The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of local binaries within the ~/.claude/skills/gstack/bin/ directory to handle configuration (gstack-config), telemetry (gstack-telemetry-log), and session state management.
[REMOTE_CODE_EXECUTION]: Utilizes the codex exec tool to run independent plan critiques using an external model, providing an adversarial 'outside voice' to challenge the current design.
[DATA_EXFILTRATION]: Implements a telemetry and analytics system that logs skill usage, duration, and outcomes. While this involves data collection, it is implemented as an opt-in feature for the user and is restricted to usage metadata rather than source code or credentials.
[PROMPT_INJECTION]: Exhibits an indirect prompt injection surface by ingesting and processing untrusted project data (design documents and plan files) to perform architectural reviews.
Ingestion points: Reads project-specific design docs (*-design-*.md) and active plan files.
Boundary markers: Does not utilize explicit XML-style delimiters for the ingested content.
Capability inventory: Uses Bash, Write, Read, Grep, and WebSearch tools.
Sanitization: Content is processed for review rather than direct execution, but no formal input sanitization is applied.
[COMMAND_EXECUTION]: Automatically modifies the project's CLAUDE.md file to include 'skill routing rules' and performs git commits to automate subsequent skill invocations based on user intent.

plan-eng-review