The Agent Skills Directory

[DATA_EXFILTRATION]: The skill captures terminal output history (up to 400 lines) from all active tmux panes within a project's git repository scope. This data, which may contain sensitive environment variables, secrets, or proprietary code displayed in the terminal, is ingested into the agent's context for summarization via the tmux capture-pane command.
[COMMAND_EXECUTION]: The skill relies on local execution of tmux and git commands to list sessions, capture pane content, and send keystrokes. It utilizes execFile in Node.js to interface with these system utilities, which is generally safer than shell-based execution but still interacts directly with the host environment.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it ingests and summarizes terminal output from external processes. Maliciously crafted output in a monitored tmux pane could attempt to manipulate the agent's summary, cross-analysis results, or suggested propagation messages.
Ingestion points: The scripts/tmux-capture.mjs script reads terminal buffers directly into the agent context.
Boundary markers: The workflow attempts to mitigate risk by using sub-agents to summarize individual panes before the main agent processes them, providing a layer of isolation, though raw terminal text is still processed by the model.
Capability inventory: The skill can write back to any tmux pane (tmux send-keys) and has access to the agent's full capability set.
Sanitization: The scripts/tmux-send.mjs script includes a foldMessage function that collapses newlines to prevent multi-line command injection into terminal interfaces, and the skill instructions mandate human approval and dry-run verification before any write operations.

tmux-orchestrate