The Agent Skills Directory

[SAFE]: The skill defines a localized coordination framework that relies on the local filesystem rather than network-based communication, reducing the attack surface for remote interception or exfiltration.
[SAFE]: Agent name and team name validation (limiting characters to alphanumeric, underscores, and hyphens) effectively prevents shell command injection during the tmux pane spawning process.
[SAFE]: State persistence for team configurations and message inboxes uses a two-phase atomic write pattern (mkstemp + os.replace), preventing partial writes and ensuring consistency across concurrent agent processes.
[SAFE]: Concurrency safety is maintained through fcntl exclusive file locking on team and task directories, protecting against race conditions and data corruption during multi-agent operations.
[SAFE]: The health monitoring protocol implements a structured heartbeat and 2-stage stall detection system to recover stalled processes without escalating privileges or using unsafe execution methods.

agent-teams