SandboxOps: Blast Radius Zero on Vultr Compute

SandboxOps: plan → sandbox → verify → evidence. Built for lablab Vultr Agent Rush (Challenge 1: Blast Radius Zero). Give an AI agent real work without a real shell. Flagship: Vultr Cloud Compute + Docker. Live on a Chicago Ubuntu 24.04 Cloud Compute VM (2 GB, systemd sandboxops): http://64.177.13.126/ — every agent action runs in a throwaway Docker container (--network none, read-only rootfs, non-root, cap-drop ALL, CPU/memory/pids caps, auto-removed). GET /health → backend=docker, planner=vultr-inference. One-script deploy (deploy/vultr-deploy.sh). No GPUs. API: GET /health, GET /api/tasks, POST /api/run. Baseline vs flagship: local process-sandbox (rlimits, stripped env, Python socket block, timeout) is best-effort for laptops. Production isolation is SANDBOX=docker on the Vultr VM. Vultr Serverless Inference (glm-5.3 via api.vultrinference.com/v1) is the planner that turns a prompt into short Python—not the differentiator; the Compute sandbox is. Without an inference key, demo tasks still run offline. Synthetic/demo tasks (live 2026-11-05 America/Chicago): bill → VERIFIED $551.00; free-form "sum of first 50 primes" → 5117 via glm-5.3; network_attempt → CONTAINED; fork_bomb_ish → CONTAINED by timeout/CPU limit. Each run returns evidence: code, stdout, stderr, exit code, and a per-run SHA-256 of the JSON record (hashlib in app.py; see live UI after a run—no fixed hash claimed). Automated suite: test_agent.py 3/3 (bill verified, network blocked, loop killed); test_app.py 5/5 (/health, verified run, two containment runs, UI). MIT. Buyer: finance and healthcare ops needing LLM reconciliations/bill checks without risking servers or data—the containment demo (network + loop) is the Blast Radius Zero moment. Limits: free-form prompts not auto-verified; containers share the VM kernel (gVisor next). Repo: https://github.com/igoreduardoai/sandboxops