dsh-preset-agent-manager

In DSH (DeepSeek Harness) multi-agent workflows, delegating a task to a sub-agent that may only use a specific set of tools has always been a practical need. Until now, such sub-agents were scattered across prompt conventions and manual assembly, with no unified management UI, no runtime-enforced tool whitelist, and no visual audit trail. dsh-preset-agent-manager elevates the preset agent to a first-class citizen in DSH: each agent carries its own id, name, trigger-domain description, persona, and a rigorously enforced plugin and tool whitelist. The primary agent invokes preset_agent_dispatch mid-conversation to delegate autonomously; the child session runs in full isolation and is destroyed upon completion.

The security model operates on two layers. The inheritance mask (toolFilter.allow) strips every inherited tool outside the whitelist from the child-scope prompt so it simply does not exist. The self-scope guard then plugs the residual gap where delegation tools (subagent, subagent_fork) are registered in the child-scope itself and would otherwise bypass the mask. Together they guarantee a hard deny for anything outside the whitelist, and nested delegation is blocked by maxDepth: 1. If a whitelisted plugin fails to initialize, the child session degrades gracefully to a partial status with warnings rather than breaking the parent conversation. The management panel opens from the sidebar footer, adapts to the viewport, and hosts create and edit forms directly in the panel body. The plugin list shows per-row status, fiber phase, and the reason each entry is unavailable. Data persists via atomic write with automatic corruption quarantine, and hot-reload is supported without a profile restart.

Ideal for developers and admins building multi-agent pipelines in DSH who need strict tool boundaries for security-sensitive tasks, who want the primary agent to route work autonomously based on trigger-domain descriptions, or who require pure-reasoning sub-agents with a zero-tool whitelist.