The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local git commands to retrieve file history and recency metadata.
Evidence: The scripts/collect_evidence.py file uses subprocess.run(["git", "log", ...]) to collect the last_git_change for files.
Safety: The execution uses a list of arguments rather than a shell string, preventing shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it ingests and processes untrusted content from the target repository.
Ingestion points: The scripts/collect_evidence.py script reads the content of all text-based files in the repository harness (e.g., .md, .json, .yaml) to build a corpus for analysis.
Boundary markers: SKILL.md implements a 'Safety Gate' (Step 0) that explicitly instructs the agent to confirm the action is read-only and refuse any commands that write, delete, or modify the project state.
Capability inventory: The skill can read local files, execute git inspection tools, and write a single Markdown report file (AUDITORIA_SDD_<date>.md).
Sanitization: The helper script processes raw data into structured JSON before the agent analyzes it, reducing the risk of direct instruction execution from file content.

sdd-auditoria-harness