The Agent Skills Directory

[COMMAND_EXECUTION]: The 'Golden Rule' explicitly instructs the agent to implement validation logic using bash -c shell commands, promoting the creation and execution of flows that contain arbitrary command logic.
[COMMAND_EXECUTION]: The skill uses the lev exec tool to run audit and promotion workflows defined in local YAML files, which may contain further execution primitives.
[DATA_EXFILTRATION]: The instructions direct the agent to read local files from the user's home directory, specifically ~/lev/docs/ and ~/.lev/flows/, exposing internal design and configuration data.
[PROMPT_INJECTION]: The skill audits and fixes external .flow.yaml files, creating an indirect injection surface. Evidence: (1) Ingestion points: The agent reads files from a user-provided scan_path during the 'Audit' and 'Fix' intents. (2) Boundary markers: No delimiters or explicit instructions to ignore embedded content are provided. (3) Capability inventory: The skill utilizes lev exec, cat, and touch for file and execution operations. (4) Sanitization: No sanitization or validation of the ingested flow content is specified before the agent processes or 'fixes' it.

flowmind-author