The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to audit untrusted repositories, creating a surface where malicious instructions in a plugin's manifest or skill files could influence the agent's behavior during the evaluation process.
Ingestion points: The script scripts/inspect_plugin.py reads data from plugin.json, mcp.json, and SKILL.md files within the target repository.
Boundary markers: The SKILL.md file includes explicit instructions to perform audits statically and avoid executing bundled code, hooks, or scripts from the target repository without authorization.
Capability inventory: The agent has capabilities to clone Git repositories, browse the web (for specification updates), and write report files.
Sanitization: While the Python scripts validate JSON structure and schema compliance, there is no explicit sanitization or filtering of natural language content to prevent prompt injection.
[EXTERNAL_DOWNLOADS]: The skill retrieves normative specifications and schemas from the official domain agent-plugins.org and documentation from agentskills.io to ensure audit conformity.
[COMMAND_EXECUTION]: The skill invokes local helper scripts scripts/inspect_plugin.py and scripts/score.py to perform the technical audit and calculate weighted conformance scores.

agent-plugin-eval