The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits the complete codebase, which involves ingesting untrusted data from the analyzed project files. Malicious instructions embedded in the audited code (e.g., in comments, documentation, or string literals) could potentially influence the audit results or the 'Prompt para Correção Automatizada' generated at the end of the report.
Ingestion points: All project files identified during the inventory stage, including source code, configuration files, and templates.
Boundary markers: The skill does not explicitly define delimiting markers or 'ignore embedded instructions' warnings for the sub-agents that process the file contents.
Capability inventory: The skill has the capability to execute local shell scripts and write a detailed Markdown report to the project root.
Sanitization: Source code content is analyzed semantically by LLM sub-agents without specific sanitization against prompt injection patterns.
[COMMAND_EXECUTION]: The skill relies on local shell scripts to perform its deterministic scanning phase, invoking tools such as grep, ripgrep, ast-grep, and semgrep.
Evidence: The 'Etapa 0' execution flow in SKILL.md runs bash .agents/skills/sonarqube-audit/scanners/scan.sh <project_root> /tmp/opencode.
[EXTERNAL_DOWNLOADS]: The scanner attempts to utilize semgrep via Docker, which involves pulling the semgrep/semgrep image from a well-known service (Docker Hub).
Evidence: The scanners/scan.sh script contains logic to execute docker run --rm semgrep/semgrep if the tool is available.

sonarqube-audit