TheHefty/jvsl.env.agents.container

A Visual Studio Code extension that gives every project its own development container, built for
working with AI coding agents. It composes the image from the stacks a project needs, builds it,
writes the dev container configuration, and reopens the window inside it.
Installing the extension is the only setup a project needs. The image's definitions ship inside the
extension, so a project carries no template, no submodule and no setup script.
## Why
The editor stays on the host. Only the toolchains, the agents and a nested Docker daemon run
inside the container, under its CPU and memory limits. The editor is never starved by a build it
is displaying.
Agents run sandboxed by default. Claude Code and the OpenAI Codex CLI start inside
ai-jail, which confines them to the workspace.
One project, one container. Each project declares its stacks, versions and resource limits in
a small manifest at its root, and gets its own image and its own Docker daemon. Claude Code's
sign-in and settings come from the host's ~/.claude, so they are set up once, not per project.
## Requirements
| Host | Linux, with Docker |
| Editor | Microsoft's build of Visual Studio Code, 1.90 or newer |
| Extension | Dev Containers, installed automatically as a dependency |
Dev Containers is licensed only for Microsoft's build of VS Code, which is why other builds are not
supported.
## Installation
The extension is distributed as a .vsix package. Build it from source:
sh
npm ci
npm run package
code --install-extension jvsl-env-agents-container.vsix
CI also publishes the package as a build artifact named vsix on every run that changes code.
## Getting started
Open the Command Palette and run one of:
| Command | What it does |
| Agent Container: Create a Project… | Asks for stacks, versions, limits and a location. It writes the project, initialises git, builds the image and reopens the window inside the container |
| Agent Container: Open a Project… | Opens an existing folder. A folder without a manifest goes through the same questions a new project answers |
| Agent Container: Build the Image | Composes and builds the image in a terminal, without reopening the window |
| Agent Container: Configure Stacks and Limits | Changes the project's stacks and resource limits |
| Agent Container: Show What Was Detected | Reports what the extension read about the host and the project |
| Agent Container: Show the Work | Opens the project's tracker as a board, read-only, when the manifest turns the tracker on |
| Agent Container: Migrate from code-server | Moves a project built on the old code-server template to this format, one step per run: the files, then a rebuild, then the planning into the tracker. Each step shows its plan and waits for Apply |
The same entries are available from the Agent Container view in the Explorer.
## The project manifest
A project is described by .agent-container.stack.json at its root. The extension writes it and
keeps it up to date, and it can also be edited by hand:
json
"node": "22",
"python": "3.12",
"limits": {
"cpus": 4,
"memory": "6g"
Each stack is a key, and its value is the version to install. Limits apply to the whole container.