AnonyJcy/dsh-j-space

简体中文 | English

J-Space Cognition Suite SV1 native Agent Preset & standalone Cordis plugin for DeepSeek Harness (DSH).
Bringing internal thought representations, externalized workspace ledgers (.jspace/), and adaptive verification to unlock full LLM reasoning potential.

---

Compatibility: adapted to DSH 0.1.6 dsh-workflow-ptc and DSH 0.1.5 dsh-persona schema (config.prefix / config.suffix).

🌟 Overview

dsh-j-space integrates the J-Space Cognition Suite (SV1 release, continuing V3.7 evolution) into DeepSeek Harness as a native Agent Preset.

Unlike traditional flat prompt injections, this plugin provides full agent scope isolation, multi-tier reasoning routes, externalized workspace ledgers (.jspace/), and adaptive verification across any compatible LLM model (DeepSeek, Claude, GPT, etc.).

---

📊 Empirical Capability Realization Report

Full evaluation reports by the original author:

• Current Benchmark Report: GLM-5.3-Flash-J-Space-Capability-Realization-Report
• Earlier Comparative Report (Preserved Archive): DeepSeek-V4-J-Space-Capability-Realization-Report

🔬 Benchmark 1: GLM-5.3-Flash Evaluation (Latest Report)

1. Main Benchmark Table (Accuracy Comparison)

Benchmark	GLM-5.3-Flash (Baseline)	GLM-5.3-Flash + J-Space V3.7/SV1†	GLM-5.3	Opus-5	Fable 5.1
HLE (w/ tools)	55.3	59.2	62.5	64.7	65.0
Terminal Bench 2.1	84.3	88.8	88.2	*89.1	*91.4
DeepSWE v1.1	63.4	68.0	66.9	68.8	67.4
Agents' Last Exam	26.3	30.5	28.5	31.6	—
AutomationBench (Public)	48.8	51.1	48.2	50.3	—

* Note: Terminal Bench 2.1 figures for Opus-5 and Fable 5.1 are independently measured by a third party; no official entries exist.
\† Estimated, based on limited controlled experiments.

2. Speed and Token Efficiency Table (GAIA Controlled Pair)

Metric	Factor / Improvement
Speed	1.87×
Token Efficiency	1.41×

---

🔬 Benchmark 2: DeepSeek-V4-Flash Evaluation (Historical Archive)

• Base Model: DeepSeek-V4-Flash-Vision-Exp
• Harness: DeepSeek Harness (Standard Mode)
• Methodology: Rigorous A/B Testing with and without J-Space on authoritative benchmark subsets and same-type mini-sets (Terminal-Bench 2.1: 20 medium / 10 hard; DeepSWE: 10 TypeScript / 10 Python / 10 Go / 2 JavaScript / 2 Rust; GAIA: level 1 / level 3, etc.), with identical model, environment, and sampling — only the J-Space toggle differs.
• Evaluation Dimensions: ① Accuracy / Pass Rate; ② Wall-clock & Token Efficiency.

1. Main Benchmark Table (Accuracy Comparison)

Benchmark	DeepSeek V4-Flash (Baseline)	DeepSeek V4-Flash + J-Space V3.7	GLM-5.3	Kimi-K3	Opus-4.8	Fable 5 (w/ fallback)
HLE (w/o tools)	*37.8	37.8	—	43.5	49.8	53.3
HLE (w/ tools)	*51.5	51.9	62.5	56.0	57.9	63.0
Terminal Bench 2.1	83.9	85.5	88.2	88.3	85.0	88.0
NL2Repo	57.7	60.4	58.0	58.0	69.7	—
CyberGym	75.3	77.8	84.5	80.0	78.3	83.1
DeepSWE	59.3	61.8	66.9	67.5	58.0	70.0
Toolathlon-Verified	75.9	77.4	73.0	76.5	76.2	77.9
Agents' Last Exam	27.3	28.3	28.5	27.6	25.7	23.8
AutomationBench (Public)	25.7	27.6	48.2	30.8	27.2	29.1
⭐ Average Score	56.99	58.61	64.54	60.96	58.33	62.13

* Note: HLE scores were not disclosed and follow DeepSeek V4-Flash-0731. The average covers the 7 rows where all six columns have values.

2. Speed and Token Efficiency Table

Benchmark	Wall-clock τ	Speedup	Output Tokens	Total Tokens	Score per Unit Time	Cost per Successful Task
HLE (w/o tools)	*1.02	−2%	−10%	+5%	0.98×	+5%
HLE (w/ tools)	0.88	+14%	−22%	+3%	1.15×	+2%
Terminal Bench 2.1	0.79	+27%	−28%	−3%	1.29×	−5%
DeepSWE	0.78	+28%	−28%	−3%	1.34×	−7%
Toolathlon-Verified	0.86	+16%	−25%	+2%	1.19×	+0%
AutomationBench (Public)	0.76	+32%	−31%	−5%	1.41×	−12%

** Note: For HLE (w/o tools), τ=1.02 is intentionally positive (i.e. slower) because on single-turn tasks without tools, injecting the full skill entry is net overhead. On long-horizon and multi-turn coding/agentic benchmarks (e.g. Terminal Bench, DeepSWE, AutomationBench), J-Space delivers +14% ~ +32% faster execution, cuts 28%~31% of output token redundancy, and boosts score per unit time by 1.15× ~ 1.41×.*

---

🚀 Installation & Deployment

Method 1: Install from npm / pnpm (Official Registry)

# via npm
npm install -D @anonyjcy/dsh-j-space

# via pnpm
pnpm add -D @anonyjcy/dsh-j-space

# Deploy preset to ~/.dsh/.agent-presets/j-space
npx @anonyjcy/dsh-j-space install

Method 2: Direct Clone & Install (Local Use)

git clone https://github.com/AnonyJcy/dsh-j-space.git
cd dsh-j-space

# Deploy J-Space preset to ~/.dsh/.agent-presets/j-space/
node bin/cli.js install

# Check status
node bin/cli.js status

---

💡 Usage

1. In DeepSeek Harness Web UI

1. Create a new Session.
2. Select J-Space Cognition Suite in the Agent Preset dropdown.
3. Pick any compatible model (deepseek-chat, deepseek-reasoner, etc.) and start your task.

Enable model selection for spawn subagents in Web UI

J-Space spawn subagents use DSH's official tool-subagent. The public preset leaves this option off by default to remain compatible with DSH profiles that do not load the Host settings plugin. To let the Agent choose a model and reasoning effort for each delegated task, add this to the preset's tool-subagent config:

config:
  provider: spawn
  toolName: subagent
  modelSelectionSettings: true
  backgroundMode: continuable

Then open Plugins → Subagent → Model selection in DSH Web Settings, enable Allow agents to choose models for Subagents, and select permitted routes from the current DSH model catalog. Start a new session for the setting to take effect.

Routes come from each DSH deployment's own model catalog and authorization list; J-Space hardcodes no provider or model IDs. Select newly added routes in Subagent settings when needed. This option requires the Host composition to load @deepseek-ai/dsh-tool-subagent/model-selection-settings; leave modelSelectionSettings unset in profiles without that Host plugin. The setting applies to new sessions. DSH fork subagents inherit the parent session's model by design.

2. In DeepSeek Harness CLI

dsh --preset j-space "Analyze this architecture and implement feature X"

3. In Cordis Composition (cordis.yml)

- id: j-space-plugin
  name: '@anonyjcy/dsh-j-space'
  config:
    autoDeploy: true

---

🧩 Architecture & Data Flow

flowchart TD
    A[New Session] --> B[Select j-space Preset]
    B --> C[Preset Discovery: AgentPresets.list]
    C --> D[Preset Mount: AgentPresets.mount]
    D --> E[Agent Scope]
    E --> F1[Persona: J-Space SV1 Architecture]
    E --> F2[Tools: Full Coding & Reasoning Tools]
    E --> F3[Skill Filesystem: Mounted skills/j-space/]
    E --> F4[J-Space Suite: SKILL.md, 13 modules, 7 references, controller & adapters]
    E --> G[Session Model Route: Any Model]
    G --> H[Agent Executes J-Space Cognition Loop]
    H --> I[Task Workspace: Managed .jspace/ Ledger & Control State]

---

🛠️ CLI Commands

node bin/cli.js install    # Deploy J-Space preset to ~/.dsh/.agent-presets/j-space
node bin/cli.js uninstall  # Cleanly remove J-Space preset
node bin/cli.js verify     # Verify integrity of installed preset files
node bin/cli.js status     # Display current installation status

---

📄 License

MIT License. See LICENSE and THIRD_PARTY_NOTICES.md.

Maintenance

See CHANGELOG.md for release notes.