dsh-dev-crew

Strong models for spec review, cheap models for bulk implementation—splitting work that way in an AI-assisted dev loop sounds obvious, yet the orchestration, polling, review rounds, and artifact bookkeeping quickly become a maze. dsh-dev-crew packages the entire collaboration into one repeatable pipeline: three built-in roles (implementer, reviewer, researcher), each bound to a provider and model of your choice, running inside a single conversation through brainstorm, pre-check, spec, plan, task-by-task implementation, three-round review convergence, and a final on-disk report. Only the spec discussion is a mandatory human checkpoint.

Beyond dispatching tools, the plugin ships four embedded methodology skills that encode the convergence protocol, versioned-artifact naming, and fix-then-edit strategy into the model-visible text. At the tool layer, a discipline gate blocks dispatching an implementer without a valid plan path, and a loop guard rejects pathological patterns observed in real sessions—consecutive list_agents polling, empty-option questions, waiting-period interrogations, and duplicate reviewer dispatch—returning a corrective reason to the model so it recovers on its own. Artifacts are saved as versioned copies per round; every review round gets its own file, and the final report can cite exactly which round the conclusion was reached in.

Built for developers already running DeepSeek Harness who want the plan-review-implement cycle to repeat reliably across projects. Configure roles, providers, and models through the Web settings panel, or via YAML for multi-model roles and custom personas; most fields hot-reload without restarting the session.