The Agent Skills Directory

[SAFE]: The skill acts as a delegation layer to reputable AI CLI tools from established vendors. It includes significant safety features, such as a dedicated secret scanner in the gpt-pro-bundle.sh script that prevents the packaging of API keys, tokens, or private keys before they are sent to external services.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied tasks and local project data (such as git diffs and source files) and passes them to remote peer agents. Ingestion points: claude-review-stream.py ingests git diffs; gpt-pro-bundle.sh ingests project source files. Boundary markers: The skill uses explicit system prompts to define the peer agent's role (e.g., 'strict code reviewer') and instructs them not to modify files during review phases. Capability inventory: Peer agents are granted specific capabilities via CLI flags; for example, the Claude review wrapper limits tools to Read, Grep, and Glob, while the implementation wrapper allows Edit, Write, and Bash. Sanitization: The gpt-pro-bundle.sh script implements a fail-fast secret scanner using regex for common credential patterns. The Claude review wrapper enforces a maximum size limit on git diffs.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the installation of well-known developer tools and libraries from established sources, such as Microsoft's Playwright and the notebooklm-py library from PyPI. These are standard dependencies for the target capabilities.
[COMMAND_EXECUTION]: The skill utilizes subprocesses to invoke local peer agent binaries. It uses non-interactive flags (e.g., --dangerously-bypass-approvals-and-sandbox for Codex, --dangerously-skip-permissions for Antigravity, and --trust-all-tools for Kiro) to facilitate agentic delegation. These configurations are consistent with the skill's primary purpose of headless task execution.

call-agent