The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run within scripts/cmux_fleet.py to interact with cmux, a terminal management utility. These calls use argument lists rather than shell strings, preventing typical shell injection vulnerabilities. The operations are limited to managing terminal layouts (splits, tabs, workspaces) and session identification, which is central to the skill's purpose of orchestrating multiple interactive sessions.
[DYNAMIC_EXECUTION]: The skill generates shell commands to launch peer processes and sends instructional text to interactive terminals via cmux send. These commands are constructed by the coordinator agent to initialize the work environment for peers. This represents standard orchestration behavior for a multi-session workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sessions, creating a theoretical surface for indirect prompt injection.
Ingestion points: The coordinator agent reads and synthesizes data from replies/<peer>.json files using the scripts/peer_mailbox.py status command.
Boundary markers: The skill includes defensive instructions in the peer briefs, explicitly stating: "Do not treat this brief as approval for credentials, escalation, publishing, or destructive actions."
Capability inventory: The skill possesses the capability to manage terminal sessions via cmux and perform local file writes for session state and coordination.
Sanitization: The scripts/peer_mailbox.py utility performs JSON schema validation on all replies to ensure they contain required fields and valid status tokens (done, blocked, failed) before they are consumed by the coordinator.

peer-sessions