The Agent Skills Directory

[COMMAND_EXECUTION]: The framework uses subprocess.run across multiple scripts to automate evaluation tasks.
scripts/evaluation_package_gate.py executes git commands (rev-parse, ls-files) to enumerate tracked files for hashing.
evals/contract-calibration-260730/contract_copilot_micro_probe.py executes sys.executable to run Python code snippets for report rendering and integrity checks.
[DYNAMIC_EXECUTION]: The evaluation probes dynamically load modules from the skill being audited.
evals/contract-calibration-260730/contract_copilot_micro_probe.py modifies sys.path and uses importlib.reload to load and test internal logic from the candidate skill directory (e.g., scripts.report.reporting, scripts.review.review_runtime).
[INDIRECT_PROMPT_INJECTION]: The skill processes external test materials and legal documents which could contain embedded instructions.
Ingestion points: Materials located in evals/case-analysis-20260808/materials/ and evals/litigation-analysis-20260808/materials/.
Boundary markers: The skill contains explicit instructions and semantic checks (e.g., CASE-NO-FABRICATE, CASE-NO-RECALL-ARTICLE) to prevent the agent from obeying instructions within materials or fabricating information when data is missing.
Capability inventory: File system access and dynamic execution via sub-processes in evaluation scripts.
Sanitization: Uses phrase matching and regex patterns in scripts/semantic_micro_case_gate.py to validate outputs against expected boundaries.

legal-skill-evaluation