The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an indirect prompt injection surface by aggregating data from external sources and injecting it into prompts for other agents.
Ingestion points: The script scripts/watchdog_controller_proxy.sh ingests data from br (bead tracking tool) and tmux pane metadata (titles and captures).
Boundary markers: The skill uses delimiters such as <<<WATCHDOG TICK>>> and <<<END WATCHDOG TICK>>> to wrap injected content, which provides some separation but does not fully neutralize malicious instructions embedded in the ingested data.
Capability inventory: The skill uses ntm --robot-send to transmit instructions to a controller agent and tmux send-keys to interact with shell panes. If the ingested data contains malicious instructions, they could influence the controller agent's subsequent tool usage.
Sanitization: The skill performs basic sanitization using strip_ansi to remove formatting and rg for pattern matching, but it does not validate or escape the content of bead identifiers or pane titles before interpolation into the coordination prompts.

controller-proxy-watchdog