The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/cmux_fleet.py script manages peer terminal sessions by executing cmux commands and peer-specific instructions.
Evidence: The script uses subprocess.run to interact with the local cmux binary for surface management and task execution.
Evidence: The launch_command function utilizes shlex.quote and shlex.join to safely construct shell commands for the peer sessions, mitigating command injection risks.
[INDIRECT_PROMPT_INJECTION]: The coordination of multiple autonomous agents sharing data through a mailbox system creates a potential attack surface for indirect prompt injection.
Ingestion points: The coordinator agent processes structured JSON files containing peer results from the .ai-workflow/peer-sessions/<run-id>/replies/ directory.
Boundary markers: SKILL.md provides explicit templates for briefing peers that include safety instructions and scope constraints to prevent unauthorized behavior.
Capability inventory: The skill possesses the capability to launch interactive terminals, execute shell commands via cmux, and perform local file system operations.
Sanitization: Data exchange is restricted to a structured schema, and the coordinator uses shlex for command-line escaping to prevent malicious interpolation of peer data into system commands.

peer-sessions