The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on tmux send-keys to run processes such as npm run dev or redis-server in background panes. This grants the agent the ability to execute arbitrary commands in the system shell based on user instructions or detected service needs.
[INDIRECT_PROMPT_INJECTION]: The skill captures output from external processes to verify status and detect errors, which creates a potential vector for malicious instructions to influence the agent.
Ingestion points: Captured via tmux capture-pane -t ... -p -S -20 and -S -50 in SKILL.md and references/commands.md.
Boundary markers: No specific boundary markers or "ignore instructions" warnings are defined for the captured content to distinguish it from agent instructions.
Capability inventory: The skill can execute arbitrary commands (send-keys), create/kill windows (new-window, kill-window), and view process output history.
Sanitization: No sanitization or filtering is applied to the output captured from the panes before it is processed by the agent.

tmux-aware