The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill consumes external files to define behavior and track state, which creates a vulnerability if those files contain malicious instructions.\n
Ingestion points: Loads workflow.yaml, .workflow-state.json, and Backlog.md from the local environment.\n
Boundary markers: No explicit boundaries or warnings are provided to separate data from instructions within these files.\n
Capability inventory: The skill manages agent routing and gate enforcement; it does not perform direct command execution or network exfiltration but governs the context for agents that do.\n
Sanitization: The skill lacks validation logic to ensure that natural language instructions are not injected into structured data fields.

workflow-engine