The Agent Skills Directory

Command Execution: The skill executes several local shell commands via the Bash tool, including git, python3, find, and grep. These are used for repository analysis, report rendering, and patch management. Access is restricted to specific command patterns and directories, and the Security Lead role is instructed to avoid chaining commands to maintain predictable execution.
External Downloads: The skill includes a capability to search and list GitHub pull requests using the gh CLI. This is a gated feature that requires explicit user authentication and is limited to fetching metadata and diffs from the code host. The analysis of these external resources is treated strictly as data, not as executable instructions.
Dynamic Context Injection: The skill uses ! commands in its documentation to provide real-time environment data, such as current UTC timestamps and existing report listings. These commands are localized to the user's environment and do not accept untrusted external input.
**Security Control
Credential Suppression:** All Git operations are executed with GIT_TERMINAL_PROMPT=0 and global configurations are ignored (GIT_CONFIG_GLOBAL=/dev/null) to prevent the session from hanging on credential prompts or using unauthorized local settings.
**Security Control
User Confirmation:** The skill implements a mandatory AskUserQuestion confirmation step for all scans to ensure users are aware of potential token costs and session duration before a scan begins.
**Security Control
Isolation:** Patch generation and verification occur in isolated scratch workspaces (git clone --shared) that are removed after completion, ensuring the user's main working tree remains untouched.

claude-security