dsh-custom-subagents

In daily DeepSeek Harness workflows, delegating tasks like code review, documentation writing, or web research happens constantly, yet each round is improvised through free-form prompts — no fixed role definition, no explicit tool boundary, and no guard against sub-agent chains spiraling out of control. dsh-custom-subagents turns those recurring specialist roles into a single saved definition. You fill in a name, system prompt, model, and tool set through a short form in Settings > Sub Agents, and from then on the main agent can summon that specialist on demand via delegate_agent. No coding required, and no need to understand DSH internals.

Three core capabilities stand out. First, clear tool boundaries: check off read, write/edit, Shell, web_search, or skill — what you check is exactly what the sub-agent receives, and it can never exceed the existing tool set of the parent agent. Second, nesting protection: native DSH sub-agents have no default depth cap, so a single task can spiral into layers of nested agents and burn through context and time. This plugin adds two layers of defense — a global Disable Nested Delegation toggle and a hard 3-level depth ceiling that always applies, rejecting any delegation chain beyond level three outright. Third, safety and isolation: editing or deleting a definition only affects future delegations, running instances are untouched, and the DSH sandboxing, approval, and security mechanisms remain unaffected by custom prompts.

It suits two audiences well: teams or individuals who delegate to the same specialist roles (code reviewer, doc writer, research agent) every day and want to define each one once and reuse it repeatedly; and security-conscious power users who want a least-privilege tool whitelist plus a hard depth fence without touching source code. Everything is configured through the DSH settings UI — up and running in minutes.