The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Cairo source code to generate security reports. This creates an attack surface for indirect prompt injection where a malicious contract could include instructions designed to mislead the AI auditor.\n
Ingestion points: Discovers and reads all .cairo files within the provided repository scope via the orchestrator's discovery phase.\n
Boundary markers: Source code is wrapped in file path headers and markdown blocks before being bundled for specialist agents.\n
Capability inventory: Specialist agents are restricted to returning JSON, but the orchestrator agent has broad tool access including Bash and Read.\n
Sanitization: Specialist outputs are validated against a JSON schema and normalized before report generation.\n- [DYNAMIC_EXECUTION]: The skill's preflight bridge script uses dynamic Python loading to execute an optional benchmark detector script from a path computed at runtime.\n
Evidence: Uses importlib.util.spec_from_file_location and spec.loader.exec_module in scripts/quality/detector_bridge.py to load benchmark_cairo_auditor.py.\n
Context: This functionality is restricted to instances where the skill is run within its full developer repository structure and is used to extend the deterministic scanner's capabilities.\n- [EXTERNAL_DOWNLOADS]: The skill performs network operations to check for updates and fetch security threat intelligence from trusted sources.\n
Evidence: Uses curl to access the author's official GitHub repository for version checks and fetches security reports from well-known domains such as starknet.io, code4rena.com, and openzeppelin.com.\n
Context: These operations are restricted to reading public data and do not involve direct execution of the downloaded content.\n- [COMMAND_EXECUTION]: The skill relies on shell commands for repository discovery, file bundling, and running its internal analysis and testing scripts.\n
Evidence: SKILL.md and various utility scripts use subprocess.run() or direct Bash execution for pipeline orchestration.\n
Context: These operations are fundamental to the skill's primary function and appear to be implemented with appropriate safeguards for path resolution and temporary directory management.

cairo-auditor