The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes manifest files and instructions from untrusted repositories, which could potentially contain malicious prompts designed to influence the agent's evaluation output.
Ingestion points: The skill reads plugin.json, mcp.json, and SKILL.md files from the target repository (provided via the target or compare parameters).
Boundary markers: The skill includes a 'Safety boundary' section in SKILL.md that explicitly restricts the agent to static auditing and forbids running any bundled executables or scripts from the untrusted source.
Capability inventory: The skill utilizes local Python scripts (inspect_plugin.py, score.py) for analysis and scoring, and involves cloning remote repositories using git clone tools.
Sanitization: The analysis script scripts/inspect_plugin.py implements path containment checks to prevent directory traversal and performs strict schema validation on manifest fields using regex.
[COMMAND_EXECUTION]: The skill executes its own internal Python scripts to perform the audit and calculate scores.
Evidence: The workflow involves running python3 scripts/inspect_plugin.py and scripts/score.py. These scripts perform deterministic static analysis and do not execute code from the target repository.
[EXTERNAL_DOWNLOADS]: The skill retrieves code from external Git repositories to facilitate its auditing functionality.
Evidence: The workflow instructs the agent to clone Git URLs into a temporary directory for inspection. This is a core functional requirement of the auditing tool.
Note: The skill also references agent-plugins.org for specification and schema retrieval, which is the official service for the specification being audited.

agent-plugin-eval