The Agent Skills Directory

[SAFE]: No malicious instructions, obfuscation, or unauthorized network operations were detected. The agent configuration in agents/openai.yaml explicitly disables implicit invocation, which is a defensive measure to ensure the skill is only used when intended.
[INDIRECT_PROMPT_INJECTION]: The skill processes outputs from external sub-agents during its aggregation phase, which represents a potential injection surface.
Ingestion points: Phase C (Aggregate) involves reading terminal results from parallel worker sub-agents.
Boundary markers: The instructions explicitly command the agent to "not paste raw worker dumps," providing a clear instruction to avoid direct ingestion of untrusted content.
Capability inventory: The skill manages the spawning and results-processing of multiple native workers to handle code slices or race arms.
Sanitization: The instructions require the agent to generate a "compact result table" and "one-line evidenced issues," effectively acting as a summarization and filtering layer that prevents raw payloads from being interpreted as instructions by the parent agent.
[EXTERNAL_DOWNLOADS]: The MIRROR.md file provides provenance links to the official cursor/plugins GitHub repository. This documentation allows for source verification and does not involve the runtime execution of remote code.

cursor-swarm