The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill utilizes local files to store and retrieve context, which introduces a surface for indirect instructions if those files are modified by external sources.
Ingestion points: Context is loaded from .agents/em-context.md and individual report files in the .agents/reports/ directory in SKILL.md.
Boundary markers: There are no explicit delimiters or specific instructions for the agent to ignore potentially malicious content within these files.
Capability inventory: The skill is capable of reading and writing local markdown files within the defined context directory; it has no network or shell execution capabilities.
Sanitization: No sanitization, filtering, or validation is performed on the content retrieved from the local context files.

feedback