The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands via git and gh. It employs a high level of shell safety by enforcing single-quote rendering for dynamic values, using -- as an option terminator, and strictly prohibiting the use of heredocs. These measures effectively prevent command injection from attacker-controlled branch names or commit messages.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as branch names and commit logs. It mitigates indirect prompt injection risks by implementing a robust sanitization script (sanitize-pr-content.sh) that strips forbidden AI-attribution patterns and by requiring explicit user confirmation through AskUserQuestion before performing any mutating actions.
[PRIVILEGE_ESCALATION]: The sanitize-pr-content.sh script implements a sophisticated security control by establishing bash -p (privileged mode) via a re-execution logic. This protects the enforcement process from environment and shell function poisoning. Additionally, it uses umask 077 to ensure that temporary PR body files are not readable by other users on the system.
[DYNAMIC_EXECUTION]: The skill utilizes a canonical cleanup block that safely captures subshell exit codes. It includes a specific validation step using case patterns to ensure that status variables are numeric before they are evaluated in arithmetic contexts, preventing potential expression-based injections within the shell logic.

create-pr