The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash to orchestrate a complex pipeline including Python scripts, shell scripts, and Expo/React Native CLI tools. These operations are strictly scoped to the /tmp/expo-skill-eval-* and /private/tmp/expo-skill-eval-* directories via allowed-tools frontmatter. The use of subprocess.run within Python orchestrators ensures that commands like xcrun simctl and adb are children of authorized processes.
[EXTERNAL_DOWNLOADS]: The skill fetches Expo templates and dependencies using standard package managers (bun, npm) and official Expo tools (create-expo-app). These operations target official registries (NPM) and are used to create clean fixtures for testing. This is standard behavior for development and testing tools.
[REMOTE_CODE_EXECUTION]: While the skill spawns claude -p subprocesses to test other skills, these are executed with user-confirmed permission flags (--dangerously-skip-permissions or --permission-mode acceptEdits). The environment is isolated within temporary workspaces to prevent side effects on the host repository.
[OBFUSCATION]: No obfuscated code, hidden URLs, or suspicious character encodings were detected. The scripts are written in plain, well-documented Python and Bash.
[PRIVILEGE_ESCALATION]: There are no attempts to use sudo or modify system-level configurations. Operations are restricted to the user's temporary directories and standard simulator/emulator management tools.
[DATA_EXFILTRATION]: No network exfiltration patterns were identified. The generate_viewer.py script creates a self-contained local HTML file. An optional "Publish a shareable Artifact" feature uses the platform's native Artifact tool only after explicit user opt-in during the scope clarification phase.

expo-skill-eval