The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the vendor-provided @lev/index library and interacts with a local gRPC server on port 50052. Tools like index:build and index:add perform file system operations to read directories and files for the purpose of creating vector indexes.
[PROMPT_INJECTION]: The skill implements a 'Prompt Architect Overlay' which contains role-definition instructions for the agent. While intended for functional alignment, these instructions use assertive language to guide agent behavior during the execution of the skill.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) due to its core functionality of indexing and searching external data.
Ingestion points: Data enters the system from user-specified directories and files via the index:build and index:add tools.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore embedded instructions' warnings for content retrieved during semantic searches.
Capability inventory: The index:search tool retrieves potentially untrusted content from the index and presents it to the agent, creating a vector for hidden instructions to influence agent behavior.
Sanitization: There is no evidence of sanitization, filtering, or validation of the ingested text or file content before it is processed or displayed.

lev-index