Foreword

DeepSeek Harness (DSH) adopts a plugin-based architecture. In the DSH ecosystem, traditional scheduling systems are usually passive: they respond only when receiving signals and remain idle otherwise. The dsh-proactive plugin overlays an autonomous layer on top of the “perception → decision → execution → reflection → consolidation” closed loop, enabling the system to observe its own state when no tasks are present, proactively explore unknowns, predict load, and perform proactive circuit breaking.

Overview

dsh-proactive is a workflow plugin maintained by beijingwahw. It implements proactive intelligent scheduling, with core capabilities including autonomous perception, autonomous decision-making, and autonomous evolution.

The plugin has built-in Scientist/Theorist dual mind, cognitive energy symbiotic economy, and Qualitative Change Kernel, aiming to solve the problems of autonomy and evolutionary capability deficiencies in traditional scheduling.

Core Features

Perception and Decision-making

The plugin provides Sentinel multi-source signal ingestion (supporting webhooks, file watching, polling, etc.), aggregate-window deduplication, and intelligent sorting. The strategic decision engine supports four-level decisions: execute, defer, dismiss, and ask-user, and continuously calibrates using statistical learning.

Dual Mind and Cognitive Economy

  • Scientist/Theorist dual mind: The Scientist kernel handles Bayesian optimal experiment design and evaluates the information value of experiments; the Theorist kernel performs induction based on hierarchical Bayesian methods and the MDL (understanding as compression) law.
  • Cognitive energy symbiotic economy: Includes an energy ledger (double-entry bookkeeping), knowledge market, belief market (LMSR), and agent contracts, simulating a symbiotic runtime.

Qualitative Change Kernel

The plugin implements a mental substrate with a shared statistical language, covering versions 3.0 through 16.0. Core capabilities include:
* Evidence and inference: Evidence kernel (3.0), causal inference (5.0), active inference (6.0).
* Optimization and prediction: Meta-reasoning (8.0), abstraction and generalization (9.0), any-time evidence (12.0), conformal prediction (13.0).
* Safety and fairness: Quality–diversity (14.0), runtime verification (15.0), Shapley attribution (16.0).

Memory and Evolution

  • Memory system: SQLite-based persistence, supports hybrid retrieval (FTS5 + vectors) and memory graphs, and includes a hallucination-resistant short-index mechanism.
  • Self-reflection: Capable of goal generation, quality reflection, metacognition layer, and strategy evolution, and supports long-term memory consolidation.

Engineering and Deployment

The plugin supports Raft consensus, distributed synchronization, hot updates, and AES-256-GCM encrypted storage. It provides multi-tenant isolation and dependency-free WebSocket real-time progress.

Installation and Activation

The plugin itself does not hold any API Key; keys are automatically injected. It supports zero manual configuration and is ready to use out of the box.

Prerequisites: Node.js ^22.18.0 || >=24.11.0.

Installation commands:

dsh plugin add beijingwahw/dsh-proactive --profile web
dsh web

Typical Usage

After installation, you can run development or build with the following commands:

npm run dev
npm run build

Use Cases and Caveats

This plugin is suitable for scenarios that require building scheduling systems with autonomous perception, decision-making, and evolution.

Note:
1. The plugin runs with the permissions of the current dsh process. Please review the source code and license before installation.
2. Key auto-injection priority: host ctx injection → process environment.
3. The build artifact dist/ is distributed with the repository; no extra build script is required during installation.

Conclusion

dsh-proactive transforms traditional passive scheduling into a proactive intelligent system by introducing the dual mind, cognitive economy, and Qualitative Change Kernel. Developers can use it to rapidly build multi-model collaborative workflows with self-reflection and evolution capabilities.

For more information, see: GitHub | SkillHub catalog