The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” When building complex agent systems, you often need to combine runtimes, memory mechanisms, and tool chains. dsh-seed-society, as a workflow plugin, packages the Agent Society based on the Yogacara eight-consciousness architecture, memory consolidation tuning, and the MCP tool bridge into a standalone package.

Core Features

This plugin mainly consists of four parts:

  1. Memory Consolidation Fix
    A configuration fix for deep reasoning models. By disabling llm-deepseek.reasoningEffort and fixing the root cause related to autoDream, it resolves memory consolidation failures.

  2. Memory Consolidation Tuning
    Locks in the configuration for dsh-mneme: enables autoDream, sets a 32768 token context window, specifies the deepseek-chat model, and disables the session summarization feature.

  3. MCP Tool Bridge
    Establishes the mcp__society__* tool chain for connecting to various Society runtime operations.

  4. Seed Skills
    The package includes six integrated seed skill directories, including yogacara-society, alaya, manas, mano, panca, and sila.

Installation

Prerequisites

  • A DSH environment (must include @deepseek-ai/dsh-llm-deepseek).
  • Python 3.10+ (required by the MCP bridge and runtime).
  • @modusensus/dsh-mneme (Note: the dsh-mneme name in the registry is incorrect; the actual scoped package name is @modusensus/dsh-mneme).

Command-line Installation

It is recommended to use the official installation command:

dsh plugin --profile web add github:woshishadowhunter/dsh-seed-society

The installation script automatically handles dependencies, skill synchronization, and configuration dump validation. After installation, restart DSH.

Manual Installation

If you need to control the steps manually, execute the following in order:

# 1. 安装 Python 运行时
python -m pip install -e <repo>

# 2. 安装记忆巩固插件 (注意包名)
dsh plugin --profile web add @modusensus/dsh-mneme

# 3. 安装本插件
dsh plugin --profile web add github:woshishadowhunter/dsh-seed-society

# 4. 同步技能
# 将 skills/ 目录下的六个技能复制到 ~/.dsh/skills/ (或 ~/.agents/skills/)

Verification

After installing and restarting DSH, check whether the configuration took effect:

dsh --profile web --dump-config | grep -E "mcp-society|reasoningEffort|dreamMaxTokens"

After restarting, you can see the mcp__society__* tools in the model tool list, and observe in the mneme logs that the dream_runs status recovers from degraded to ok, indicating that memory consolidation is working properly.

Notes

  • Permissions and Boundaries: The plugin only composes configurations and bridges tools; it does not modify acceptance criteria, does not grant extra permissions to the model, and does not bypass approval workflows.
  • Environment Dependencies: mcp-society requires a local Python environment. The server does not use Shell and does not hold keys.
  • Audit Consistency: The LLM arbitration mechanism in mneme and its CAS/snapshot hash audit remain unchanged, ensuring auditability.

Summary

dsh-seed-society standardizes the complex Agent Society architecture and memory tuning solution into a pluggable DSH component. With a simple installation command, you can obtain an agent environment with the Yogacara eight-consciousness architecture, a dual-loop runtime, and an MCP tool bridge.

  • Project repository: https://github.com/woshishadowhunter/dsh-seed-society
  • Plugin directory: https://www.skillhub.cn/plugins/woshishadowhunter/dsh-seed-society