The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted data from agent interaction traces and behavioral documentation. Maliciously crafted content within these files could attempt to subvert the agent's auditing logic.
Ingestion points: Processes ETHOS.md behavior definitions and ATIF format trajectory.json interaction logs.
Boundary markers: Employs Markdown comment markers and enforces strict JSON schemas for all structured input and output.
Capability inventory: Utilizes Bash, Read, and Write tools to execute measurement scripts and manage resulting report artifacts.
Sanitization: Uses jsonschema for structural validation and PyYAML.safe_load for secure parsing of configuration files.
[EXTERNAL_DOWNLOADS]: The skill requires standard Python packages harbor, jsonschema, and PyYAML. These are well-known libraries fetched from official package registries and are documented as safe dependencies.
[COMMAND_EXECUTION]: The skill executes local Python scripts under its own directory using uv run to perform generation, validation, and measurement tasks. This is restricted to the skill's own code and is necessary for its intended auditing functionality.

eval-author-audit