The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute several local scripts and CLI tools, including scripts/spawn-confined.sh and the doctrine tool suite for git worktree management. These commands are integral to the skill's core function of process orchestration and are used for creating forks, importing changes, and managing the development lifecycle.
[INDIRECT_PROMPT_INJECTION]: The skill describes a multi-agent architecture where an orchestrator prepares prompts for worker agents. It mandates specific safety instructions within those prompts, such as forbidding destructive git commands like reset or stash. While this creates an injection surface where a sub-agent receives instructions, the skill includes explicit guidance on prompt distillation and maintaining strict boundaries.
Ingestion points: The <PROMPT_FILE> passed to scripts/spawn-confined.sh (SKILL.md).
Boundary markers: The skill instructs the orchestrator to provide a "pre-distilled and self-contained" prompt to ensure the worker inherits no previous conversation state (SKILL.md).
Capability inventory: Includes process spawning, file writing (via worktree import), and git operations (SKILL.md).
Sanitization: Relies on kernel-level OS confinement (e.g., bwrap) to limit the worker's file system access and prevent unauthorized commits (SKILL.md).
[DYNAMIC_EXECUTION]: The skill facilitates the runtime execution of AI harnesses (claude or pi) inside a confined environment. The execution flow is managed by scripts/spawn-confined.sh, which enforces mandatory timeouts and restricts the process to a read-only view of the coordination tree, mitigating risks associated with executing untrusted logic.

dispatch-spawn