The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads agent-browser binaries from Vercel Labs' official GitHub repository. This is handled by platform-specific scripts in references/browsers/agent-browser.md. The process is secured by verifying each download against hardcoded SHA-256 checksums before execution, ensuring the integrity of the external dependency.
[DYNAMIC_EXECUTION]: The skill implements a descriptor-based architecture where operations for trackers and browsers are defined in markdown files within the repository's .ai/ directory. These files contain shell commands that are dynamically loaded and executed by the agent. This mechanism is the intended design for extensibility, allowing repository owners to customize operations without modifying the core skill code.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with external data from issue trackers (e.g., issue bodies, PR comments) and repository files. To mitigate risk, references/agentic-setup.md establishes a strict 'Untrusted content boundary,' instructing the agent to treat such content as data rather than instructions, to report suspected injections, and to validate/quote all externally-sourced values before use in shell commands.
[PRIVILEGE_ESCALATION]: During the setup of agent-browser on Linux, the installation script may attempt to install system dependencies using sudo -n. This is a non-interactive check that only proceeds if the environment already has passwordless sudo privileges, intended to facilitate headless environment setup for trusted browser tools.
[COMMAND_EXECUTION]: The skill executes various system and CLI commands, including the GitHub CLI (gh) for repository management and user-defined validation commands (e.g., pnpm test, cargo test) to verify the health of the project during the setup process.

om-setup-agent-pipeline