ai-coding-agents-safety-envelope
AI Coding Agents Safety Envelope
Use this skill to design or review the two layers that bound what a coding agent may do: the approval system (who decides: permission modes, rule sources, plan-mode transitions, prompt routing, leader-worker handoff) and the execution sandbox (what is reachable: process isolation, filesystem mounts, network policy, environment exposure, destructive-command controls).
Approval policy decides who decides; the sandbox decides what is reachable. Neither substitutes for the other, and both are defined here so their interaction (auto-allow inside the boundary, escalation at its edge, worker envelopes) is designed once.
This skill does not own remote execution or managed cloud sandboxes (ai-coding-agents-cloud-sandboxes), hook automation (../agents-hooks/SKILL.md), full settings-source layering and managed policy (../ai-coding-agents-settings-policy/SKILL.md), plugin trust (../ai-coding-agents-plugins/SKILL.md), or the tool execution pipeline (../ai-coding-agents-runtime-core/SKILL.md).