dsh-plugin-subagent-roles

In DSH multi-agent pipelines, handing off code review, documentation, or architecture analysis to specialized subagents is the norm, yet keeping each role's persona text, model route, and tool scope synchronized across every delegation usually means hard-coding fragments into prompts. A single edit can silently widen a role's tool surface, and there is no single source of truth to audit. dsh-plugin-subagent-roles folds all of that into one Markdown file per role: the YAML frontmatter declares a routing description, an optional LLM route, and a tool policy, while the body is the persona the child agent runs with. Project-scoped roles live in /.dsh/roles/, global roles in ~/.dsh/roles/, and a project role takes precedence on id collision.

The plugin registers a single delegation tool and renders a compact catalog line during prompt assembly so the top-level agent can route by description alone. Role files are re-read both at assembly and at delegation time, meaning edits take effect immediately without restarting DSH. Tool policies support allow and deny lists with glob wildcards; any tool not explicitly granted is invisible to the child. The persona body may reference the three harness variables {{cwd}}, {{model}}, and {{provider}}, interpolated on the child side. A child-prompt trim step then strips out tool-guidance sections the child's policy does not grant, keeping the context lean and focused.

Built for developers who want composable, auditable multi-agent workflows in DSH: a handful of human-readable role files close the loop on delegation, model selection, tool isolation, and prompt trimming in one declarative layer, backed by a diagnostic tool and a browser settings card for live tuning.