Yolobox – Run AI coding agents with full sudo without nuking home dir

Executes AI agent commands with full sudo privileges within an isolated container environment that prevents destructive filesystem operations from affecting the host system. Uses container-level isolation (likely Docker or similar) to create a bounded execution context where agents can run privileged commands without risk of corrupting the host home directory or critical system files. The sandbox intercepts and restricts filesystem write operations to designated safe zones while allowing read access to necessary system resources.

Manages the lifecycle of AI agent commands from parsing agent intent through execution, output capture, and result feedback. Implements a command execution pipeline that accepts directives from AI models (likely via structured prompts or function-calling APIs), translates them into shell commands, executes them in the sandboxed environment, and returns structured results back to the agent for iterative decision-making. Handles command queuing, timeout management, and error propagation.

Implements a filesystem access control layer that intercepts write operations and restricts them to explicitly allowlisted directories while blocking writes to sensitive paths (home directory, /etc, /root, system binaries). Uses either kernel-level syscall filtering (via seccomp or AppArmor) or filesystem-level hooks to enforce a whitelist of safe working directories where agents can create/modify files. Reads are generally unrestricted to allow agents to inspect system state.

Provisions isolated working directories for each agent execution with automatic cleanup after completion. Creates a temporary filesystem namespace or directory tree for each agent run, ensuring agents cannot access files from previous executions or other concurrent agents. Implements automatic garbage collection to remove temporary files, preventing disk space exhaustion from repeated agent runs. May use copy-on-write (CoW) filesystems or union mounts for efficient isolation.

Monitors running agent processes for resource consumption (CPU, memory, execution time) and enforces configurable timeout limits to prevent runaway processes from consuming system resources indefinitely. Implements process-level monitoring with hard kill capabilities when timeouts are exceeded or resource limits are breached. Captures execution metrics and logs for debugging and auditing purposes.

Parses command execution output and structures it for agent consumption, extracting exit codes, stdout/stderr streams, and execution metadata into a format suitable for agent reasoning. Implements output sanitization to remove sensitive information (API keys, credentials) before feeding results back to agents. May include output truncation for very large results to prevent context window overflow in LLM-based agents.

Manages concurrent execution of multiple AI agents within the same sandboxed environment while preventing resource contention and ensuring fair allocation. Implements process scheduling, CPU/memory quota enforcement per agent, and isolation mechanisms to prevent one agent's resource consumption from starving others. Uses cgroup-based resource partitioning to enforce per-agent limits while allowing efficient resource sharing.