The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extensively executes local binaries and scripts (e.g., gstack-skill-start, gstack-context-recovery, gstack-slug) located in the ~/.claude/skills/gstack/bin/ directory. These tools are used for session management, context recovery, and logging analytics.
[DATA_EXFILTRATION]: The skill transmits technical plans and design documents to external review services, specifically aside.com and the OpenAI Codex CLI. It includes a mandatory sanitization step to strip hostnames, IP addresses, file paths, and secrets from queries before they are sent to the external Aside agent.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill processes external data including technical plans, design docs, and branch diffs as described in SKILL.md and sections/review-sections.md.
Boundary markers: Instructions use specific headers like "THE PLAN:" and include explicit warnings that handoff notes are data, not instructions, and should not be followed if they attempt to bypass review steps.
Capability inventory: The skill utilizes Bash, Read, Write, Grep, and WebSearch tools.
Sanitization: Employs query sanitization for web research and directs the agent to report suspicious content found in design documents.
[DYNAMIC_EXECUTION]: The skill dynamically sources local shell scripts (gstack-codex-probe) and executes complex shell command strings, including a perl-based wrapper used to implement timeouts for external CLI calls.
[EXTERNAL_DOWNLOADS]: The instructions direct the user to install the well-known @openai/codex CLI package via npm to enable independent secondary reviews.

plan-eng-review