dsh-task-orchestrator

When a single request spans research, architecture, frontend and backend implementation, testing, and documentation, a lone agent path quickly hits context and parallelism ceilings, and manual subtask splitting lacks dependency tracking and safe context reclamation. dsh-task-orchestrator keeps the parent model as the sole orchestrator: a deterministic detector scores request complexity, a strict JSON dependency graph is produced, and the minimum useful set of workers is packed by context budget so every step returns structured evidence without context drift.

The scheduling layer is where the plugin earns its keep. Independent read-only roles are admitted incrementally by context tier; as soon as one role settles, the next fitting role starts without waiting for its siblings. Cross-role dependency reports are compacted to bounded facts before handoff, and only a NEED_MORE_CONTEXT flag triggers a single bounded context escalation. Writes are disabled by default and plans are reviewed by a human before execution. Up to six workers run alongside the parent, with oversized or unbounded requests rejected or truncated at configured limits. The entire tool group is package-owned, so a single Market toggle atomically enables or disables all five model-facing tools without touching the web profile standard rows.

This plugin suits teams running multi-component development pipelines on DeepSeek Harness, particularly those deploying on constrained GPUs such as an RTX 3090 where parallel generation must be tightly bounded. Whether the workload is read-only research or full-stack write tasks, conservative defaults deliver controlled parallel speedups without unsafe memory reservations, leaving filesystem permissions and approval policy to the surrounding DSH profile as intended.