Overview

The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When developing agents with instructional capabilities, developers often need to handle complex interaction flows, such as how to provide learning support without interrupting normal conversation. The wsnxxxs/deepseek-harness-interactive-learning plugin provides a solution for this.

What It Is

@dsh-portable/interactive-learning adds a user-explicitly selected learning Agent preset and native learning activities to DeepSeek Harness. Through a declarative protocol and non-blocking flows, it enables the agent to provide instructional support within the default conversational flow, including semantic visualization, Reflective Pause, and fine-grained instructional state management.

Core Features

  • User-explicitly selected learning Agent preset: Supports multiple presets, such as Standard, Code, and Minimal, while keeping the tool schema and standing prompt unchanged.
  • Learn intent and first-turn routing: The system distinguishes between “building understanding” and ordinary tasks. Requests such as what is X, ELI5, flashcards, and study guide belong to learn intent and are routed to the learning path.
  • Non-blocking learning flow: The default path remains normal conversation, and learning_visual is called only when manipulation parameters can aid understanding. The old Question → Reveal dual-tool structure has been removed, preventing model turns from remaining in a running state while waiting for the user.
  • Reflective Pause: Implemented via dsh-learning/checkpoint@1. There is at most one pending checkpoint per session, used only when answering would change the next instructional action.
  • Multiple semantic visualization types: Supports types such as plot, node_link, scene_2d, relation, formula_steps, study_map, recall_deck, data_table, state_transition, sequence_buffer, sequence_diagram, code_trace, field_2d, and causal_loop via dsh-learning/visual@4.
  • Session-scoped LearnerState: Maintains instructional state, including the current goal, exposed priors, misconceptions or gaps, scaffolding needs, urgency/stuck evidence, assessment context, and a bounded failed-move history.
  • Design system and accessibility: All learning interfaces share a common set of design tokens and support keyboard and assistive technology usage.

Installation and Activation

The package is marked as private in package.json. Installation is primarily done through the provided binary dsh-learning-preset. To build from source, run the following commands:

pnpm install
pnpm run build
pnpm run test
pnpm run check
pnpm run test:package:purity
pnpm run test:package
pnpm run pack:check

Typical Usage

  • Default path and visualization: The default path remains normal conversation, and learning_visual is called only when manipulation parameters can aid understanding. The title of the visualization being prepared is shown where the tool call appears, instead of a generic waiting prompt.
  • Learn Intent routing: The system first distinguishes between “building understanding” and ordinary tasks. Requests such as what is X, ELI5, flashcards, and study guide belong to learn intent and are routed to the learning path.
  • Reflective Pause: Implements a non-blocking reflective pause via dsh-learning/checkpoint@1. The card title displays the cognitive action, such as prediction, explanation, or comparison, rather than internal mechanism labels.

Use Cases and Notes

  • Use cases: Scenarios where the agent needs to explain concepts, provide learning paths, generate flashcards, or perform structured debugging diagnostics.
  • Notes:
    • The plugin runs with the current DSH process privileges. Before installation, inspect the source code and license (no explicit license information was found in the provided materials).
    • At most one pending checkpoint is allowed per session.
    • Refresh and resume are folded into the same log.
    • The old Question -> Reveal dual-tool structure has been removed.

Brief Conclusion

This plugin provides native, non-blocking instructional capabilities to the DeepSeek Harness ecosystem. Through dsh-learning/visual@4 and dsh-learning/checkpoint@1, developers can build more refined interactive teaching flows. For more details, see the plugin catalog or the GitHub repository.