The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to manage local state files in ~/.gstack, perform git operations, and execute a suite of helper binaries located within the skill's installation directory at ~/.claude/skills/gstack/bin/. This use of shell commands is essential for its core functionality as an automated code reviewer.
[DATA_EXFILTRATION]: Contains a telemetry system that tracks skill usage, duration, and success outcomes. This functionality is explicitly gated behind a user-consent process using AskUserQuestion, where the user is informed that no sensitive data like code, repo names, or file paths are transmitted.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted content from git commit logs and pull request descriptions (git log, gh pr view) to synthesize the 'stated intent' of the code changes. This intent then guides the subsequent automated review and remediation suggestions.
[EXTERNAL_DOWNLOADS]: Implements an update-check mechanism through a local binary (gstack-update-check) and an 'inline upgrade flow' that can read and execute updated instruction sets from within the skill's own directory structure.
[REMOTE_CODE_EXECUTION]: Provides integration with the external Codex CLI for performing cross-model adversarial reviews. This involves executing system commands via the Codex tool if it is installed and authenticated by the user.

review