dsh-prompt-rewriter (DSH Input Box One-Click Prompt Rewriter)

DeepSeek Harness Plugin: A whale-tail button appears on the right side of the input box. Click it to use the current model to expand a draft into a complete task description with goals, format, and boundaries. The text is replaced in place but not sent; press Ctrl+Z to undo if you are unsatisfied.

DeepSeek Harness Plugin: A whale-tail button appears on the right side of the input box. Click it to use the current model to expand a draft into a complete task description with goals, format, and boundaries. The text is replaced in place but not sent; press Ctrl+Z to undo if you are unsatisfied.

Star count 37

Reasons

  • Only modifies the draft without sending the message: For example, "Help me write a Douban movie scraper" will be expanded to include scraping titles, ratings, and links from the Top 250, outputting CSV, complying with robots.txt, and enforcing a minimum request interval of 2 seconds; the rewritten result remains in the input box for further editing.
  • By default, it references previous context (indicated by a small dot at the bottom-right of the button), reading only images currently in the input box and not images from historical messages; you can specify a dedicated model for rewriting in settings, ensuring that switching conversation tasks does not interrupt the process.
  • The author documented implementation details: Injected history is truncated based on a budget of min(6000 tokens, context window × 5%), with history injected as a JSON array and declared as "data, not instructions"; released on 10-28, version 1.6.1 (10-30) is available on npm, marking the first inclusion in this column.

Usage

  • Requires Node ≥ 22: Run dsh plugin --profile web add dsh-prompt-rewriter (use --profile desktop for the desktop version), then restart Harness after installation (no hot reload).
  • Write a draft in the input box → Click the whale-tail icon to the left of the model name; the shortcut key is disabled by default and can be enabled in Settings → Prompt rewriter (Ctrl+E on desktop, Ctrl+Alt+E in browser).

Risks

  • Each rewrite triggers a real model call; referencing previous context consumes additional tokens, making it slower and more expensive in long conversations; the replacement is not streamed, so you must wait for the entire result to load.
  • Installing from GitHub source runs build steps; unit tests are not published with the repository; no license was found for the DeepSeek whale logo within the repository, which the author has explicitly excluded in the LICENSE file.

Open

GitHub

Explore the latest AI Radar updates