dsh-next-step

Grey ghost suggestions in the DeepSeek Harness composer: after a turn, the plugin shows the message you are likely to send next; Tab accepts it, deleting the draft brings it back.

┌──────────────────────────────────────────────┐
│ 帮我看看当前项目的结构                          │
└──────────────────────────────────────────────┘
        grey until you press Tab

How it works

The plugin is client-only (a dsh.bundle.patch package with a browser half), because a bundle installed from the Plugins panel cannot import harness packages on the Host side. Prediction runs in a small local service instead:

• the browser half watches the composer, and while the draft is empty asks the service for a suggestion every 15 seconds;
• the service receives the sessionId, reads that Session's own log under ~/.dsh/sessions, and predicts from the last human and assistant texts (injected context such as AGENTS.md is excluded);
• with DEEPSEEK_API_KEY set it makes one small model call (reasoningEffort: off, at most 160 output tokens); without it, a zero-token heuristic answers.

The suggestion never enters the conversation history.

Install

1. Install from the registry:

   npm install -g dsh-next-step   # optional: run the service by hand
   

2. In DeepSeek Harness open Plugins → Add plugin, enter dsh-next-step, install, enable, and restart the app. The bundle starts the service through the harness MCP client, so nothing has to be launched by hand.

   To run it manually instead: npx dsh-next-step (prints next-step: http://127.0.0.1:61760).

3. Open a new Session, send a message, and clear the input: the suggestion appears within 15 seconds.

Configuration

variable	default	meaning
NEXT_STEP_HTTP_PORT	61760	service port the browser half calls
DEEPSEEK_API_KEY	unset	enables model suggestions; otherwise the heuristic answers
NEXT_STEP_MODEL	deepseek-chat	model used for the suggestion
NEXT_STEP_LOCAL_ONLY	unset	1 forces the heuristic even with a key

Limitations

• A local service is required. Third-party bundles cannot call the model host-side yet; the platform gap is tracked in discussion #12742.
• Polling. Suggestions refresh every 15 seconds, not exactly at turn/end.
• Overlay positioning. The ghost is placed by measuring [data-composer-input]; multi-line drafts and theme changes may offset it.
• Heuristic mode is generic. Without a key the service can only offer continuations like "继续完成下一步".

License

MIT