The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs extensive local operations including git repository management (commit, add, rm), remote URL inspection, and interaction with the GitHub (gh) and GitLab (glab) CLI tools to retrieve PR/MR context. It also executes specialized local utility scripts from its own binary directory to manage project configuration and slugs.\n- [REMOTE_CODE_EXECUTION]: It leverages codex exec to send plan content to a remote model for analysis. This is a core architectural feature used to achieve dual-voice consensus in the review process.\n- [DYNAMIC_EXECUTION]: Uses eval to process the output of a local utility script (gstack-slug). This pattern is employed to dynamically set environment variables (such as project slugs and branch names) required for subsequent pipeline steps.\n- [PROMPT_INJECTION]: The skill implements an autonomous decision-making framework that instructs the agent to bypass standard user interaction for routine 'Mechanical' or 'Taste' decisions. However, it explicitly preserves manual gates for high-level 'Premises' and 'User Challenges' where model recommendations conflict with user intent.\n- [DATA_EXPOSURE]: The skill reads local environment data, such as git remotes and CLI authentication status (gh auth status), to determine the project context and hosting platform. This information is used for internal routing and is not exfiltrated to untrusted third parties.\n- [INDIRECT_PROMPT_INJECTION]: As the skill processes user-supplied plan files, it implements a security boundary by explicitly instructing sub-tools not to read or execute SKILL.md files or skill definition directories. \n
Ingestion points: Processes plan markdown files and design documents from the repository and local .gstack storage.\n
Boundary markers: Includes a mandatory 'IMPORTANT' prefix for Codex prompts to prevent tool-calling loops or logic subversion from skill files on disk.\n
Capability inventory: Full access to shell (Bash), file system (Read/Write/Edit/Glob/Grep), and remote model execution (Codex/Agent).\n
Sanitization: Relies on prompt-based boundary markers to delineate instructions from untrusted plan data.

autoplan