dsh-cbx-orch

When you are driving multiple external coding CLIs such as codebuddy, opencode, omp, cline, and qwen, what is usually missing is not more prompts but a unified orchestration and governance layer: who runs first, who runs after, who reviews the output, what happens on crash, and how quota is enforced. dsh-cbx-orch ports the persistent orchestration engine from cbx-orch into the DeepSeek Harness plugin mechanism. Task state machines, queues, retries, review gates, and artifact archives are all persisted to disk; after a process crash, dead workers are reclaimed, completed stages are skipped, and remaining tasks resume automatically instead of replaying the entire chain.

The routing layer is another core capability. The plugin detects locally installed coding CLIs, filters candidates by declared capabilities such as auto-approve, plan-mode, and sandbox, then scores them with multi-factor strategies like capability-best, cost-aware, fastest, and round-robin to pick the most suitable executor. A sliding-window health tracker continuously down-weights repeatedly failing executors. Seven governance gates-approval, review, per-task invocation budget, global concurrency and quota limits, policy-drift fingerprinting, event-tamper audit, and workspace whitelist with isolated worktrees-provide layered safeguards so delegated agents operate within controlled boundaries.

It is designed for engineers and teams whose Harness workflows involve multi-executor collaboration, task traceability, and crash recovery. If your daily work is not single-turn Q&A but a submit-queue-execute-review-retry pipeline, this plugin eliminates the hand-rolled state management that used to glue it all together.