The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository content, creating a vulnerability surface where malicious code in audited files could attempt to influence the agent's behavior.
Ingestion points: The tool reads repository file lists and contents via git ls-files and the references/detect.sh script.
Boundary markers: Workflow rules explicitly require citing specific evidence and disclosing any skipped or failed checks.
Capability inventory: The orchestrator can execute local scripts, dispatch sub-workers, and write files to the project directory.
Sanitization: Analysis is performed by separate, specialized lane agents, which isolates the primary processing from the data ingestion.
[COMMAND_EXECUTION]: The skill invokes a local shell script (references/detect.sh) to perform technology discovery. This script relies on standard system utilities like git and grep and does not engage in network activity or access sensitive files outside the audited scope.

repo-audit