The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of system commands to interact with Android devices and development tools.
Evidence: Multiple scripts in the scripts/ directories execute commands such as adb for device control, ./gradlew for building and testing, keytool for keystore generation, and emulator/avdmanager for virtual device management.
[EXTERNAL_DOWNLOADS]: The skill dynamically fetches instruction files from a remote repository at runtime.
Evidence: state-droid-remote-device/scripts/device-check.sh attempts to download the latest SKILL.md from https://raw.githubusercontent.com/hereisderek/StageDroid/ to resolve server capabilities.
[CREDENTIALS_UNSAFE]: The skill collects and stores sensitive user credentials locally to facilitate automation.
Evidence: state-droid-remote-device/scripts/device-wake-unlock.sh prompts for and caches test device passcodes in ~/.config/android-skill-derek/devices.json. While it applies restricted file permissions (0600), it represents a local store of sensitive data.
[DYNAMIC_EXECUTION]: The skill implements logic by dynamically executing Python code blocks embedded within shell scripts.
Evidence: Scripts such as device-pool.sh, device-ui-audit.sh, and device-wake-unlock.sh use heredocs to execute Python snippets for complex JSON and XML processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data which could contain malicious instructions targeting the agent's behavior.
Ingestion points: philipp-android-skills/pr-quiz/scripts/publish_quiz.py reads GitHub PR diffs and descriptions from the gh CLI; state-droid-remote-device/scripts/device-ui-audit.sh parses Android UI hierarchy XML obtained from uiautomator dump.
Boundary markers: No explicit sanitization or instruction-ignoring delimiters were identified for the natural language data processed.
Capability inventory: The skill possesses significant capabilities including device interaction (adb), repository modification (gh), and build execution (gradlew).
Sanitization: Content is parsed using standard libraries, but no specific security filters for prompt injection patterns were observed.

android-skill-derek