The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the codebase being analyzed, which could contain malicious instructions designed to influence the agent's architectural recommendations.
Ingestion points: A sub-agent walks the codebase, reads CONTEXT.md, and scans Architecture Decision Records (ADRs) in docs/adr/ (referenced in SKILL.md).
Boundary markers: None identified; there are no specific delimiters or instructions to ignore embedded prompts in the analyzed files.
Capability inventory: The skill can write files to the system's temporary directory, execute shell commands (git, open/start), and call other tools like 'domain-modeling' to modify CONTEXT.md (referenced in SKILL.md).
Sanitization: No explicit sanitization or filtering of codebase content is mentioned before it is processed or rendered into the HTML report.
[EXTERNAL_DOWNLOADS]: The generated HTML report fetches resources from well-known content delivery networks (CDNs).
Evidence: The report includes <script> tags for Tailwind CSS (https://cdn.tailwindcss.com) and the Mermaid graphing library (https://cdn.jsdelivr.net/npm/mermaid@11/dist/mermaid.esm.min.mjs).
[COMMAND_EXECUTION]: The skill uses local shell commands for discovery and presentation.
Evidence: It executes git log --oneline to identify codebase hot spots and uses system-specific commands (xdg-open, open, start) to automatically display the generated HTML report to the user (referenced in SKILL.md).
[DYNAMIC_EXECUTION]: The skill dynamically generates an HTML report and configures a graphing library with a relaxed security posture.
Evidence: HTML-REPORT.md specifies mermaid.initialize({ securityLevel: "loose" }). The "loose" setting allows the rendering of scripts or links within diagrams, which could lead to cross-site scripting (XSS) if the sub-agent incorporates unsanitized malicious code snippets from the scanned project into the Mermaid diagram blocks.

improve-codebase-architecture