dsh-subagent-tools

Out-of-the-box, the subagent delegation tools in DeepSeek Harness offer a minimal surface: you can describe a task, write a prompt, and choose background execution. This static setup becomes a bottleneck when you need to dynamically change the underlying LLM, switch the execution provider, or temporarily adjust the agent's persona for each specific sub-task.

dsh-subagent-tools is built to shatter this limitation. It provides an enhanced set of tools that replace the default ones, maintaining full backward compatibility while injecting powerful per-call overrides. The core innovation lies in four customizable parameters: you can override the child agent's model for a single call (using composite IDs like "kimi-code/k3" to switch provider and model simultaneously), change the delegation provider, inject new persona text or load a saved preset on-the-fly (via @preset: references), and fine-tune the tool allow/deny filter. All overrides are ephemeral—if not specified, the system gracefully falls back to the instance's default configuration, ensuring both flexibility and stability.

This plugin is ideal for users who require granular, context-aware control over AI sub-agents. Whether you are a researcher testing different models across project phases, a developer who needs to swiftly adapt an agent's persona for varying tasks, or an administrator managing multiple team-specific presets, it offers significant utility. It transforms workflow orchestration from a rigid "one-size-fits-all" model into a dynamic, customizable menu, greatly expanding the practical versatility of sub-agents without altering the core system.