The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill monitors and relays output from worker agents running in tmux sessions. This data enters the agent context and could potentially contain malicious instructions from a compromised or adversarial worker agent designed to influence the orchestrator.
Ingestion points: Content is ingested from tmux panes using capture-pane via the scripts/wait_for_idle.py and scripts/broadcast.sh helpers.
Boundary markers: The orchestrator uses a split completion marker (e.g., TAC_DONE_$suffix) to identify the end of a worker's reply, as defined in references/delivery-and-waiting.md.
Capability inventory: The skill has the capability to execute tmux commands, including send-keys to inject input into panes and kill-session to terminate processes.
Sanitization: The skill filters captured content to identify new lines (new_lines function in scripts/wait_for_idle.py) but does not perform content-based sanitization of instructions.
[COMMAND_EXECUTION]: The skill relies on executing tmux commands and helper scripts to perform its functions. This involves extensive use of shell execution and Python subprocess calls.
Evidence: scripts/preflight_send.py and scripts/wait_for_idle.py use subprocess.run to call the tmux binary for session inspection and pane capture.
Evidence: scripts/broadcast.sh uses tmux send-keys to dispatch messages to multiple agents concurrently.
Evidence: references/tmux-recipes.md provides shell command patterns for resolving script paths and spawning new tmux sessions with specific agent commands.

tmux-agent-comms