The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides various tools (send, paste, broadcast, dispatch) to execute commands in tmux panes. These are protected by scripts/safety-check.sh and a is_dangerous_text regex that identifies destructive operations (e.g., rm -rf, mkfs, git reset --hard), though users can override these with a --force flag.
[DATA_EXFILTRATION]: The skill captures the contents of tmux panes for audit and evidence collection purposes. To mitigate the risk of data exposure, it implements a redact function in scripts/auto-tmux.sh and scripts/remote-readonly.sh that masks common secret patterns, such as OpenAI/Anthropic API keys and Bearer tokens.
[PROMPT_INJECTION]: The skill facilitates multi-agent coordination using role-play instructions in its "Codex Pilot Mode" (references/codex-pilot-mode.md). While these instructions use override patterns (e.g., "you are now the master Codex"), they are used to establish structured agent roles for the intended automation workflow.
[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it reads data from interactive tmux panes and executes subsequent commands based on those observations.
Ingestion points: scripts/auto-tmux.sh (capture/inspect) and scripts/remote-readonly.sh.
Boundary markers: None identified beyond instructional warnings.
Capability inventory: Arbitrary shell command execution via tmux send-keys across the script library.
Sanitization: Automated redaction of secrets and dangerous command pattern matching in scripts/safety-check.sh.

auto-tmux