dsh-thinking-summary

The built-in "Think" row in DeepSeek Harness shows a single scrolling line while the model reasons, and a wall of raw text when expanded. For multi-thousand-token deep-thinking traces, extracting the key conclusions is nearly impossible. dsh-thinking-summary replaces that row with a three-stage collapsible view: collapsed shows a one-line summary of the last completed paragraph, semi-expanded chains per-paragraph summaries on a vertical rail, and fully expanded reveals the raw thinking text. Summaries are produced by a user-selectable model called with thinking disabled, low temperature, and a terse system prompt.

Paragraph splitting is tuned for real-world reasoning streams: blank-line-separated fragments under roughly 160 characters are merged forward into a single unit so that frequent line breaks do not fragment the summary. The in-progress tail always displays a "Thinking…" placeholder until it settles into a completed paragraph. Messages that were already finalized before the plugin mounted keep the plain raw-thinking disclosure unless the host already holds cached summaries.

Ideal for developers and AI engineering teams who work in the Harness web profile and regularly review model reasoning. You can judge whether the model is on track in three seconds, then drill into individual paragraphs or the full raw trace when needed.