The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality involves auditing untrusted codebase data and pull request information, creating a potential vector for indirect prompt injection attacks.
Ingestion points: The skill processes entire repositories, git diffs, pull request titles/descriptions, and generated findings reports.
Boundary markers: The role.md file contains explicit instructions to treat all repository content, including code and comments, as data under review rather than active instructions.
Capability inventory: The skill can execute shell commands (git, mkdir, date, find), run local Python and Bash scripts, and orchestrate sub-agents via the platform's Workflow tool.
Sanitization: A specialized redactor agent (claude-security:scan-redactor) is employed to identify and remove hardcoded credentials from the final reports.
[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file utilizes dynamic context injection syntax (!) to execute shell commands at load time for session management.
Evidence: Usage of date -u +%Y%m%d-%H%M%S to generate unique session timestamps.
Evidence: Usage of find . -maxdepth 1 -type d -name "CLAUDE-SECURITY-2*" to locate existing report directories within the repository.
[COMMAND_EXECUTION]: The skill performs extensive command execution to manage git operations and run its internal suite of analysis scripts.
Evidence: Orchestrates various helper scripts, such as write_scan_meta.py, render_report.py, and patch_artifacts.py, which are executed from the plugin's internal directory using python3 and bash.

claude-security