Skip to content
a-mcorpPublic

About

Enforce (a) skill

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

4 Commits

Folders and files

Repository files navigation

This skills bundle has been generated using 'A-Skills'.

A-Skills

A-Skills is an agent system, which doubles as an interface to the 'a-mcorp' repositories. It was designed to provide a sandboxed container allowing the user to either elevate their own skills or study LLMs in live production scenarios. The assistant enhances the user's own knowledge and skills, rather than replacing them.

I do strongly emphasise the last part

Its skills define how an assistant writes, names, comments and contributes, and its container images give that assistant a ready, sandboxed workspace. Together, they describe how my repositories are designed and how to contribute to them.

If you ever wish to contribute to my 'emporium', simply load 'A-Skills', open a repository, and the assistant follows the (almost) same conventions as the projects' Maintainer.

Metadata

  • Name: 'A-Skills'.
  • Purpose: An agent system and contribution interface for the 'a-mcorp' repositories - skills, container images and their configuration.
  • Format: Agent Skills framework / Docker Container.
  • License:

Features

Features
  • Bundled skills:
    • 'style': answer length, spelling and text formatting, comments, references, contributing, CHANGELOG.md and CODEOWNERS conventions.
    • 'ops': naming notation, Terraform and Python conventions.
    • 'hardware': dimensions, bills of materials, 3D printing and photo conventions.
    • 'obsidian': Obsidian Markdown conventions.
  • Two container images: GitHub Copilot and Claude Code, each with Visual Studio Code, the desktop app and the CLI. These images will continuously receive dedicated image extensions, the first one being a hardware module.
  • Hardened container: runs with --cap-drop=ALL and no-new-privileges, with only the selected folders mounted.
    • --userns=keep-id maps the container's 'agent' user to the host user, so root inside the container is not root on the host.
    • Only the chosen folders are mounted, and /run/user/1000 is an in-memory tmpfs that disappears with the container.
    • --security-opt label=disable turns off SELinux labelling, because the default labels block the Wayland socket and /dev/dri, which the window needs. The isolation above still applies.
  • Native Wayland window passthrough, without X11.
  • Optional microphone passthrough via the host's PipeWire/PulseAudio socket, for voice input, per Dockerfiles/README.md. voice commands
  • Pre-configured Visual Studio Code settings, documented in SETTINGS.md.
Podman, before logging in Copilot, with the user's settings.json Claude Code, with the user's settings.json
A-Skills Agent container with Wayland passthrough, right after provisioning A-Skills Copilot Agent container with Wayland passthrough and fully configured A-Skills Claude Agent container with Wayland passthrough and fully configured

Setup

Dependencies Usage
  • None (Any OS, no isolation)
  • GNOME Boxes/VM (Linux, full isolation)
  • Podman/Container (Linux, full isolation)

Note: The images are Linux containers only, on both Linux and Windows hosts. GUI passthrough is not possible in native Windows containers by design (Windows-Containers #306).

AGENTS.md tells the agent which skill to apply and when, while CLAUDE.md imports it for Claude Code. Each SKILL.md holds the general rules and points to its extension files, which the agent reads only when the task requires them.

To add a custom skill, create a new folder under .github/skills/ with a SKILL.md file (with name and description in its frontmatter), then reference it in AGENTS.md.

The skills can be used in one of the following modes:

  • 'no isolation' (not recommended)
    • In this mode the skills in this repository can be used as-is and the agents roam free across the entire host operating system, performing the magic of rm or the like on random files 👍 You do it at your own risk. If your fridge runs out and the electric car in the garage starts smoking, it is your problem
  • GNOME Boxes (not great)
    • In this mode the skills in this repository can be used inside a VM. GNOME Boxes offers the majority of Linux distributions in a drop-down list. The Operator can select which files on the host are shared with the guest OS. This is not a very performant setup, as there are significant lag and I/O issues when operating inside the guest OS's VSC and CLIs. Mostly a 'click-ops' type of setup.
    • The setup of the VM is on the Operator. This repository does not cover it.
    • This option supports simple host-mirrored configuration of the IDE and support software, along with compatibility with immutable distributions and Flatpaks.
  • Podman (recommended, works as designed)
    • In this mode the skills in this repository can be used inside a container. Podman may build one of the included .dockerfile images:
      • GitHub Copilot: Visual Studio Code with the GitHub Copilot extension, the GitHub Copilot App and the GitHub Copilot CLI.
      • Claude Code: Visual Studio Code with the Claude Code extension, the Claude Desktop App and the Claude Code CLI.
    • This is most likely the ideal solution, as folders can be mounted directly to the container, while leveraging the native Wayland display server. The concept offers supreme convenience at low risk.
    • This mode supports multiple entrypoints:
      • GitHub Copilot: copilot-vsc, copilot-app, copilot-cli, browser, terminal.
      • Claude Code: claude-vsc, claude-app, claude-cli, browser, terminal.

Installation

  • clone the repository duh
  • (optionally) rename the folder to .github or .claude or .agents, depending on the harness, per documentation by GitHub.
    • e.g. Claude Code only loads skills natively from the .claude/skills directory. They still work just fine through the AGENTS.md instructions.
  • (optionally) adjust Visual Studio Code settings per the settings.json snippet (pre-loaded automatically for the container-sandboxed agent).
    • these settings are overwritten by the user-specific settings.json file synchronised from the logged-in user!

Linux-only

  • (optionally) install GNOME Boxes (e.g. Flathub)
  • (optionally) install Podman Desktop (e.g. Flathub)
    • check whether the Podman socket is running
      • systemctl --user enable podman.socket
      • systemctl --user start podman.socket
      • systemctl status --user podman.socket
    • build the image using Podman with Dockerfiles/ as the context directory.
      # Replace 'Copilot' with 'Claude' for the Claude Code image
      podman build -f Dockerfiles/Copilot/askillsAgent.dockerfile \
        -t docker.io/library/askillsagent Dockerfiles/
    • run the image using the podman run command from Dockerfiles/README.md.
    • share files with the agent by placing them in the host's $HOME/askillsagent directory, which is mounted as /workspaces/main-workspace inside the container.

Repository structure

.
├── .github
│   ├── skills
│   │   ├── hardware           # Dimensions, bills of materials, 3D printing, photos
│   │   ├── obsidian           # Obsidian Markdown conventions
│   │   ├── ops                # Naming notation, Terraform, Python
│   │   └── style              # General styling and contributing rules
│   ├── CHANGELOG_TEMPLATE     # Template for CHANGELOG.md
│   ├── CODEOWNERS_TEMPLATE    # Template for CODEOWNERS
│   ├── PULL_REQUEST_TEMPLATE  # Templates for pull requests
│   └── pull_request_template.md # Default pull request template
├── Dockerfiles
│   ├── Claude                 # Claude Code image, its hardware image and entrypoint
│   ├── Copilot                # GitHub Copilot image and entrypoint
│   ├── PROMPTS.md             # Prompts for maintaining the settings files
│   ├── SETTINGS.md            # Visual Studio Code settings reference
│   ├── vsc-claude.json        # Pre-loaded Claude Code settings
│   └── vsc-settings.json      # Pre-loaded Visual Studio Code settings
├── docs                       # Documentation assets (screenshots)
├── AGENTS.md                  # Skill workflow for all agents
├── CLAUDE.md                  # Imports AGENTS.md for Claude Code
├── LICENSE                    # AGPL-3.0
├── NOTICE                     # Copyright and the additional terms (attribution, name)
└── TO-DO.md                   # Planned work, as user stories

Troubleshooting

Podman socket

socket unresponsive
  • the command starts Podman Desktop (Flatpak) in CLI mode with an ls command to read the socket; the spawn command creates a test Podman process.
  • flatpak run --command=sh io.podman_desktop.PodmanDesktop -c \
      'ls -l $XDG_RUNTIME_DIR/podman/; flatpak-spawn --host podman --version'
  • test if there are any stale connections using podman system connection list
  • if nothing seems wrong, then congrats. Perhaps, like me, you managed to spawn so many processes on the OS that the system has random issues. I occasionally experience this in GitKraken's "file watching failed to start for this repository" error, but also in GNOME System Monitor's "Unable to find default local file monitor type".
  • close some VSC instances or browser windows, or restart the host. Essentially increase fs.inotify.max_user_watches and fs.inotify.max_user_instances, per my article.

About

Enforce (a) skill

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages