The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/build_evidence_inventory.py uses subprocess.check_output to interact with git for repository history analysis.
Evidence: The script calls git rev-parse, git rev-list, and git hash-object to identify historical file matches.
Security Note: These commands are executed using a list of arguments without shell=True, preventing shell injection vulnerabilities. The operations are scoped to the project root provided via the --site argument.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external websites and reference materials to find plagiarism risks.
Ingestion points: The skill reads site text, metadata, and assets via the build_evidence_inventory.py script (specifically in SKILL.md Step 2).
Boundary markers: The script outputs findings in a structured JSON format, and SKILL.md instructs the agent to treat these as "leads for human review, not an automatic verdict."
Capability inventory: The skill uses git for history inspection and standard file read operations for hashing and text shingling. It does not contain eval() or exec() calls targeting the ingested data.
Sanitization: Data is processed into hashes and normalized text shingles for comparison purposes; it is not executed or interpreted as instructions.

audit-reference-originality