dsh-subagent-pro

When you spin up subagents in a DeepSeek Harness multi-agent workflow, they vanish into a black box—you can't see which model is running, what tools were injected, or whether the persona actually took effect. dsh-subagent-pro bundles three capabilities that previously required separate plugins into one: a live subagent monitoring panel, role-based routing with four-layer fallback, and Claude Code-style declarative persona files. A single HUD icon in the conversation input area gives you a draggable, resizable floating panel showing every active and finished subagent, its parent chain, model, and status—no sidebar toggling, no raw API polling.

Role routing is the other half of the story. The subagent_role tool lets the main agent delegate with a call > role > default > inherit fallback, injecting persona, provider, model, and toolFilter per call. Roles can live as JSON entries in the settings panel or as markdown files in ~/.dsh/agents/.md and project-level .dsh/agents/.md, where frontmatter carries metadata and the body becomes the subagent's system prompt. A built-in subagent_providers tool lets the main agent query the live LLM service for available providers and models at runtime, so nothing is hard-coded and no session restart is needed.

The plugin is zero-intrusion by design: with no roles or default model configured, it is indistinguishable from an uninstalled state. Settings changes hot-reload instantly, and agent markdown files re-scan on save. It is a natural fit for developers and small teams that regularly orchestrate multiple subagents in Harness, prefer managing agent personas as version-controlled markdown rather than JSON, and want a live dashboard to keep an eye on background runs without leaving the conversation.