The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses the agent's global configuration directory (~/.claude/) to read guidelines and check for orphaned files. While this involves sensitive system paths, the access is restricted to the skill's primary function of maintaining the agent's environment, and it explicitly requests user permission before performing deletions in this directory.
[INDIRECT_PROMPT_INJECTION]: The skill processes audit-report.md as a data source for determining code refactors. This presents a potential surface for indirect prompt injection if the report contains malicious instructions. However, the risk is mitigated by a mandatory manual review step (Step 3) where the agent must present a plan and wait for explicit user approval (Confirm to proceed?) before any files are modified. Additionally, the skill performs a self-check (Step 5) to verify the integrity of changes after they are applied.
[COMMAND_EXECUTION]: The skill includes logic to automatically invoke its companion skill, config-audit, if its required input file is missing or outdated. This is a standard functional dependency and does not involve arbitrary command execution. The skill also provides the user with instructions to run a local install.sh script, but it specifically prohibits the agent from executing the script automatically to ensure human oversight.

config-refactor