The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/test_evidence_utilities.py uses subprocess.run to execute the skill's bundled Python utilities during regression testing. While this is used for internal validation and employs a safe argument list with the current Python interpreter, it represents a pattern of dynamic command execution that could be targeted if the environment is compromised.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves processing external data sources—such as repository files, execution artifacts, and user-defined scenarios—which constitutes an attack surface for indirect prompt injection.\n
Ingestion points: Data is ingested via file read operations in scripts/hash_artifacts.py and scripts/verify_review_bindings.py to facilitate evidence hashing and verification.\n
Boundary markers: The skill's core contract recommends the use of an 'Isolated review launcher' and mandates a 'fail-closed filesystem boundary' for higher-assurance validation tasks.\n
Capability inventory: The skill possesses repository-wide file read access and the ability to execute local Python scripts via its bundled test suite.\n
Sanitization: The skill instructions explicitly require the redaction of secrets, credentials, and private identities before any data reaches logs or the human/Codex reviewer, mitigating potential data exfiltration risks.

build-scenario-validation-loop