Foreword¶

When running multi-step tasks in Harness, there are three common sources of chaos: tasks scattered throughout the conversation, making them hard to find later; unclear whether a task was proposed by the model or manually copied, making it hard to distinguish; and automatically checking off tasks when the model says “completed” without anyone actually confirming. dsh-task-manager addresses these three issues: collects tasks into a floating panel, labels the source of each annotation, and requires you to make the final decision when the model proposes completion.

What is this¶

dsh-task-manager is maintained by caipanjin, a plugin for DeepSeek Harness (DSH), published on npm, current version 1.1.5. It provides a floating task panel in the bottom right corner, placing model-suggested tasks and manually added tasks in the same place, and additionally handles three things: double confirmation is required to click “Start” before sending the task into the conversation; when the model reports completion, it only proposes without automatically finalizing; and the panel position is remembered by the browser. The README describes it as “a dependency-free small plugin.” The license is MIT.

Core Features¶

1. Two task sources: Tasks written by the model via the task_write tool are labeled “Model Suggested,” and annotations added in the panel input box are “Manually Added.” Each has a source badge.

2. Select text to create tasks: Select a paragraph of text in the conversation, and a “+ Add as Task” button floats nearby. When the selected content is a rendered list (-, 1., 2. etc.), the button changes to “+ Add N tasks,” creating a task for each list item. Since list markers are drawn by CSS ::marker and are not included in the selected text, the plugin reads the DOM <li> structure to identify items, rather than regex matching the text.

3. Floating panel: Located above the shell overlay, it expands from a compact capsule, is draggable to any position, and its position is saved by the browser.

4. Status dot and status text: Each row displays a status dot and the text “To Do / In Progress / Completed.” Colors are auxiliary (Gray for To Do, Amber for In Progress, Green for Completed).

5. Explicit completion, confirmation to finalize: Each row has only one action button that changes with the status (“Mark In Progress” → “✓ Complete” → “Reopen”). When the model judges a task complete, it only proposes; a banner appears on the panel saying “Confirm All / Ignore,” and it will not truly finalize until you confirm.

6. Double confirmation to start: Clicking “Start” on a row first pops up a confirmation bar; clicking “Send” actually sends “Start executing this task: