@custom/dsh-plugin-j-space

English | 中文说明

J-Space Cognition Suite (V3.6) is an inference-time cognitive control plugin designed for DeepSeek Harness (DSH). It enhances the reasoning fidelity, context persistence, and verification discipline of LLMs (especially DeepSeek-V4-Pro / Flash and Kimi models) during complex, long-horizon tasks.

🌟 Why J-Space?

During long multi-step reasoning, coding, and autonomous workflows, large language models frequently suffer from four major inference-time losses:

1. Working-Set Overload: Too many active constraints dilute attention.
2. Representation Drift: Global invariants, architectural definitions, or goals gradually mutate across steps.
3. Uncontrolled Retry: Repeating failed routes without carrying diagnostic hypotheses.
4. Premature Completion: Mistaking fluent conversational output for verified execution.

J-Space turns the model's accessible working memory into a structured, actively managed internal workspace without changing model weights or requiring fine-tuning.

⚙️ 6 Core Cognitive Mechanisms

🚀 Installation

Compatible with all operating systems (macOS, Linux, Windows).

Method 1: DSH CLI (Recommended for Local & Server)

Inside your DSH profile or project workspace, run:

# Install directly from GitHub
dsh plugin add github:kolawong/dsh-plugin-j-space

# Or link from a local folder
dsh plugin add ./dsh-plugin-j-space

Method 2: Manual Setup (Local / Server / Container)

If you are cloning manually into your user directory (~/.dsh):

macOS / Linux:

# 1. Clone into your DSH plugins directory
mkdir -p ~/.dsh/plugins ~/.dsh/skills
git clone https://github.com/kolawong/dsh-plugin-j-space.git ~/.dsh/plugins/dsh-plugin-j-space

# 2. Symlink skill to the global skill directory
ln -sfn ~/.dsh/plugins/dsh-plugin-j-space/skills/j-space ~/.dsh/skills/j-space

# 3. Restart DSH
# On local desktop/CLI: restart your 'dsh web' or app process
# On systemd server: systemctl restart deepseek-harness

Windows (PowerShell):

# 1. Clone into your user profile DSH plugins directory
New-Item -ItemType Directory -Force -Path "$HOME\.dsh\plugins", "$HOME\.dsh\skills"
git clone https://github.com/kolawong/dsh-plugin-j-space.git "$HOME\.dsh\plugins\dsh-plugin-j-space"

# 2. Create directory junction for skill
New-Item -ItemType Junction -Path "$HOME\.dsh\skills\j-space" -Target "$HOME\.dsh\plugins\dsh-plugin-j-space\skills\j-space"

# 3. Restart your DSH process

🎛️ 4 Runtime Modes & Activation Policy

You can switch the operating mode in the DSH Web UI (Settings ➔ Plugins ➔ J-Space) or in ~/.dsh/settings.yaml:

j-space:
  mode: on-demand # Options: on-demand | always-on | auto | off

• 🎯 On Demand (on-demand) [Default]: Automatically activated upon explicit user prompts or during complex multi-step reasoning.
• ⚡ Always On (always-on): Global cognitive workspace injected into every turn context.
• 🤖 Auto (auto): Autonomously triggered based on task complexity and code depth.
• ⛔ Off (off): Completely disables the cognitive framework.

💬 Usage Examples

In DSH conversations or Taskboard AI drawer:

• Implicit activation: In on-demand mode, the model automatically loads j-space when you ask for architecture refactoring, bug tracing, or theorem proving.
• Explicit trigger:

  "Use the J-Space cognitive framework to plan and verify this database migration." "Activate J-Space Dense Track to audit cross-file consistency for the auth module."

🌐 Internationalization (i18n)

The plugin UI card (client.js) features built-in automatic language detection (English / 中文), adapting automatically to your DSH locale and browser language without any configuration needed.

🔬 Scientific Background & Original Attribution

This project is built upon empirical research on language-model internal representations and packages the original open-source J-Space Cognition Suite V3.6 by Tiger3807861189 as a standard DeepSeek Harness plugin.

Scientific Foundation

• Gurnee et al., Anthropic (2026) — Research on privileged internal representational workspace ("poised-to-say" representations).

Citation

@software{j_space_cognition_suite_2026,
  author = {Tiger3807861189},
  title = {J-Space Cognition Suite: Model-Agnostic Inference-Time Control Suite},
  year = {2026},
  version = {3.6.0},
  url = {https://github.com/Tiger3807861189/J-Space-Cognition-Suite-V3.6}
}

📄 License & Open Source Compliance

Licensed under the Apache License, Version 2.0 (the "License").

• You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.
• Includes upstream work from J-Space Cognition Suite V3.6 (c) 2026 Tiger3807861189.
• See the LICENSE and NOTICE files for complete attribution details.