dsh-agent-hub

Coordinating multiple AI agents on a real project often means chasing scattered outputs, opaque intermediate states, and failures that are hard to reproduce. dsh-agent-hub collapses the entire workflow into a single group-chat room: a requirement comes in, a verified deliverable comes out, and every step of decomposition, dispatch, execution, acceptance, and knowledge capture is visible and traceable in the message stream—no external services and no manual backend setup needed.

At the centre of the system sits a coordinator brain that you assign to any member. Each round follows a strict decision-dispatch-collect-accept-memorise state machine; before emitting instructions the brain must answer four questions—goal, current state, available members, and definition of done—and output structured JSON. Tasks are split into atomic units where one member owns exactly one deliverable plus one acceptance criterion, and independent tasks run in parallel. Members connect through any OpenAI-compatible API, so DeepSeek, Tongyi, Kimi, GPT, and others can coexist in the same team. After delivery the brain verifies every acceptance criterion, issues a precise rework list on failure, and writes passing results into a persistent group memory that carries across rounds.

The plugin ships with a self-contained backend, carries zero npm runtime dependencies, and is ready the moment it is installed. It is built for developers and engineering teams who need multiple agents to collaborate end-to-end: scaffolding a feature from scratch, comparing and implementing technical options, reviewing and fixing code, or running parallel research and consolidating it into a final document—each scenario closes the loop from requirement to verified delivery inside a single chat window.