NanmiCoder/dsh-agent-teams

What it solves

dsh-agent-teams transforms a single DeepSeek Harness (DSH) session into a coordinated multi-agent team. It solves the problem of managing complex goals that require multiple specialized roles, dependency-aware task sequencing, and persistent state without needing a separate external workflow engine.

How it works

The plugin establishes a "Captain" (the main session) that manages a team of durable sub-agents.

1. Planning: The Captain splits a goal into a Directed Acyclic Graph (DAG) of tasks with explicit dependencies.
2. Delegation: The Captain assigns roles to members, who are continuable sub-agents. These members can be configured with specific LLM providers and models.
3. Execution: A shared scheduler atomically assigns ready tasks to idle members. Members can communicate via direct, durable mailbox messages to teammates or the Captain.
4. Monitoring: A live Web UI provides an activity panel showing the task DAG, member roster, and progress.
5. Quality Control: Optional quality gates allow for a structured pipeline of requirements $\rightarrow$ implementation $\rightarrow$ verification $\rightarrow$ review $\rightarrow$ integration.

Who it’s for

Developers and power users of DeepSeek Harness who need to automate complex, multi-step workflows involving multiple AI personas (e.g., reviewing code from performance, security, and product perspectives simultaneously).

Highlights

• Captain-led Coordination: Centralized delegation and result consolidation.
• Dependency-Aware Scheduling: Tasks are only claimed once their prerequisites are finished.
• Durable Sub-agents: Members are persistent and can be woken for follow-up turns.
• Integrated Web UI: Native workspace tab for visualizing task progress and member activity.
• Direct Messaging: Teammates can communicate directly without relaying through the Captain.
• Staged Planning: Allows users to review and revise the team roster and task DAG before execution.