dsh-agents-md

English | 中文

A plugin for dsh that brings opencode-style custom agents to your conversations. You define agents as markdown files — each with its own system prompt and model — and then call them from chat with an @mention. The mention dispatches a subagent running under that agent's instructions and model, and the agent's reply comes back into the conversation as context.

How it works

Drop one file per agent into the agents directory (default ~/.dsh/agents/). The file name becomes the mention name; the YAML frontmatter carries the description (shown in the @ menu and to the model), an optional provider, and an optional model; the body is the agent's system prompt.

---
description: Reviews code for bugs and edge cases
provider: google
model: gemini-3-flash-preview
---
You are a senior code reviewer. Check for correctness, edge cases, and
security issues. Explain each finding and suggest a fix.

provider and model are the names dsh's own model routing uses — the ones your Models page (or a settings.yaml llm-* section) configures. Omit both to inherit the parent agent's route. A missing or unknown route fails the child's first request, so use ids your configured providers actually advertise.

Then type @reviewer <your request> in any conversation:

• The @ menu (same trigger the built-in subagent reference uses) lists your agents with their descriptions — pick one and the mention lands in the draft as @name.
• The main model reads the roster from its system prompt and, when a request names one of the agents, calls the call_agent tool itself with the agent's name and the task. The tool runs the agent as a subagent on the subagent provider (spawn by default): the agent's body becomes the child's system prompt, the frontmatter provider/model become the child's route, and the child inherits the parent's tool set — so the agent can call tools itself to do the work. You can watch it in the subagent catalog while it runs.
• The tool returns the agent's reply to the main model, which continues with it.

# In chat
@reviewer 检查这段代码有没有 bug

# The main model calls call_agent(agent: "reviewer", prompt: "检查这段代码有没有 bug")
# and receives the agent's reply from the tool result.

Install

# Install from npm (requires dsh >= 0.1.0-rc.6)
dsh plugin --profile web add @dingyi222666/dsh-agents-md
# Restart dsh web; the '@' agent source mounts automatically
dsh web

Notes:

• dsh plugin behaves like adding a dependency to your web profile. A bundle plugin is loaded once its full package name appears in the profile's dsh.profile.bundles list (adds automatically on recent dsh builds; add it manually if your build does not); the bundle patch applies on the next boot.
• With the repo source-launched CLI, run the args through the bin directly (node --import tsx/esm apps/cli/src/bin.ts ...).

Configuration

The plugin row accepts the usual cordis config keys (set them in your profile's cordis.patch.yml under the dsh-agents-md row's config, or with !!js expressions where needed):

Model Experience

• Added prompt content: one system-prompt section (order 95) listing the loaded agents (@name — description (route)), telling the model to call call_agent when a request names one. Empty roster → the section renders nothing.
• Added tools: call_agent (registered only when at least one agent is loaded) — agent (enum of the roster) + prompt. The model decides when to dispatch.
• Token costs: the roster section is proportional to the number of agents; each dispatch costs the child's own turn.

What's missing

• The model decides when to dispatch. Nothing forces a call_agent call; a strong model may answer a @name request directly. Keep the roster descriptions precise so the model routes correctly.
• Strict {{…}} persona rule. The agent body is used as the child's persona, where {{name}} is a strict prompt-variable reference. A file whose body contains a complete {{...}} group is skipped at load with a logged reason — a lone {{ without a later }} is fine. This matches dsh's own deployment-persona semantics.
• Only description, provider, and model in frontmatter. opencode's temperature, mode, and tools fields are not honored (dsh's subagent request has no temperature channel, and tool scoping is a separate capability).
• Live reload. The agents directory is watched: adding, editing, or removing an *.md file re-registers the call_agent tool's enum and refreshes the roster section automatically (debounced ~200 ms) — no restart needed. Editing the body or route of an existing agent updates the definitions the tool dispatches immediately.
• Mention boundary is name characters only. @types/react matches an agent literally named types — pick distinctive names.

Development

• yarn run build — builds the browser bundle (lib/client.js) and the Node half (lib/index.js + lib/invariant.js).
• yarn test — vitest suite: agent parsing, mention detection, dispatch orchestration, the real apply path, and the browser @ source.
• yarn typecheck — strict TypeScript over src and tests.