The Agent Skills Directory

[COMMAND_EXECUTION]: The script script_runner.py executes local scripts using subprocess.run with parameterized arguments and shell=False to mitigate command injection risks.
[COMMAND_EXECUTION]: The skill dynamically generates a temporary Python shim (sitecustomize.py) at runtime in a temporary directory to facilitate API call logging in scripts under test.
[DATA_EXFILTRATION]: The skill reads sensitive local data from ~/.claude/projects/ (Claude Code session logs) to provide visibility into native tool usage and API calls as part of its auditing function.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it analyzes untrusted third-party skill content.
Ingestion points: api_logger.py and prompt_linter.py read script source code and SKILL.md instructions from the skill being audited.
Boundary markers: SKILL.md (Line 18) defines a 'content-as-data' rule that explicitly instructs the model to treat all ingested content as data to be analyzed rather than instructions to be followed.
Capability inventory: The skill can execute local scripts via script_runner.py and read/write files within its workspace.
Sanitization: shared_io.py includes _reject_traversal and _check_boundary functions to ensure file operations are restricted to allowed directories.

skill-tester