dsh-swarm-orchestrator

When you hand a big goal to a single agent, it can only work serially: research queues behind quick edits, context balloons, and nothing independent checks the output. dsh-swarm-orchestrator takes a different approach - give the goal a team. An architect decomposes the work into a dependency-aware task graph, independent subtasks run in parallel, builders hand off to reviewers who check against the task brief and bounce rejections back with feedback, and an integrator ships the final result. All of it moves on a live kanban board you can watch from the web GUI, dispatched with a single chat message and no config files to hand-edit.

Core capabilities center on parallelism, role-based model routing, and traceability. Every role can be pinned to any model in your DSH catalog - DeepSeek, GLM, Kimi, Claude - with an ordered fallback chain and a per-role reasoning-effort ladder. Concurrency caps adapt when the provider strains. Failure is a tracked state, not a mystery: timeouts, quota exhaustion, and missing evidence are all detected, retried with resume hints, or rotated to the next model automatically. State is a deterministic fold over an append-only JSONL event log, so a hostile or duplicated stream cannot resurrect an aborted run. The Swarm tab in the web host exposes task columns, reviewer feedback, retry trails, and an optional human approval gate that keeps the final verdict in your hands.

Built for users running multi-step, long-horizon tasks in DeepSeek Harness: batch code audits, multi-document research synthesis, parallel front-end component development - anything that decomposes into independent subtasks with cross-checks can be dispatched in one sentence and tracked to a delivery report on the board.