Autonomous Execution In Restricted Environments

Modern AI agents often fail in sandboxed or restricted terminal environments because they rely on direct process execution. This skill replaces “fragile execution” with a robust, Health-Gated architecture designed for reliability in hostile or restricted shells (Windows AppContainers, sandboxed cloud IDEs, etc.). The skill implements a “Direct Launcher Pattern” and a mandatory health-check cycle. Instead of guessing if a command worked, the agent generates auditable, repo-resident scripts and performs multi-dimensional health checks before and after every action. If the system state is unhealthy, the agent stops and reports rather than causing cascading failures. By using file-based script orchestration, it works even when run_terminal_command is blocked or unreliable, and it implements machine-readable health gates that prevent the agent from pushing code or running builds on a broken system. The output includes structured AGENT PUSH RESULT blocks containing commit hashes, composite health scores across multiple dimensions, and explicit “streaks” of successful operations, providing a clear audit trail for parent agents or human supervisors.