The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and display execution 'traces', which contain untrusted data from external AI interactions. This constitutes a potential attack surface for indirect prompt injection and cross-site scripting (XSS).
Ingestion points: The review application reads trace data from traces.jsonl files.
Boundary markers: The UI implementation in app.js correctly uses textContent instead of innerHTML when rendering trace content, ensuring that data is treated as literal text and not executable code.
Capability inventory: The bundled servers (server.py, server.mjs) can write annotations and taxonomy revisions to the local filesystem as JSON files.
Sanitization: The local servers implement a robust defense-in-depth strategy, including a strict Content Security Policy (CSP) that blocks external scripts/styles and a per-session token requirement (X-Session-Token) for all write operations to prevent unauthorized access.
[DYNAMIC_EXECUTION]: The setup_workspace.py script performs automated template installation by copying provided scripts and HTML/JS resources to a project directory. During this process, it performs simple string replacements on the template code (such as updating default file paths) to match the user's workspace configuration. This is a standard installation procedure and does not involve the execution of arbitrary external code.
[EXTERNAL_DOWNLOADS]: The skill's documentation and managed workflow guides reference the author's official website (confident-ai.com) and GitHub organization (confident-ai). These references are used for integration with the author's managed services and official CI/CD tools, representing legitimate vendor functionality within the skill's intended scope.

eval-discover