dsh-anchored-subagent

DeepSeek was RL-trained under a specific Minimal condition: only bash and str_replace_editor visible, paired with a one-line persona. A typical DSH session, however, ships with the full tool catalog and a verbose system prompt, which can pull the model away from the "We need..." reasoning trajectory it was most strongly optimized for. dsh-anchored-subagent solves this with a "secret handshake": both the main agent and every inherited subagent start on that exact Minimal bootstrap, then seamlessly unlock the complete toolbox after the first meaningful turn.

The plugin installs as an agent preset that DSH composes into the parent session and each child subagent automatically. On request #1 the model sees only bash and str_replace_editor with a terse Minimal persona; once the first durable tool-call or assistant message lands, the full tool catalog and any pre-configured subagent persona are restored. The promotion trigger (tool-call, assistant-message, or either) and other details such as suppressed context sources are all tunable in agent.cordis.yml, so no manual prompt engineering is required.

Ideal for DSH users who frequently spawn subagents, delegate multi-step tasks, or simply want their main agent to open every conversation in DeepSeek's optimal reasoning regime. Select the preset and the bootstrap is in effect — zero configuration, immediate benefit for multi-agent workflows.