JevDo: A Next-Generation Harness Architecture with Jev as the Scheduling Core

Introduction

During development, developers often find themselves repeatedly thinking through startup processes, searching for scripts, and assembling commands. JevDo is a next-generation plugin architecture built on DeepSeek Harness, designed to change this situation by introducing an independent scheduling core. It turns the large model from the sole executor into one of the options: Jev decides the next step, Actions accumulate experience, and the number of calls to the main model is reduced in repetitive work.

Core Features and Architecture

The core of JevDo lies in separating scheduling logic from execution logic.

  1. Jev (scheduling core): Reads the session history and decides whether the next step is to invoke an existing Action or to ask the main model to handle a new problem.
  2. Action (operation persistence): Stores reusable commands or scripts, including their purpose, project binding, and acceptance criteria, and persists them across sessions.
  3. Main model (problem solving): Handles only new problems and actively maintains the Action library during tasks.

The plugin provides two main tools:
* jevaction_create: Creates and stores an operation.
* jevaction_validate: Validates and activates an operation.

The workflow usually follows these steps: Jev selects an operation -> Harness executes and accepts it -> the result returns to the history; if an unhandled exception or new problem is encountered, the work is handed back to the main model.

Typical Use Cases

JevDo is suitable for development workflows that require repeatedly executing fixed operations.

1. Routine bug fixing
When fixing an order-list pagination bug, the flow is:
* Jev starts the backend and checks that the service is ready.
* The main model reads the issue, locates, and fixes the pagination logic.
* Jev runs unit tests and the build, and checks the exit code and artifacts.
* The main model explains the changes and verification results.

2. CI failure
When the CI lint check fails:
* Jev prepares dependencies and reproduces the lint check according to the CI configuration.
* The main model modifies the code based on the error messages.
* Jev reruns lint and performs pre-commit checks.
* The main model explains the changes and provides a commit message.

3. Release preparation
When preparing a release:
* Jev runs the full regression suite and checks release conditions.
* The main model reads the merge records, and updates the version number and CHANGELOG.
* Jev creates a tag, builds, and validates the artifacts according to project scripts.
* The main model organizes the release notes.

It has been verified that in subsequent turns, Jev can independently complete some tasks without calling the main model.

Usage

Using JevDo requires the prerequisite: Node.js 24.14+.

Because no explicit installation command is provided in the materials, integration usually needs to be done via source code or a package manager. After obtaining the source code, you can run the demo to understand its behavior:

  1. Clone the repository:
    git clone https://github.com/yinhong-zhou/jevdo.git
    cd jevdo
  1. Install dependencies:
    npm ci
  1. Run the demo (this step does not require an API key and is used to experience cross-session reuse):
    npm run demo

When connecting to a real environment, refer to the official documentation to configure Jev and the main model.

Notes and Limitations

The current features of JevDo are mainly focused on fixed commands and project binding:

  • Version requirement: Prioritized support for the headless use of DeepSeek Harness (DSH) 0.2.0-rc.1.
  • Package name: The tool prefix remains jevaction.
  • Validation scope: Arbitrary free parameters, a full Web UI, and multi-Agent combination compatibility are not within the verified scope.
  • Core entry point: During development and debugging, the core logic is located in src/dsh/index.ts.

Summary

By separating the scheduling logic, JevDo enables Jev to decide based on historical experience whether to reuse an Action or call the main model, thereby reducing invocation costs in repetitive development tasks. It is suitable for scenarios that require solidifying development scripts, but the current implementation boundary is mainly focused on fixed commands and project binding.

  • Project: https://github.com/yinhong-zhou/jevdo
  • Plugin catalog: https://www.skillhub.cn/plugins/yinhong-zhou/jevdo
羽毛球分组比赛记分
小程序二维码

欢迎使用《羽毛球分组比赛记分》微信小程序

Xiaoye