The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content such as source code, READMEs, and design documents from audited repositories. This data is interpolated into agent prompts. To mitigate this risk, the skill includes a specific rule (Rule 6) requiring the agent to treat all audited content as data, ignore any embedded instructions (e.g., 'ignore previous instructions'), and report them as security findings.\n- [CREDENTIALS_UNSAFE]: The skill establishes strict data handling procedures (Rule 4) for discovered secrets. It prohibits the agent from copying or reproducing API keys, tokens, or credentials, instead requiring only references to their file path and line number, preventing accidental leakage into generated plans or version control.\n- [COMMAND_EXECUTION]: The skill employs common developer CLI tools (e.g., git, gh, npm audit) for read-only repository analysis. It includes built-in safety checks for external reporting, such as verifying repository visibility and obtaining user confirmation before creating GitHub issues for potentially sensitive findings.\n- [DYNAMIC_EXECUTION]: The skill implements a coordinated multi-agent workflow, delegating tasks to specialized sub-agents for auditing and execution. The executor agents are constrained to isolated git worktrees, which ensures that modifications and tests are performed in a sandboxed environment separate from the user's primary source code.\n- [EXTERNAL_DOWNLOADS]: The skill utilizes standard auditing tools like npm audit and pip-audit, which fetch vulnerability information from official package registries. These operations are restricted to standard diagnostic functions and target well-known, trusted services.

improve