The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute platform-specific system commands such as xdg-open, open, or start to automatically display the generated report in the user's browser.
[EXTERNAL_DOWNLOADS]: The generated report template includes references to external script and style assets hosted on well-known CDNs (Tailwind CSS and JSDelivr for Mermaid.js). While these are reputable services, they represent external dependencies in the rendered report.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted data from the user's codebase (source files, ADRs, and glossary files).
Ingestion points: Codebase exploration via the Explore agent and reading project-specific documentation files.
Boundary markers: No explicit delimiters or instructions are provided to the agent to differentiate untrusted codebase content from its own instructions.
Capability inventory: Writing files to the temporary directory, shell command execution, and utilizing the Agent tool for recursive exploration.
Sanitization: No explicit sanitization or escaping logic is defined for codebase content before it is included in report diagrams. Furthermore, the report template explicitly initializes the Mermaid diagramming library with securityLevel: "loose", which disables certain security protections and creates a surface for cross-site scripting (XSS) if malicious data from the codebase is rendered within a diagram.
[DATA_EXFILTRATION]: Because the generated architectural report is viewed in a browser and loads external CDN resources, any sensitive codebase information included in the report could theoretically be exposed to those external providers.

improve-codebase-architecture