dsh-octo

English | <a href="README_zh.md">简体中文</a>

<a href="#prerequisites">Prerequisites</a> · <a href="#install">Install</a> · <a href="#what-it-does">What It Does</a> · <a href="#how-it-works">How It Works</a> · <a href="#documentation">Documentation</a> · <a href="#license">License</a>

Deepseek Harness (dsh) natively supports headless invocation of Claude Code and Codex as subagents through its built-in subagent backends, and these subagents can reuse the local account login state and the coding plans that come with the accounts. For example, if you have a Claude Code Max plan, you can have dsh call Fable-5 as a subagent.

Inspired by this, I designed this skill specifically for dsh, with these ideas in mind:

• Use built-in capabilities wherever possible for the most seamless invocation of heterogeneous subagents.
• Make dsh aware that it can call subagents, so it can aggregate multiple agents into greater intelligence.
• Design dedicated multi-agent workflows for different stages—such as planning, coding, and acceptance—to achieve high-quality collaboration.

Prerequisites

• dsh 0.1.0-rc.6.
• A deepseek-official provider configured for the main dsh model.
• Node.js, npm, and a pnpm environment available to dsh plugin.
• Claude Code and Codex installed and signed in locally, with active coding-plan subscriptions on the corresponding accounts.

Install

dsh-octo currently targets dsh 0.1.0-rc.6. Install it from a local tarball so dsh can resolve the bundle and all of its dependencies:

git clone https://github.com/tokentopo-ai/dsh-octo.git
cd dsh-octo
npm pack
dsh plugin --profile web add ./dsh-octo-0.1.0.tgz

Replace web with the profile you use, then restart it.

Once installed, use dsh exactly as usual: describe your task in normal conversation. No trigger phrase, subagent selection, or stage-by-stage command is required. The skill starts multi-agent collaboration only when it is useful.

What It Does

dsh-octo is a prompt-only orchestration contract, not an orchestration engine: it ships no workflow runtime and does not depend on ultracode or dsh Dynamic Workflow capabilities. The bundle only installs dependencies, registers the skill, and enables the delegation tools; every routing and stage decision is made by the main agent as it follows the skill's written instructions.

It only activates itself for tasks that genuinely need multi-stage collaboration:

• Activates — tasks that need a plan → code → review pipeline (with optional experiments), or that benefit from comparing multiple solutions or an independent review pass.
• Stays out of the way — single-point fixes, direct Q&A, and small changes that don't need a design decision. These continue on the main agent's usual path, with no overhead.

When it does activate, the main agent delegates to a fixed pool of official dsh subagent backends, one per stage:

Agent	Composition	Delegated tool
fable	Claude Code + Fable (xhigh)	subagent_claude_code
gpt-5.6-sol	Codex + gpt-5.6-sol (xhigh)	subagent_codex
deepseek-v4-pro	DSH child + deepseek-v4-pro	subagent_deepseek_v4_pro
deepseek-v4-flash	DSH child + deepseek-v4-flash	subagent_deepseek_v4_flash

A typical run walks through:

…