The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a workflow to launch new agent instances via a shell command: tmux new-window -n "$TITLE" "AGENT_MAILBOX_DIR='$ROOT' AGENT_MAILBOX_ID='$child_id' $AGENT \"$PROMPT\"". This represents a dynamic execution surface where the shell command is assembled at runtime using variable interpolation.
[COMMAND_EXECUTION]: The variables $TITLE and $PROMPT are derived from untrusted sources, including the operator's conversation and the current focus context. The lack of explicit instructions to sanitize or escape these variables creates a significant risk of command injection. An attacker or malicious input could use shell metacharacters (e.g., ;, |, $(...)) to break out of the intended command string and execute arbitrary code in the host shell environment.
[COMMAND_EXECUTION]: Analysis of indirect injection surface (Category 8 rules):
Ingestion points: Operator's request ("dispatch to ", "dispatch this to a new agent") and the focus context used for the kebab-case slug (file: SKILL.md).
Boundary markers: None present in the tmux command template or the opencode prompt example.
Capability inventory: Shell process spawning via tmux and the $AGENT CLI (file: SKILL.md).
Sanitization: Absent; the instructions do not describe any escaping, validation, or filtering logic for the input variables before they are interpolated into the executable command string.

dispatch