The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute internal scripts defined in its protocol manifest (skill-protocol.json). These executions use shell=False and target specific Python scripts within the skill directory, minimizing risks of command injection.
[DYNAMIC_EXECUTION]: The runtime uses importlib.import_module to dynamically load a diagnostic contract helper. This is used to maintain compatibility across different execution environments and does not ingest untrusted paths.
[DATA_EXPOSURE]: The skill implements strict path resolution and validation. The seal_audit.py script verifies that any file path provided in the audit bundle for evidence resides within the specified repository root, preventing directory traversal attacks.
[PROMPT_INJECTION]: The skill's documentation (references/audit-contract.md) explicitly requires the agent to treat repository content and issue prose as evidence rather than instructions, mitigating the risk of indirect prompt injection from the audited codebase.

audit-codebase