The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts claude_md_layering_audit.py and settings_linter.py utilize subprocess.run to call the git binary. This execution is limited to git ls-files for the purpose of checking whether sensitive configuration files (like .claude/settings.local.json) are incorrectly tracked by version control, which is a key security check for team environments.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the local filesystem, including CLAUDE.md and .mcp.json files. While these files are scanned for potential security risks, the scripts operate deterministically using regular expressions and do not feed raw untrusted data back into the LLM's own prompt logic during the analysis. The scripts identify ingestion points like @import statements and check for boundary markers, helping mitigate the risk of instructions being accidentally overridden by project-level data.
[SAFE]: All scripts included in the skill are implemented using the Python standard library and do not perform any network operations. The adoption_tracker.py script specifically emphasizes privacy by aggregating usage counts while excluding prompt text, file contents, and tool inputs, ensuring that the usage audit does not expose sensitive session data.

coding-agent-rollout