dsh-next-step
Find your next step. Refine your prompt.
简体中文 · Download · English video · 中文视频
dsh-next-step — inline predictions for DeepSeek Harness
A community plugin for DeepSeek Harness desktop and web. After a reply finishes, a suggested next message appears in gray inside the composer. Press Tab to turn it into an editable draft. Have a draft already? Click the magic-wand icon to refine it, then review and send when ready.
See it work
Watch the demo
- Inline suggestions: same composer, same font, no extra suggestion panel.
- One-click prompt improvement: write a draft, then click the magic-wand icon to turn it into a well-described, actionable spec. Explicit requirements stay intact; routine gaps receive reasonable Proposed defaults so work can begin with minimal questions. Review and edit before sending.
- Tab to accept: the prediction becomes a normal draft. It does not send itself.
- Esc to dismiss: keep writing your own message.
- Your draft comes first: typing, attachments, queued messages, and a new turn hide the prediction.
- Plugin-controlled: predictions are active whenever the plugin is enabled. To turn them off, disable the plugin in DSH’s plugin manager.
- Existing model setup: uses DSH's configured provider and credentials. No separate API key to enter.
The real magic-wand icon and an improved, unsent draft
dsh-next-step enabled in Plugin Manager
The improvement screen comes from v0.3.0 verification. Prediction text details are crops of an earlier demo; the old toggle is excluded. Media sources · All assets
Install
No build step or runtime npm dependencies are required.
- Download the source archive and extract it.
- Add the extracted plugin directory in DSH’s Plugin Manager.
- Enable dsh-next-step there. Use the same manager to disable it.
Both features follow the plugin state, with no separate composer toggle. Reload an updated local installation through Plugin Manager; no separate startup entry point is needed.
Improve a draft
The optimizer turns a brief into a compact, executable spec: goals, requirements, a default approach, behavior, and observable acceptance checks. It may add proportionate, low-risk defaults for ordinary gaps, labeled Proposed defaults so they remain easy to adjust. It preserves explicit exclusions and does not expand a small task into an unrelated product.
For a vague website request, it can propose a first-version page structure, visual direction, responsive behavior, and implementation approach instead of returning a questionnaire. Existing project conventions take priority. Missing real names, content, or data use labeled placeholders, not fabricated facts.
The optimizer does not request a confirmation section. Missing sensitive values use placeholders and safe execution constraints; routine gaps use proposed defaults. If a returned spec still has a recognized confirmation-section heading, it receives one repair request using the original draft. A second such result is rejected without replacing the draft. Both attempts share the existing timeout. Complete, simple requests stay concise.
The button displays a working label while a request is pending. Successful edits appear directly in the composer without a success banner. Unchanged output still reports that the model suggested no changes and leaves the draft revision untouched.
Click the magic-wand icon beside the composer controls. The request uses your configured model and sends only the current draft. The result replaces the draft without sending it. If you edit, send, switch sessions, or disable the plugin while a request is pending, its result cannot overwrite your new text. Errors preserve the original draft.
Failures explain whether the cause is a timeout, missing model, authentication, exhausted quota, rate limiting, network/service errors, or invalid/truncated output. DSH terminal stream failures keep their original error category. Only safe error categories and HTTP status codes are displayed; provider response bodies and credentials are never forwarded. Cancelling by editing or switching sessions stays silent. Service errors are not automatically retried. A recognized confirmation section can trigger one additional model request to repair the spec, using the same configured service and shared timeout.
The icon is disabled for empty drafts, drafts over 12,000 characters, active turns, and drafts containing structured file/session reference chips. Separate attachments are retained. Each click makes an additional model request and uses the service’s normal quota or billing.
Keyboard
| Key | Action |
|---|
| Tab | Accept the visible prediction as a draft |
| Esc | Dismiss the current prediction |
| Shift + Tab | Keep normal keyboard focus navigation |
Predictions appear only when the session is idle and the draft, attachment list, and message queue are empty. IME composition does not trigger acceptance. Gray text is a display layer; it is not inserted into the editor document until accepted.
Model calls and data
Generating a prediction makes an additional request to your configured model service and consumes that service's normal quota or billing. This plugin does not make inference local or free.
The request contains up to 10 recent user/assistant text messages, capped at 12,000 characters. Reasoning blocks, tool results, and attachments are excluded. Predictions and the auxiliary prompt are not appended to the conversation transcript. Results are cached per completed turn in host memory; a restart clears the cache. Errors, timeouts, and unsuitable results leave the composer unchanged.
Configuration
By default, the plugin reuses the current session's provider and model, with the model's default reasoning configuration. To choose another already-configured route, add this entry to your profile's cordis.patch.yml:
- id: next-step
config:
timeoutMs: 25000
# provider: your-configured-provider-id
# model: your-model-id
# reasoningEffort: a-supported-effort-id
Set provider and model together. Timeout range: 1–60 seconds. Output cap: 512 tokens; accepted suggestions: at most 280 characters. No tools or agent actions run during prediction.
Compatibility and development
Verified with macOS DSH desktop 0.2.1-rc.2. The media demo uses the matching DSH web runtime. Other versions may change plugin APIs.
npm run build # JavaScript syntax checks
npm test # behavior tests
npm pack
index.js: authenticated Connection Fetch RPC, shared by desktop IPC and web HTTP.engine.js: context selection, model call, deduplication, and invalidation.client.js: native DSH client module, composer ghost text, and prompt-improvement icon.optimizer.js: draft-only rewriting, cancellation, and output validation.
The screenshots and videos use the actual DSH frontend and this plugin, with a scripted demo provider and synthetic conversations. They are not evidence of live-model prediction quality. See media provenance.
English launch copy · 中文宣传文案
MIT. Community project; not an official DeepSeek or OpenAI product.
