Prompt for Me

中文

Prompt for Me (中文名：Prompt 嘴替) predicts the next message you may want to send from the DeepSeek Harness composer. After a completed agent turn, it quietly offers one suggestion as ghost text. Your draft stays empty until you accept it, and the plugin never submits on your behalf.

What it does

• Generates only after a newly completed turn when the session is idle, plan mode is inactive, and the composer has exactly empty text, no images, and no queued intent.
• Retains that completed turn while Host policy loads or the composer is temporarily ineligible, then generates once the requirements are met.
• Retires suggestions from the preceding context when a newer turn completes, including after submissions that settle too quickly for an intermediate composer state to render.
• Shows native inline ghost text on compatible Harness clients. Older clients use a small preview card without mutating the draft.
• Accepts a ghost with Tab, Right Arrow, or the adjacent check control. Enter never accepts ghost text.
• Hides the ghost when you type, offers it again if you clear the draft, and dismisses it with Escape.
• Keeps the Sparkles Trigger and Mod+Shift+Space shortcut. An explicit Trigger always requests a fresh suggestion and writes it directly into the draft, even while automatic generation is in flight.
• Sends skipped suggestions with the next explicit request so the model avoids repeating or paraphrasing them. The bounded cycle stores at most ten.
• Never bypasses Harness approvals, never invokes tools, and never sends a message automatically.

Interaction reference

Install

The release tarball is the simplest option because it contains prebuilt Host and Client artifacts:

dsh plugin --profile web add https://github.com/ChuanTianML/prompt-for-me/releases/download/v0.5.1/dsh-prompt-for-me-0.5.1.tgz

Restart dsh web after installation.

You may also install a pinned Git tag:

dsh plugin --profile web add github:ChuanTianML/prompt-for-me#v0.5.1

pnpm 10 may ask you to allow the package's prepare script for a Git install. Add dsh-prompt-for-me: true under allowBuilds in the Web profile's pnpm-workspace.yaml, then run the command again. The script only copies the checked-out Host files and wraps the checked-out Client factory; it performs no downloads.

Update or remove it with:

dsh plugin --profile web update dsh-prompt-for-me
dsh plugin --profile web remove dsh-prompt-for-me

Current DeepSeek Harness builds provide the native inline suggestion API. If that API is absent but completed-turn projection is available, the plugin falls back to an explicit preview card with Use and dismiss controls. The manual Sparkles workflow remains direct-fill in either presentation.

Model and API key

The plugin calls ctx.llm on the Harness Host. By default it reuses the current session's provider and model, falling back to the Harness default selection. The provider therefore uses the API key already configured in DeepSeek Harness. The browser never receives or reads that key, and this plugin has no separate key.

To pin an auxiliary model, set both provider and model in cordis.patch.yml or an overriding profile patch:

- id: prompt-for-me
  name: dsh-prompt-for-me
  config:
    provider: deepseek-official
    model: deepseek-chat

Data and privacy

On each automatic or explicit generation request, the Host may send these bounded text fields to the selected model provider:

• the current draft;
• the last three direct-human/assistant turns from the current session;
• current-session submitted suggestion edits, exact accepts, and rejected suggestions;
• bounded raw examples from manual prompts and suggestion interactions in up to 20 earlier sessions;
• up to ten suggestions skipped during the current cycle, so the model can avoid repeating or paraphrasing them.

The current draft and recent turns determine the task, intent, and message content. Current-session feedback adjusts the immediate wording. Cross-session memory may influence only durable style, detail, and workflow preferences. Manual prompts and submitted suggestion edits carry more weight than exact accepts; rejected suggestions are weak negative evidence. Editing a suggestion and triggering again rejects the original suggestion but does not treat the unsubmitted edit as a positive preference.

Harness records injected workspace instructions, runtime context, and skill catalogs in user-role events. The plugin excludes these non-human sources from conversation turns and preference memory. If both the draft and direct-human conversation are empty, generation stops before the model call.

Common API-key, token, password, and Bearer-token patterns are replaced with [REDACTED_SECRET] before the model call. Attachments, tool arguments, files, credentials, and binary blocks are not collected. The plugin has no analytics endpoint and sends data only to the model route already selected in Harness.

Interaction records are stored only in this browser's localStorage under dsh.prompt-for-me.outcomes.v2. Each record contains its session ID, final action, origin, and the relevant original/final text. Merely seeing or adopting ghost text is not positive feedback; it becomes evidence only after submission. Version 1 records migrate automatically and remain untouched. Clear both versions in the browser console with:

localStorage.removeItem('dsh.prompt-for-me.outcomes.v1')
localStorage.removeItem('dsh.prompt-for-me.outcomes.v2')

DeepSeek Harness 0.1.0-rc.6 does not expose downstream registration for custom durable session-event types. For that reason, the standalone plugin does not append its auxiliary model request or outcomes to the Harness session log; doing so would make persisted sessions unreadable to the stock runtime. This is the main difference from the experimental in-tree implementation and will be revisited when a public event-registration API exists.

The generation RPC uses NDJSON. Each complete candidate is validated before it reaches either the transient suggestion surface or the draft; partial model tokens and incomplete JSON never enter the composer. Background failures stay quiet. Explicit failures remain visible on the Sparkles control.

The auxiliary request always uses reasoning effort off. The model receives one system instruction plus one JSON user message with current, currentSessionFeedback, userPreferenceMemory, and currentCycleSkipped. It receives no tool schemas or attachments.

The Host retains the latest 50 privacy-safe performance records in memory and logs each record as prompt-for-me metrics. Records contain model route, text byte/item counts, history/input preparation time, first model chunk/reasoning/text times, suggestion arrival time, total model/request time, and provider token usage. They contain no prompt, candidate, or outcome text. Query the current process with:

curl -sS -X POST -H 'content-type: application/json' \
  -d '{"method":"metrics"}' \
  http://127.0.0.1:3080/dsh-prompt-for-me/rpc

Configuration

All generation limits are configurable in cordis.patch.yml:

Development

Requires Node.js 22.19 or newer.

npm run check

The command rebuilds the static Host/Client artifacts, runs the Node test suite, and verifies the npm package contents.

License

MIT