The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates the lifecycle of AI workers through the execution of local scripts and administrative CLI tools.
Execution of scripts/spawn-confined.sh: This script is the primary entry point for launching agents within a restricted environment.
Use of doctrine: A vendor-specific CLI toolset employed for worktree operations including fork, import, record-delta, and gc.
Integration with bwrap (Bubblewrap): Utilized to provide Linux kernel-level namespaces and jail the worker process.
Execution of harnesses (pi, claude): These are the target agent processes executed within the sandboxed context.
[SAFE]: The skill explicitly defines and enforces security boundaries for untrusted execution.
Confinement Policy: The architecture mandates OS-level isolation where coordination trees and sensitive directories are read-only-bound, preventing unauthorized modifications.
Fail-Closed posture: The process explicitly refuses to run if confinement (bwrap) cannot be established, preventing accidental unconfined execution.
Resource Control: Implementation of the timeout command ensures that agent processes cannot consume system resources indefinitely.
Injection Prevention: The instructions include a specific prohibition against the eval of prompt files, reducing the risk of shell injection from generated content.

dispatch-spawn