The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze user-provided source code, test suites, and repository configurations, which creates a potential surface for indirect prompt injection if the analyzed content contains adversarial instructions.
Ingestion points: Processes project source files, test definitions, and environmental configuration as part of its audit workflow defined in SKILL.md.
Boundary markers: The SKILL.md file contains explicit 'Non-negotiable evidence rules' and safety boundaries, requiring separate authority for costly experiments and the use of recoverable baselines.
Capability inventory: The skill utilizes the target project's native test tooling and performs internal verification using subprocess.run in scripts/test_skill.py.
Sanitization: The instructions emphasize the use of synthetic or approved data and isolated copies of fixtures to reduce the impact of potentially malicious inputs.
[COMMAND_EXECUTION]: The preflight script scripts/test_skill.py uses the subprocess module to execute unit tests.
Evidence: The script invokes sys.executable -m unittest to verify the integrity of the skill's own fixtures located in evals/fixtures/parser-weak-oracle/. This execution is limited to internal verification logic.
[DYNAMIC_EXECUTION]: The skill performs runtime execution of local Python code as part of its internal validation and preflight checks.
Evidence: Both scripts/test_skill.py and scripts/test_validator_regressions.py are executed to ensure the reliability of the skill's behavioral evidence and validator logic. These operations use static paths relative to the skill's installation directory.

adversarial-test-sweep