The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the platform's hook system to execute local Python scripts via uv run. These scripts, such as ds-post-subagent-guard.py and ds-pre-subagent-clear.py, are located within the plugin root and serve as security controllers that monitor tool usage and maintain subagent boundaries.
[SAFE]: The skill enforces a strict protocol requiring all data analysis and engineering tasks to be performed by isolated subagents. This design prevents the main agent from executing code directly in the orchestrator context, providing a layer of execution isolation.
[SAFE]: The skill manages an indirect prompt injection surface by using structured markdown templates to separate instructions from untrusted plan data.
Ingestion points: Untrusted task descriptions from PLAN.md and project objectives from SPEC.md are interpolated into the subagent dispatch prompts.
Boundary markers: Templates use clear markdown headers (# TASK, ## CONTEXT) to delimit instructions from user-provided data.
Capability inventory: The skill dispatches agents with access to Read, Write, and Bash tools for data processing.
Sanitization: The protocol relies on the structural separation within the prompt and explicit negative constraints ('MUST NOT DO') provided to the subagents.

ds-delegate