The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the management of external processes through tmux. It allows the primary agent to launch new processes, send inputs, and restore sessions from snapshots.
Evidence: The scripts/tmux_agent.py helper uses subprocess.run to invoke tmux commands like new-window, split-window, and send-keys. These commands are used to execute agent harnesses provided by the user.
[SAFE]: The tool incorporates several security best practices and safety boundaries to mitigate risks associated with process orchestration.
Evidence: Prompt delivery is handled via stdin pasting (load-buffer/paste-buffer) rather than shell interpolation, reducing the risk of shell injection.
Evidence: The helper script includes a check (SHELL_COMMANDS) to refuse sending multi-line agent input to panes running a standard shell by default, preventing accidental command execution in an unmanaged environment.
Evidence: The operating protocol (references/operating-protocol.md) explicitly instructs agents to capture the minimum necessary scrollback and to summarize rather than repeat sensitive data like credentials.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data generated by a supervised 'worker' agent, creating a surface for indirect prompt injection.
Mitigation: The protocol addresses this by mandating independent verification of worker claims and maintaining a clear boundary where the primary agent remains the authority and must not blindly trust worker output.

tmux-agent-orchestrator