Containment Foreman - sandboxed agent runtime

Containment Foreman is a web-based agent that performs real work - writing and running code - with every action contained inside a sandbox designed to run on a Vultr VM. How it works: the foreman plans a task (for example: analyze the bids CSV and produce a JSON summary), dispatches each step to an isolated Docker worker, and returns a receipt: completion status, duration, artifact metadata, content and SHA-256 hashes. The centralized control layer plans, the disposable execution environment acts, and the output is verifiable. Defense in depth, all visible in the demo video: 1. Plan-level policy gate. An untrusted instruction pasted into the task is caught by the plan scanner - rejected_by_policy with the matched pattern - before anything is dispatched. 2. Sandbox containment. network=none, memory=512m, pids=64, secrets_passed=false, read-only root, docker --rm (destroy after run). A containment fixture that attempts to write a host sentinel leaves it unchanged. 3. Lifecycle caps. A sleep-10 task with a one-second cap times out cleanly; the host stays responsive and docker ps -a shows no leftover workers. Vultr usage: the project ships a one-file Vultr deployment kit (cloud-init + docker-compose) so the whole foreman runs on a Vultr VM, plus a provider-neutral LLM adapter (LLM_BASE_URL / LLM_API_KEY / LLM_MODEL) that points at Vultr Serverless Inference - planning happens on Vultr infrastructure too. Business value: safe automation for research, CSV analysis and internal operations - agents can execute, and unsafe execution no longer destroys trust.