The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from codebases and project artifacts (e.g., SQL schemas, OpenAPI specs). This creates an inherent surface for indirect prompt injection, where malicious instructions inside the analyzed files could attempt to manipulate the agent's responses.
Ingestion points: Untrusted content enters the context through the codebase_search and codebase_context_search tools.
Boundary markers: No specific delimiters or safety instructions (like 'ignore embedded commands') are mandated in the skill instructions for handling this data.
Capability inventory: The skill supports file reading, semantic search, and launching an interactive HTML-based visualization tool.
Sanitization: The instructions do not describe any sanitization or validation of the indexed codebase content.
[COMMAND_EXECUTION]: The codebase_graph_visualize tool, when used in interactive mode, generates an HTML file in the system's temporary directory and automatically invokes a system command to open it in the default web browser. This behavior is a documented feature for providing interactive visualizations but involves direct interaction with the host operating system's file system and browser controls.

codebase-exploration