and it never answers the prompt — it only rewrites it.

---

Why not just retype it

Retyping works. This exists for the cases where it doesn't:

• You're thinking out loud. Voice input, dictation, or a half-finished sentence produces a draft that is right in substance and wrong in form.
• You're on the wrong keyboard. Switching input methods mid-thought breaks flow.
• You don't know what the agent needs. A good prompt states the goal, the subject, the constraint, and what "done" looks like. Knowing that helps; writing it every time doesn't.
• Rewrites are cheap and reversible. The result lands in the draft, not in the send buffer. Edit it, undo it, or ignore it.

---

Requirements

Requirement	Notes
DeepSeek Harness	Desktop app or a web profile. Built against the bundled bundles; no source checkout needed.
A working model route	The plugin calls ctx.llm.stream() on an explicitly configured provider/model. Whatever you already chat with will do.
dsh-client-ui-conversation	Ships with DSH; declared as the client dependency.
dsh-commands	Ships with DSH; the plugin needs the commands service.

Nothing else. No build step, no dependencies, no npm packages of its own — the plugin is three source files.

---

Install

Step 1 — get the package into your profile

git clone https://github.com/AntMan-xtalgh/dsh-prompt-enhance.git
cd dsh-prompt-enhance

Then install it into a profile. Replace desktop with your profile name if different (ls ~/.dsh/profiles/).

dsh plugin --profile desktop add "$(pwd)"

If that fails with a pnpm workspace-root error, use the DSH-bundled pnpm directly:

PNPM="/Applications/DeepSeek Harness.app/Contents/Resources/runtime/primary-runtime/dependencies/pnpm/bin/pnpm.cjs"
cd "$HOME/.dsh/profiles/desktop"
node "$PNPM" add -w "file:/absolute/path/to/dsh-prompt-enhance"

Use DSH's own pnpm. A Homebrew or system pnpm with a different store major version fails with ERR_PNPM_UNEXPECTED_STORE because your profile's node_modules is linked from pnpm 11's store.

Step 2 — register the bundle

Open ~/.dsh/profiles/desktop/package.json and add the package to dsh.profile.bundles:

{
  "dsh": {
    "profile": {
      "bundles": [
        "@deepseek-ai/dsh-base",
        "@deepseek-ai/dsh-web-app",
        "… your existing bundles …",
        "dsh-prompt-enhance"
      ]
    }
  }
}

pnpm add puts the package in dependencies; this step is what actually tells DSH to load it.

Step 3 — point it at a model

Append to ~/.dsh/profiles/desktop/cordis.patch.yml:

- id: prompt-enhance
  name: "dsh-prompt-enhance"
  config:
    provider: moxing
    model: deepseek-flash-1017

Use the route you already run. If unsure, open the model picker in the composer and read the name — the left part is usually the provider, the right part the model.

The id must match the one in this package's cordis.patch.yml. That is how a profile-level entry overrides the bundle's inserted row rather than adding a second one.

Step 4 — restart DSH

Cmd+Q to quit fully (closing the window does not unload plugins), then relaunch.

The ✨ appears left of the model selector. If it does not, see Troubleshooting.

Verify it loaded

ls ~/.dsh/profiles/desktop/node_modules/dsh-prompt-enhance/
# client.js  cordis.patch.yml  index.js  locale  package.json

---

Configuration

All configuration lives in the profile patch from step 3.

Field	Type	Required	Meaning
provider	string	yes	Provider route id, e.g. moxing, deepseek-official
model	string	yes	Model id on that route, e.g. deepseek-flash-1017

Both must be set together. ctx.llm.listProviders() reports provider ids but not their models, and the session's own resolved route is not exposed to plugin code — so this pair cannot be auto-detected. See Known limitations.

The rewrite instruction is a constant in index.js (SYSTEM_PROMPT). Fork and edit it if you want a different rewriting policy — for example, always produce English, or always emit a bullet list.

---

Usage

1. Type anything into the composer — a fragment, a garbled sentence, a rough plan.
2. Click ✨.
3. Wait a moment. The draft is replaced in place with the rewrite.
4. Edit it if you want, then send as usual.

Notes:

• The draft is replaced, not appended. Undo (Cmd+Z) restores the original.
• Multi-line drafts are flattened for the model call, because a command line may not contain line breaks. Line breaks are collapsed to single spaces, not dropped.
• An empty draft does nothing except show a short hint.
• The button is per-session. Each session has its own composer and its own button state.
• Concurrent sends are safe — if a submission is in flight, the button is disabled rather than racing it.

---

How it works

Client (browser)                                Host (process)
─────────────────                               ──────────────

  ✨ click
      │
      ├─ read document.querySelector('[data-composer-input]').innerText
      │
      ├─ ctx.remote.commands.execute(
      │      sessionId,
      │      '/enhance <draft>',
      │      []
      │  )                                    →  ctx.commands.register({
      │                                             name: 'enhance',
      │                                             recordInput: false,
      │                                             handler: async ({ agent, rawInput, signal }) => {
      │                                               // resolve provider/model from config
      │                                               for await (chunk of ctx.llm.stream({ … })) {
      │                                                 if (chunk.type === 'text-delta') text += chunk.text
      │                                               }
      │                                               return { kind: 'success', text }
      │                                             }
      │                                           })
      │
      ├─ unwrap { ok, value: { commandId, result: { kind, text } } }
      │
      └─ inputActions.setDraft(rewrittenText)

The command uses recordInput: false, so the draft is not written into the session log as a user message. Nothing reaches the model's conversation history.

---

Design decisions

These are the choices that shaped the plugin, and why.

Registered into a list slot, not the activity slot

conversation.input.activity is exactly where WorkBuddy puts its sparkle — between the model selector and the send button. It is also, in DSH:

"conversation.input.activity": { kind: "single", scope: "session" }

Single-valued. It is the slot the bundled voice-input plugin claims for its microphone, and a control there can expand to fill the whole toolbar while active. Putting the button there would make the two mutually exclusive and squat on a reserved interaction.

"conversation.input.right": { kind: "list", scope: "session" }

List-valued. It sits immediately left of the model selector, takes multiple registrations, and conflicts with nothing. The button lands one control to the left of where WorkBuddy puts it — a fair trade for never breaking voice input.

A command as the transport, not a custom Remote

A browser plugin cannot call ctx.llm directly; that service lives in the Host process. Two transports exist:

Transport	Leaves a UI trace	Effort
Session command (ctx.remote.commands.execute)	Yes — one flow node	Small
Custom typert Remote (how voice input works)	No	Large

The command route was chosen to get a working plugin first. The cost is documented in Known limitations and is the one known deviation from WorkBuddy's behaviour.

Reading the draft from the DOM

The provide channel gives plugins inputActions — the write side (setDraft, insertText, submit) — but not the draft text. [data-composer-input] on the editor's contenteditable root is the only available source. It is official DSH markup (role="textbox", aria-multiline="true"), not an incidental class name, but it is still a coupling point.

Explicit model route

ctx.llm.listProviders() returns { id, name } only — no models. The session's resolved route lives behind request/header internals that are not reachable from plugin code. Rather than guess at unstable paths, the route is explicit config: predictable, debuggable, and one line to change.

---

Troubleshooting

Every failure path has its own message. Hover the red button to read it.

cannot get property "remote.commands" without inject

Your client.js declares inject without the qualified sub-service path. Cordis gates service access on declared paths, and parent and child are separate entries:

inject: ['slots', 'remote', 'remote.commands', 'locale']  // ✅
inject: ['slots', 'remote', 'locale']                     // ❌

Confirm with grep -n "inject: \['slots'" client.js.

No /enhance command: the Host half is not mounted.

The command did not resolve, so the Host half is not running. Check, in order:

1. dsh-prompt-enhance is in dsh.profile.bundles in the profile's package.json — installing the package is not enough.

2. ~/.dsh/profiles/desktop/node_modules/dsh-prompt-enhance/ exists.

3. You quit with Cmd+Q and relaunched. Closing the window is not a restart.

4. The Host half imports cleanly:

   cd ~/.dsh/profiles/desktop/node_modules/dsh-prompt-enhance
   node --input-type=module -e "import('./index.js').then(m => console.log(Object.keys(m)))"
   # → [ 'apply', 'inject', 'name' ]
   

The model returned an empty prompt.

The stream produced no text-delta. Usually one of:

• Reasoning model ate the budget. maxTokens is 4096 in index.js; raise it for a chatty reasoning model.
• A chunk-field mismatch after a DSH update. If DSH renamed chunk.type, the loop silently collects nothing. See Development notes.
• The route rejected the request but reported a normal finish. Try a different model in the config.

Prompt enhancement failed: …

The message after the colon is the underlying error. Common ones:

Suffix	Cause
NO_ADAPTER	provider in your config does not match any registered route
MISSING_CREDENTIAL / AUTH	The route's API key is missing or rejected
RATE_LIMIT	Backing off and retrying is worth it
CONTEXT_WINDOW_EXCEEDED	The draft is larger than the model's window

No model route: set provider and model …

Step 3 is missing or the id does not match. The patch entry must read id: prompt-enhance.

The enhance command was refused.

Admission rejected the line. Usually a mangled sessionId — reload the page.

The button is missing entirely

• Bundle not in dsh.profile.bundles, or DSH was not fully restarted.
• A client-side crash. Check the DevTools console (Cmd+Option+I in the desktop app) for errors containing dsh-prompt-enhance.

ERR_PNPM_UNEXPECTED_STORE

You installed with the wrong pnpm. Use DSH's bundled one — see step 1.

Edits to the source do nothing

file: dependencies are copies, not symlinks. Re-copy after editing:

cp index.js client.js "$HOME/.dsh/profiles/desktop/node_modules/dsh-prompt-enhance/"
# then Cmd+Q and relaunch

---

Known limitations

The command leaves one line in the transcript

DSH renders every admitted command as a persistent flow node. dsh-client-ui-commands hardcodes the fallback:

text ?? (node.outcome.kind === "error" ? t("command.failed") : t("command.done"))

So returning empty text still shows a generic "command done" row. Removing it requires a real typert Remote instead of the command channel — a deliberate larger change, not a bug.

The model route is explicit, not inherited

The plugin does not read the session's own request/header route. Configured once per profile; change it when you switch models.

Draft reading depends on a DOM attribute

[data-composer-input] is the only source. If DSH renames it, the button reads an empty draft and reports "type something first".

Multi-line drafts are flattened

Command lines cannot contain line breaks, so paragraphs are joined with spaces before the model call. The model usually restores structure; occasionally it does not.

Rewrites are not diffable

The draft is replaced outright. Undo restores it, but there is no side-by-side compare. A preview mode would need a confirm step, which trades away the one-click appeal.

---

Development notes

Everything below cost real debugging time. It is written down so it costs you none.

inject must name the full service path

Cordis gates access on declared paths. Reaching ctx.remote.commands needs the qualified path beside its parent namespace — declaring only the parent throws at call time, not at load:

inject: ['slots', 'remote', 'remote.commands', 'locale']  // ✅
inject: ['slots', 'remote', 'locale']                     // ❌

Reference: @deepseek-ai/dsh-client-ui-plan declares the same pair.

Stream chunks are discriminated by type

chunk.type === 'text-delta'   // ✅
chunk.kind === 'text-delta'   // ❌ silently yields empty text forever

BlockAssembler switches on chunk.type — same field. Reasoning routes emit reasoning-delta chunks first; only text-delta is the answer. Accumulating reasoning into the result produces a polluting preamble.

The terminal chunk carries reason

{ type: 'finish', reason: { kind: 'stop' | 'error' | 'aborted' | 'max-tokens', failure? } }

Not chunk.finish. Read reason.kind before reading reason.failure.

Command results nest three deep

{ ok: true, value: { commandId, result: { kind: 'success' | 'error', text, data? } } }

The wrapper layers differ between the Remote edge and the executor. Unwrap by locating kind, not by a fixed path:

let node = execution
for (let depth = 0; depth < 4; depth += 1) {
  if (typeof node?.kind === 'string') break
  node = node?.value ?? node?.result
  if (node == null) break
}

data is dropped in transit

dsh-commands' settle() forwards only kind, text, and sourceEventSeq:

return Object.freeze({ commandId, result: Object.freeze(result) })

The rewrite must travel in text. Putting it in data looks correct and arrives empty.

maxTokens, not maxOutputTokens

GenerateOptions sampling is limited to temperature, maxTokens, and stop. Unknown keys are ignored without complaint. Reasoning routes spend output budget on thinking first — 2048 can leave visible text empty; this uses 4096. An error settlement requires non-empty text; a success may omit it.

There is no top-level system option

Pass the instruction as a message:

messages: [
  { role: 'system', content: [{ type: 'text', text: SYSTEM_PROMPT }] },
  { role: 'user',   content: [{ type: 'text', text: draft }] },
]

Slot kinds are not interchangeable

Slot	Kind	Use
conversation.input.right	list	Extra toolbar buttons (this plugin)
conversation.input.left	list	Left-side toolbar buttons
conversation.composer.dock	list	Below the composer card
conversation.input.activity	single	Reserved: voice input
conversation.input.model	single	Reserved: model selector

Client plugin constraints

The browser half runs in a sandbox: plain JavaScript, no JSX, no TypeScript, no module imports. Its globals are React, console, styles, and (for dynamic plugins) host. Register factories via window.__ModuleLoader__.load({ id, factory }) with the id equal to the package name.

---

Project layout

dsh-prompt-enhance/
├── package.json          # manifest: exports, dsh.bundle, dsh.client
├── cordis.patch.yml      # inserts the plugin row into the profile
├── index.js              # Host half — registers /enhance, streams the rewrite
├── client.js             # Client half — the ✨ button and the RPC call
├── locale/
│   ├── en.json           # English strings
│   └── zh.json           # 中文字符串
├── docs/
│   └── composer.png      # the screenshot above
├── README.md
└── LICENSE

Both halves are small: index.js is the command registration plus one stream loop; client.js is the button component plus one RPC call and its result unwrapper.

---

Uninstall

cd ~/.dsh/profiles/desktop

Then:

1. Remove "dsh-prompt-enhance" from dsh.profile.bundles in package.json.
2. Remove the - id: prompt-enhance block from cordis.patch.yml.
3. Delete the package:

   rm -rf node_modules/dsh-prompt-enhance
   

4. Cmd+Q and relaunch.

The plugin writes nothing outside the profile: no state, no cache, no telemetry, no network calls except the model request you configured.

---

FAQ

Does this send my draft anywhere besides my configured model? No. One request to the route in your config, through ctx.llm.stream(). No telemetry, no analytics, no third-party endpoints.

Does the rewrite count against my context window? No. recordInput: false keeps it out of the session log, so neither the draft nor the rewrite enters the conversation history. Only the model call itself is billed.

Can I have more than one button here? Yes — conversation.input.right is a list slot. Copy this package, change name / id / the command name, and both will render side by side.

Can I change what the rewrite does? Edit SYSTEM_PROMPT in index.js. It is a plain constant. Re-copy into the profile and restart.

Does it work in the web profile? It should — the client half declares platform: "web" and uses no desktop-only APIs. It has been exercised on the desktop app only.

Why does English README describe a Chinese-first feature? The rewrite follows the language of your draft, whatever that is. locale/ carries both English and Chinese UI strings.

---

Contributing

Issues and pull requests welcome. Two asks:

• Report the exact hover message. Every failure path has its own string; it identifies the layer that broke.
• Say which DSH version you are on. Chunk field names and slot kinds are internal APIs and may move.

The highest-value contribution right now is the typert Remote replacement for the command transport, which would remove the transcript line. See Known limitations.

---

License

MIT