The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the niri command-line tool to interact with the compositor.
Evidence: scripts/niri.py and scripts/niri_ctl.py use subprocess.run() and subprocess.Popen() with list-formatted arguments (e.g., ['niri', 'msg', '--json', ... ]), which effectively prevents shell injection vulnerabilities.
[PROMPT_INJECTION]: The skill provides access to desktop metadata that could be exploited for indirect prompt injection.
Ingestion points: scripts/niri_ctl.py and scripts/niri.py fetch window titles and application IDs directly from the compositor environment.
Boundary markers: Absent. There are no explicit markers or 'ignore' instructions wrapped around the window metadata when it is retrieved.
Capability inventory: The skill allows for significant session control, including the ability to close windows, switch workspaces, and execute new processes via the niri msg action spawn command.
Sanitization: Absent. Window titles, which may contain arbitrary text from web browsers or other applications, are passed to the agent without sanitization or filtering.

niri-ipc