The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands to perform its primary functions. In verify mode, it runs project-specific launch scripts (e.g., pnpm dev, npm start, make) and uses tools like curl to exercise the application. In review mode, it executes git commands (git diff, git rev-parse, git merge-base) to scope code changes.
[PROMPT_INJECTION]: The review subagent is vulnerable to indirect prompt injection as it processes untrusted code changes. Ingestion points: Reads git diff output and changed file contents into the subagent context. Boundary markers: Uses Markdown sections to separate rubrics from code, but lacks explicit instructions to ignore commands within the code. Capability inventory: The subagent is granted Read, Bash, Grep, Glob, and Write tools. Sanitization: No evidence of input sanitization or escaping was found.
[DATA_EXFILTRATION]: The skill accesses configuration files to detect the active model family. It reads .claude/settings.json and .claude/settings.local.json from both the current project directory and the user's home directory. While used for functional routing, this involves reading files outside the target source code.

check