The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses system commands (xdg-open, open, start) to automatically open the generated HTML report in the user's browser. This executes a shell command with a constructed file path, which is a potential risk if the path generation is manipulated.
[EXTERNAL_DOWNLOADS]: The generated HTML report fetches external scripts and styles from well-known CDNs (cdn.tailwindcss.com and cdn.jsdelivr.net). While these sources are generally trusted, their inclusion in an architectural review file creates a dependency on external servers and introduces potential tracking or availability risks for internal development work.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection.
Ingestion points: It scans local codebase files, git log, and ADR documents (SKILL.md).
Boundary markers: None. There are no instructions to the agent to distinguish between its commands and potential instructions embedded in the codebase files it reads.
Capability inventory: The agent can read files, write files to the OS temp directory, and execute the system browser (SKILL.md).
Sanitization: None. The skill does not instruct the agent to sanitize or escape data extracted from the codebase before rendering it into the HTML report, allowing malicious content in the codebase to manipulate the report's output.
[REMOTE_CODE_EXECUTION]: The skill generates an HTML report that initializes Mermaid diagrams with securityLevel: "loose" (HTML-REPORT.md). This setting specifically permits HTML tags within diagrams, creating a Cross-Site Scripting (XSS) vector if the diagram content—sourced from the codebase—contains malicious scripts or specific Mermaid exploits.
[DATA_EXFILTRATION]: The architectural review extracts domain models and internal structures into a file that triggers external network requests (CDNs). Opening the report could potentially leak information via HTTP Referer headers or metadata to the CDN providers.

improve-codebase-architecture