The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructs the agent to execute sudo apt-get install commands to install system-level dependencies including cage, wtype, grim, wf-recorder, and ffmpeg.
[EXTERNAL_DOWNLOADS]: The installation instructions include cargo install --git https://github.com/asciinema/agg, which downloads and compiles code from a remote repository at runtime.
[DYNAMIC_EXECUTION]: The skill documents patterns for dynamically constructing shell command strings (e.g., building complex tctl launch commands with resolved paths and session arguments) to be executed by worker agents.
[COMMAND_EXECUTION]: The skill relies on extensive use of system binaries and CLI tools (tctl, tuistory, asciinema, agent-browser) to operate the environment and perform desktop control.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to interact with browsers and terminal TUIs, it has a significant attack surface where untrusted data from web pages or application outputs could influence agent behavior.
Ingestion points: Browser DOM, terminal byte sequences, application state snapshots.
Boundary markers: None explicitly implemented to separate untrusted content from control logic.
Capability inventory: File system access via /tmp/, network operations through agent-browser, and shell command execution via tctl.
Sanitization: No explicit sanitization of application output before it is used to determine subsequent actions.

droid-control