dsh-agent-teams, a plugin that turns one DeepSeek Harness session into a coordinated agent team, reached 1,424 GitHub stars in 26 days

dsh-agent-teams, a plugin that turns a single DeepSeek Harness session into a coordinated team of sub-agents, has reached 1,424 GitHub stars since developer NanmiCoder created the repository on 12 August. The counts come from the GitHub REST API, retrieved 7 September: 1,424 stars and 122 forks in 26 days, under an MIT licence, with the package published on npm.

The design makes the current session a captain. The captain assembles durable members, each backed by a continuable sub-agent, splits the goal into tasks with explicit dependencies, and consolidates the final result. Members message each other and the captain through durable mailboxes with no relay through the main conversation, and the plugin exposes eleven coordination tools in total. Team state lives on disk under the workspace's .agent-teams directory, and a live web panel reads that state and renders segmented progress, the member roster and an interactive task graph.

The scheduler is where the engineering sits. It claims one ready task per idle member atomically, using real running, idle and ready states, and a task cannot be claimed before its dependencies finish. An interrupted attempt stays parked and can resume through a direct message, reassignment revokes the old attempt and waits for the old worker to stop before a new one starts, and after a cold process restart the scheduler retries stranded open work with a fresh attempt.

Two features aim at control. A normal run stages an unspawned roster and task graph first, and only an explicit Approve and Run step creates members and starts scheduling, so the user sees the plan before any agent exists. Opt-in quality gates add per-task contracts that walk work through requirements, implementation, verification, review and integration, with automatic repair and re-review on failure.

The version pairing is the adoption cost, and the README states it plainly: the recommended combination is DeepSeek Harness 0.1.2-rc.1 with plugin 0.1.16-rc.1, and both are still prereleases. Desktop users must also check the app's embedded Harness core separately, because upgrading the global CLI does not upgrade it.

For a team using or watching DeepSeek Harness, the traction is the signal: 1,424 stars in 26 days for dependency-aware multi-agent scheduling says harness users want parallel teamwork with recovery semantics, and they are getting it from a community plugin ahead of the harness itself. The disk-backed state and the approval step carry the practical weight: the team survives a crash, and nothing runs before a person has seen the plan.

Source

Primary source: NanmiCoder/dsh-agent-teams on GitHub. Star and fork counts from the GitHub REST API, retrieved 2026-09-07.