The Agent Skills Directory

Filesystem Access and Secret Detection: The skill includes a Node.js utility, pack.mjs, which reads source code from the project directory to include in implementation plans. This utility implements a 'fence' to ensure it only accesses files within authorized directories (such as the repository root). Additionally, it contains a blocklist for sensitive file paths (like .ssh, .aws, and .env files) and a regular expression to detect and exclude content that appears to contain credentials or private keys.
Restricted Command Execution: The pack.mjs script executes the git command to retrieve specific file versions. The implementation strictly limits the subcommands to rev-parse and cat-file blob, while disabling git hooks and external monitors through specific configuration flags. This approach facilitates necessary repository inspection while minimizing the risk of executing unintended side effects.
Indirect Prompt Injection Guidance: The skill instructions address the handling of feedback responses by emphasizing that they should be treated as data rather than executable instructions. It explicitly cautions the agent against performing risky actions—such as running commands or fetching external URLs—based on comments or edits provided in the feedback, which serves as a mitigation against potential injection attempts via processed data.

html-plan