The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external inputs such as GitHub issues and pull request reviews. It includes a specific "Retrieved Evidence Boundary" protocol to ensure that natural language from these sources is treated as data, not as instructions that could override the skill's protocol or permissions.
[DATA_EXFILTRATION]: The skill implements robust guidelines for handling sensitive data, specifically instructing the agent to exclude passwords, tokens, and other credentials from logs, traces, and reports. It also utilizes Git-ignore checks to ensure that tool-generated data in the .tigerkit/ directory is not accidentally pushed to remote repositories.
[COMMAND_EXECUTION]: Local shell operations, including Git commands and test execution, are restricted to isolated worktrees. The skill enforces mandatory verification of workspace state and ignore rules before any product mutation occurs, preventing interference with unrelated user work.
[OBFUSCATION]: Client-side utility scripts in the assets/ directory use Base64 encoding to bundle evidence files within generated HTML reports. This is a functional requirement for a client-side ZIP generation feature and does not represent an attempt to hide malicious code.

tk-prep