The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local system commands including git and codex doctor to aggregate diagnostic data. These calls are implemented using subprocess.run in scripts/scan_workspace.py without shell=True, and the arguments are strictly controlled to prevent command injection from external workspace data.
[DATA_EXFILTRATION]: The scanner identifies and flags untracked files with sensitive names (like .env or auth.json) and configuration keys that resemble credentials. It is designed to alert the user to potential exposures for manual review without printing or transmitting the actual secret values, acting as a defensive audit tool.
[PROMPT_INJECTION]: The skill contains comprehensive instructions to prevent the agent from being manipulated into deleting safety boundaries or business rules. It explicitly defines 'protected semantic classes' in references/checks-and-repair-policy.md and requires individual human approval for any suggested file modifications.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data from the workspace (e.g., AGENTS.md and config.toml). It mitigates risk by applying strict semantic classification, requiring manual review for deletion candidates, and enforcing byte-limit checks to prevent context overflow attacks.

codex-doctor