Introduction

The core design philosophy of DeepSeek Harness (DSH) is plugin-based modularity, building agents by combining different capabilities. In real-world development, you sometimes need to integrate local code tools with DSH agents for collaborative development. The dsh-codex-peer plugin provides this capability, allowing DSH agents to collaborate as peers with a locally installed OpenAI Codex CLI. Each one maintains its own session, tools, and sandbox, and they interact through files in a shared worktree and tool calls—for example, one writes code while the other reviews it.

Plugin Overview

This is a DeepSeek Harness plugin for coordinating a DSH agent with a local OpenAI Codex CLI in a shared working directory.

  • Author: wuanthony397-hash
  • License: MIT
  • Category: workflow plugin

The plugin contains no code from OpenAI or DeepSeek; it only drives the Codex CLI that you install and log in to yourself.

Core Features

The plugin provides five core tools used to control the collaboration flow:

  • codex_ask: runs one Codex turn in the session working directory and decides whether to ask, plan, or implement based on the mode parameter.
  • codex_review: runs a read-only Codex turn and requires Codex to output structured review results (including verdict, severity ranking, files, and line numbers).
  • codex_status: reports the path and version of the Codex executable, effective settings, the shared work item list (task summaries and open tasks), and recent run records.
  • codex_plan: agrees on and records the split of objectives between DSH and Codex, generating a shared work item list that both read.
  • codex_task: manages the task list and supports operations such as listing, viewing, claiming, updating, recording evidence, setting budgets, and running.

Installation and Enablement

Environment Requirements

Before installing, make sure the following conditions are met:

  • DeepSeek Harness version >= 0.2.0-rc.2 (requires @deepseek-ai/dsh and @deepseek-ai/schemastery ^3.18.1)
  • Node.js version >= 20
  • Codex CLI is installed and logged in (the desktop Codex application does not automatically add it to PATH)

Installation Steps

This plugin is not published on npm. After installation, restart DSH to load the bundle.

# Install from the marketplace (requires marketplace listing)
dsh plugin --profile <profile> add dsh-codex-peer

# Download and install from GitHub (recommended)
dsh plugin --profile <profile> add github:wuanthony397-hash/dsh-codex-peer#v0.3.2

# Install from a local path (suitable for development or offline scenarios)
dsh plugin --profile <profile> add /absolute/path/to/dsh-codex-peer

After installation is complete, restart DSH. After the restart, the five tools mentioned above will appear in the available list.

Typical Usage

1. Validate the Environment

First, confirm that Codex CLI is correctly installed and discoverable.

# Have the agent execute this in the conversation
"check the codex peer status"

The agent calls codex_status and outputs the location and version of the Codex executable, as well as the current state of the shared work item list.

2. Plan Tasks

Decide who does what. The plugin supports multiple planning modes, for example:

  • Assigned division of labor: If you already know the split, you can ask the agent to record it directly.
    • Example instruction: “Split this between you two: Codex does the retry logic, you do the tests.”
    • Mechanism: The agent calls codex_plan with mode: "assigned" and writes the task split into the shared list.
  • Self-organizing collaboration: If the split is unclear, the two parties can negotiate.
    • Example instruction: “Work out between yourselves how to do X.”
    • Mechanism: The agent calls codex_plan with mode: "self-organizing", and both parties decide the division based on their capabilities.

3. Execute and Review

After planning is complete, enter the execution phase.

  • Implement: Let Codex perform the concrete code writing.
    • Example instruction: “have Codex implement the retry logic, then review what it wrote.”
    • Mechanism: The agent calls codex_ask in implement mode; Codex edits the file tree and reports the results.
  • Review: Review the generated code.
    • Mechanism: The agent calls codex_review, performs a read-only inspection of the workspace, and requires Codex to output a structured review report.

4. Manage the Task Flow

During execution, you can use the codex_task tool to manage individual tasks. For example:

  • codex_task action: list: view the current task list.
  • codex_task action: claim: have Codex claim and start processing a task.
  • codex_task action: run: run Codex under a specific task and record its results as evidence.

Applicable Scenarios and Notes

  • Applicable scenarios: Suitable for scenarios where a large, complex task needs to be split between two different agents (DSH and Codex) for collaboration, such as a “planning–implementation–review” pipeline workflow.
  • Notes:
    • The plugin only drives an external CLI and does not include any code; users are responsible for using Codex CLI.
    • Because the desktop Codex is not in PATH, ensure Codex CLI is correctly configured before installation, or that the plugin can locate it through a configuration file.
    • The plugin packages components at startup, so DSH must be restarted after installation for the plugin to take effect.

Summary

dsh-codex-peer uses a shared working directory and tool-call mechanisms to allow DSH agents and local Codex CLI to collaborate like two developers. By establishing the division of labor with codex_plan, and using codex_ask and codex_review to complete implementation and inspection, you can build a more efficient local agent workflow.