The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes subprocess.run across several modules to orchestrate its compliance testing workflow. It invokes the claude CLI to generate specifications, build scenarios, and classify behavioral traces. Additionally, it executes setup commands for test environments within a dedicated sandbox directory (/tmp/skill-comply-sandbox). These operations are essential for its function as a developer benchmarking tool and are constrained by path validation logic.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it reads external Markdown files (agent skills or rules) and interpolates their content into prompts for internal LLM processing.
Ingestion points: Markdown skill/rule files are read in scripts/spec_generator.py and scripts/scenario_generator.py, and JSONL trace files are parsed in scripts/parser.py.
Boundary markers: Prompt templates use triple-dash (---) delimiters to wrap external skill content.
Capability inventory: The skill maintains extensive subprocess capabilities, allowing it to run shell commands and the claude agent CLI across multiple scripts (scripts/run.py, scripts/runner.py, scripts/classifier.py, scripts/spec_generator.py, scripts/scenario_generator.py). It also performs file system writes for generating compliance reports.
Sanitization: The implementation includes a sandbox environment that uses path resolution checks (via Path.resolve().relative_to()) to ensure file operations stay within the intended directory, and it restricts the agent's capabilities by explicitly defining allowed tools during scenario runs.

skill-comply