The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures and processes output from a tmux terminal, which serves as a primary ingestion point for untrusted data that could contain malicious instructions.\n
Ingestion points: Terminal output is captured via tmux capture-pane in internal/tty/session.go and through tmux control mode %output bytes in internal/tty/control_*.go.\n
Boundary markers: Output is rendered through a structured tty.Model and screenmodel adapter, though the skill instructions do not explicitly define delimiters for the agent when interpreting this data.\n
Capability inventory: The skill uses exec.Command to interact with the tmux CLI, facilitating session management, pane resizing, and input injection. It also supports asynchronous editor sessions for file manipulation.\n
Sanitization: The documentation explicitly mandates that source OSC (Operating System Command) sequences be treated as untrusted, requiring a sanitizer and fuzz oracle to handle complex escape sequences and prevent terminal-based attacks.\n- [COMMAND_EXECUTION]: The skill relies on executing the tmux binary to perform its core functions.\n
Evidence: Uses os/exec to run tmux for querying cursor positions (display-message), resizing panes (resize-pane), and sending user input (send-keys). The use of the -l flag with send-keys is noted as a security measure to prevent command interpretation of printable characters.

shell-integration