dsh-prompt-optimize

When you keep re-reading a prompt to tighten the wording, dsh-prompt-optimize puts a small sparkle button right next to the composer. One click fires an auxiliary LLM call that rewrites the current draft under a fixed instruction, then overwrites the editor via inputActions.setDraft—and nothing else. No message is sent, no session event is appended, no new inference turn is started.

The safety boundaries are deliberately tight: the button disables itself when the draft is empty, contains a reference chip, the session has been removed, or the input machine is not in the plain phase. If you edit the text while the LLM is still working, the plugin abandons the replacement and shows a toast instead of silently overwriting your changes. Every request enforces configured byte and token ceilings and sets reasoningEffort to off, so the auxiliary call never incurs extra reasoning overhead.

It is built for developers who compose and iterate on prompts frequently within the DeepSeek Harness Web interface—anyone who wants a one-click polish without sending an extra message or breaking their flow. All configuration fields are explicitly declared (maxInputBytes, maxOutputTokens, timeoutMs, and an optional provider/model route) with no hidden defaults, making behavior predictable and easy to audit.