The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands to manage agent orchestration.
spawn_subagent.sh executes a user-provided or default command (e.g., devin --permission-mode dangerous) inside a new tmux window using tmux new-window.
send_command.sh transmits literal text to a shell running inside a tmux window using tmux send-keys -l followed by an Enter key signal.
exit_subagent.sh sends termination commands and potentially kills tmux windows to ensure cleanup.
[PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection common to orchestration tools.
Ingestion points: The send_command.sh and spawn_subagent.sh scripts accept command strings that are often derived from agent reasoning or data processing.
Boundary markers: No specific boundary markers or instructions are provided to distinguish between orchestration commands and potentially malicious instructions embedded in the data processed by subagents.
Capability inventory: The skill has the capability to execute shell commands, manage processes via tmux, and write/read files from the /tmp directory as seen in the workflow examples in SKILL.md.
Sanitization: The scripts use send-keys -l to prevent the tmux binary from interpreting special characters within the command string, though they do not sanitize the content for the target shell.

tmux