Introduction

In DeepSeek Harness, session-level interactions usually rely on text input. To introduce voice interaction, we need a product layer that can manage session state, drafts, and authorization. It needs to work within the current session, maintain the draft in the main input area, and execute the submission only after explicit user authorization, while also supporting full-duplex conversations with the Agent.

Plugin Overview

dsh-session-assistant is a full-duplex voice assistant product layer for the current Session. The plugin is maintained by AlexKaiqi and is licensed under the MIT License. Its core value is providing secure voice discussion and drafting capabilities: users can maintain drafts by voice and submit them to the main Agent after confirmation, with optional support for knowledge organization.

Core Features

  1. Session-level Voice Interaction: Provides a product layer for voice discussion and drafting in the current Session.
  2. Draft Maintenance: Maintains the draft in the main input area and supports editing before final confirmation.
  3. Secure Submission: Submits the draft to the main Agent only after explicit user authorization.
  4. Full-Duplex Conversation: Starts a full-duplex voice conversation with the Agent, supporting real-time interruption and feedback.
  5. Delegated Knowledge Organization: Supports delegating knowledge organization tasks to the Agent (requires dsh-personal-knowledge-base).

Installation and Dependencies

This version targets DeepSeek Harness 0.1.1-rc.2. DSH does not automatically install, activate, or update dependent plugins based on peer declarations; they must be added manually to the same profile.

Installation Command:

dsh plugin --profile web add dsh-multi-model-provider dsh-realtime-voice dsh-session-assistant

Dependency Requirements:
* dsh-multi-model-provider@^0.1.0-rc.11 (required)
* dsh-realtime-voice@^0.3.1 (required)
* dsh-personal-knowledge-base@^0.3.2 (optional, for knowledge integration)

Typical Usage

  1. Enable Knowledge Projection: If dsh-personal-knowledge-base is installed in the same profile, knowledge projection and organize_notes are enabled automatically.
  2. Voice Guided Tour: On the settings page, the “Voice Guided Tour” feature uses the selected Realtime voice to introduce the plugin and keeps the microphone enabled, allowing users to ask about capabilities, boundaries, and recommended workflows.

Notes and Architecture

  • Version Compatibility: Ensure that you are using DeepSeek Harness 0.1.1-rc.2.
  • Browser Dependency: Browser recognition depends on the browser’s SpeechRecognition capability; Chrome/Edge have the best compatibility.
  • Security: Long-lived Provider credentials remain on the Host only; the browser receives only validated route/profile data and short-lived session data.
  • Audio Handling: Unmatched audio is discarded immediately and is not persisted.
  • Wake Word: The standby wake word can be modified in settings.
  • Knowledge Base Integration: If dsh-personal-knowledge-base is enabled, the Host Remote reads the bounded personalKnowledge.project() before establishing a connection.

Summary

This plugin integrates session-level voice interaction, draft maintenance, and secure submission through a product layer, lowering the barrier to using voice input with the Agent. It relies on the underlying capabilities provided by dsh-realtime-voice and can extend knowledge organization functionality through the optional dsh-personal-knowledge-base.

GitHub Repository | Community Directory