dsh-agent-team-gui

When you work with DeepSeek Harness on a daily basis, a single model often struggles to act as planner, coder, reviewer, and specialist all at once — forcing one model to do everything either muddles your prompts or widens tool permissions too far. dsh-agent-team-gui turns "multi-model division of labor" into a persistent, reusable product object: you create a team once in Settings → Teams, then reuse it across projects and conversations instead of assembling a one-off dispatch scheme every time. A team is not a temporary dispatch form; it is a first-class choice sitting right beside the conversation composer.

Each team member gets its own primary model, role prompt, fallback route, token limit, and tool allow/deny list. By default, the active conversation's model acts as a bounded planner: it breaks the request into structured assignments and produces an acyclic dependency graph (DAG). Members run with finite concurrency once their dependencies are ready, passing only bounded handoff summaries between stages, while complete raw output stays in durable run history. You can also enable fixed order to turn the team into a repeatable serial pipeline. The Team / Solo / Inherited conversation modes work alongside a crash-safe "next message" one-shot choice, and all of these states survive page refresh, cold Host startup, and live reconnect.

Safety and observability are core to this plugin. Planner, member, reviewer, and repair prompts all follow an exclusive-role, no-delegation contract; team children cannot spawn another team, and detected DSH subagent tools — including renamed registrations — are denied inside team sessions. One durable claim binds both automatic and model-tool dispatch to the latest human message, so repeated tool calls cannot create dozens of duplicate teams. An optional quality gate names one reviewer, one repair owner, explicit criteria, and at most two repair rounds; a rejection reruns only that repair owner. The Run Center exposes foreground/background state, live phase, elapsed time, child IDs, complete outputs, stop, and retries with linked immutable runs. Token usage reuses DSH's official projection with four buckets — uncached input, cache read, cache write, and output — coverage is explicitly labeled full, partial, or unavailable, and no invented prices are ever shown.

Who is this for? If you use DeepSeek Harness for complex tasks that need several specialist roles to collaborate, or you want to codify a repeatable workflow, reproduce it across projects, or share it with collaborators without exposing credentials, this plugin fits well. Team version snapshots, recipe import (Copy/Merge with route remapping), definition backup, and preview-first restore make team definitions as manageable as code. Upgrading never silently deletes existing run history or team versions — retention is deliberately opt-in. As a final note, follow the official README for installation; this plugin targets advanced DSH scenarios, so it helps to be familiar with provider/model route configuration and the Web profile structure before building your first team.