The Agent Skills Directory

[COMMAND_EXECUTION]: Uses sudo apt-get to install system packages such as cage, wtype, ffmpeg, and grim, which allows for privilege escalation to root to modify the system environment.
[COMMAND_EXECUTION]: Dynamically constructs and executes shell commands for terminal automation, including the use of absolute file paths and session management in global temporary directories (/tmp/tctl-sessions/).
[EXTERNAL_DOWNLOADS]: Fetches and installs multiple external dependencies from npm (tuistory), pip (asciinema), and cargo. Specifically, it downloads and compiles the agg tool directly from a remote GitHub repository using cargo install --git.
[PROMPT_INJECTION]: Operates as an automation driver that processes output from external applications (TUIs and web apps) to determine its next actions. This architecture is vulnerable to indirect prompt injection, as a malicious target application could produce output designed to influence the agent's behavior.
Ingestion points: Terminal output via tuistory/tctl and web page content via agent-browser.
Boundary markers: No specific delimiters or safety warnings for ignoring embedded instructions in external data are implemented.
Capability inventory: Extensive subprocess execution (tctl, bash, remotion), file system writes, and network-enabled browser automation.
Sanitization: The skill does not define sanitization or validation protocols for the data ingested from target applications.

droid-control