The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill's primary function is to execute shell commands (gemini, node) with arguments derived from user prompts and AI responses. This includes a 'sandbox' mode (-s) intended for code execution and a 'change mode' for structured file edits, both of which are high-privilege operations.
[REMOTE_CODE_EXECUTION] (HIGH): High susceptibility to Indirect Prompt Injection (Category 8).
Ingestion points: Reads files and directories via @path and @. syntax (referenced in SKILL.md).
Boundary markers: None identified in the skill instructions to distinguish between trusted instructions and untrusted data from the processed files.
Capability inventory: Can execute shell commands, run sandbox scripts, and perform 'deterministic' file modifications (OLD/NEW blocks).
Sanitization: No evidence of sanitization or validation of the content retrieved from local files before it is processed by the AI or acted upon by the agent.
[DATA_EXFILTRATION] (MEDIUM): The skill is designed to send local project data to an external API (Google Gemini). While this is the intended use case, it presents a data exposure risk for sensitive configuration files or proprietary source code if the user or agent inadvertently includes them in the context window.

gemini-cli