@aiwayds/dsh-subagent-registry

Configuration

Config field	Default	Description
agentsDir	~/.dsh/agents	Directory holding .md agent definitions.
provider	spawn	Subagent provider the child runs through (reuses dsh-base's spawn).
toolName	use_agent	Name of the registered tool.
leafDenyTools	[] (computed default)	Explicit tool-deny list installed on deep: 0 (leaf) children. Empty = computed default (every agent-spawning tool in the dsh base distribution plus toolName).
resume	auto	When use_agent continues a prior interrupted run: auto resumes whenever one exists, opt-in only when the caller passes resume: true, off never (an explicit resume: true still overrides).

Usage example

A multimodal example — ~/.dsh/agents/rubber-duck.md:

---
name: rubber-duck
display_name: 小黄鸭
description: "小黄鸭：多模态视觉 agent，看图识别、OCR、画 plotext/mermaid 图……"
model: digitalvolvo/kimi-k2.7-code
thinking: max
extensions: ["*"]
---
你是「小黄鸭」——多模态视觉 agent。……（这里写完整的 system prompt）

Then ask in a conversation:

用 rubber-duck 看一下这个浏览器截图，描述页面状态并提取文字

The main model calls use_agent(agent: "rubber-duck", prompt: "…"), the child runs with the file body as its persona and an image-capable model, reads the screenshot, and its result comes back into the conversation.

A text-only example — ~/.dsh/agents/workhorse.md:

---
name: workhorse
display_name: 牛马狗
description: "牛马狗：干活的主力……"
model: opencode-go/deepseek-v4-flash
deep: 0
---
You are 牛马狗，干活的主力。……（这里写完整的 system prompt）

Then ask in a conversation:

用 workhorse 把今天的发布清单整理成表格

The main model calls use_agent(agent: "workhorse", prompt: "…"), the child runs with the file body as its persona, and its result comes back into the conversation.