The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands via the command node type in pipeline definitions and the command field in Markdown checklist items. This functionality allows for the execution of scripts and binaries in the user's environment to validate agent work.
[DYNAMIC_EXECUTION]: The skill dynamically parses and runs shell commands extracted from the YAML frontmatter of Markdown files and from pipeline configuration files. This behavior creates a reliance on the integrity of external files for secure operation.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting external data from user-editable files and interpolating it into agent prompts. Specifically, the $goal variable and checklist items are fed into agent turns without visible sanitization mechanisms, potentially allowing malicious instructions in those files to influence agent behavior. 1. Ingestion points: YAML pipeline files (goal strings) and Markdown checklist files. 2. Boundary markers: No specific delimiters or 'ignore instructions' guards are mentioned for the interpolated goal text or checklist data. 3. Capability inventory: The skill allows for arbitrary shell command execution and file system modifications across multiple agent turns. 4. Sanitization: The instructions do not document any sanitization or validation of the content being interpolated into the agent's context.

tractor