The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation script evals/run_behavior_suite.py uses subprocess.Popen to execute the codex platform CLI for behavior testing. The command arguments are passed as a list to avoid shell injection, and the execution is confined to the development testing infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill ingests complex JSON workflow specifications, representing a potential attack surface. Ingestion points: The scripts scripts/graph_tool.py and evals/run_behavior_suite.py parse JSON graph artifacts provided in the project environment. Boundary markers: Core instructions in SKILL.md and references/security-and-governance.md mandate treating all external data as untrusted and suggest implementing context trust labels. Capability inventory: The skill provides deterministic tools for graph validation and rendering (scripts/graph_tool.py) and a test runner for automated evaluations (evals/run_behavior_suite.py). Sanitization: All artifacts are validated against strict JSON Schemas, and the rendering logic in graph_tool.py uses html.escape to sanitize data before output.
[DYNAMIC_EXECUTION]: The test runner evals/run_behavior_suite.py dynamically loads the graph_tool.py module using importlib during the test lifecycle, which is a standard procedure for local testing.

graph-engineering