DSH (DeepSeek Harness) uses a plugin-based architecture. Traditional agents are often one-off, ending once a specific task is completed. This wastes cache hit rates and makes them difficult to manage. The dsh-lead-panel plugin introduces a “Lead” template, treating agents as independent conversation sessions. They can share updates and submit progress, similar to a “feudal” arrangement—the lead manages workers, but is not in turn controlled by the boss.

Core Capabilities

This plugin adds a Lead button next to the input box in the DSH web version. Clicking it opens a dialog where you can talk to the “Lead,” which breaks down tasks and dispatches them to worker AIs.

Panel Area Features

  • Model switching: The top shows the lead’s current status (running/idle) and a model dropdown. The list is grouped by provider, contains full model names with the current item checked, and selection switches the model.
  • Plan card: The lead summarizes four sections: 【Requirements】【Effect】【Division of labor】【Estimated time】, which are automatically extracted into a card for display.
  • Worker area: Displays the worker AIs dispatched by the lead. The list includes the name, status (running/inactive/provisioning/failed), assigned task, token usage and estimated cost for each, and the final conclusion.
  • Real-time metrics: The bottom shows five metric blocks: tokens / estimated cost / elapsed time / estimated remaining / tok/s, plus two buttons: Summarize and dispatch and Send.

Prerequisites

Team capabilities come from official Agent Teams (experimental). You must enable the “Agent Teams” toggle in the official settings. The plugin depends on ctx.agentTeams.listMembers and the usage projection service; otherwise, the panel cannot retrieve worker information.

Installation and Enablement

dsh plugin --profile web add dsh-lead-panel

After installation, you must restart DSH to load the Host half. The plugin follows the MIT license.

Typical Usage

1、Talk to the “Lead.” State requirements directly in the panel, and the lead will understand and record them.
2、Use the “Summarize and dispatch” command. This gives the lead a fixed instruction requiring it to answer in four sections and immediately use Agent Teams tools to assign work to worker AIs.
3、Switch the lead model. Use the model dropdown at the top to select from all local providers/models.
4、Monitor the process and cost. Use the worker area and bottom metric blocks in the panel to monitor token usage, estimated cost, and estimated remaining time in real time.

Limitations and Notes

  • Worker model switching: Currently, the worker model cannot be switched after creation. To change a worker’s model, the lead must be asked to create a new one.
  • Estimates: The cost and estimated remaining time are estimates. The plugin includes a built-in DeepSeek price list and converts values based on token count; estimated remaining time is extrapolated from the number of completed workers and their elapsed time.
  • Code modifications: DSH must be restarted when code changes are made. The Host module caches entries by the name in package.json, so simple modifications to cordis.patch.yml may not take effect unless you restart.
  • Official dependency: This plugin does not implement any team logic. It simply creates a regular DSH session as the “Lead” and displays the process from official Agent Teams.