dsh-plugin-task-runner

A locally hosted mid-size model often carries only 64K or fewer context tokens, so anything past a single document can trigger memory overflow or a sharp drop in output quality. dsh-plugin-task-runner adds a Task-Runner mode to DeepSeek Harness: the orchestrator breaks a large job into independent sub-tasks and hands each one to a sub-agent that works in its own fresh, short context window. The orchestrator never accumulates raw material; it only keeps the plan and each worker's compact summary, so no single window ever has to hold the entire task.

Budgets are enforced at execution level, not just in the prompt. A real semaphore caps concurrency; a pre-dispatch token estimate rejects sub-tasks that would exceed the worker context budget; any worker result that overflows the result-bandwidth cap is silently truncated and persisted to disk. Sub-agents run in fully isolated sessions, so secrets and history in the parent session never leak into a worker. A structured .task-runner/ directory externalizes long-running state, letting the orchestrator write heavy context to disk and keep only a tight summary in memory, with the ability to resume after an interruption.

This preset is built for developers running OpenAI-compatible local endpoints on 48 GB or 64 GB Macs who want a 27B-class model to handle multi-document research, large refactors, or any job that outgrows one context window. Turn on Task-Runner mode, set concurrency to 1 or 2, pick your worker and fallback models, and a single machine can now carry work that would otherwise demand a much larger context.