/\
      .______.                     /  \
      |\    /|                    /    \
      |/____\|                   /      \
      |\    /|                  /        \
      |/____\|                  |        |
      |\    /|                  | /---\  |
      |/____\|=================>| |O/_|  |
      |\    /|                  | \---/  |
      |/____\|                  |        |
      |\    /|                  | #####  |
      |/____\|=================>| #####  |
      |\    /|                 /|        |\
      |/____\|                / |        | \
      |      |               /__|________|__\
      '______'
  ________________________________________________

     ____     _     _   _  _____  ____  __   __
    / ___|   / \   | \ | ||_   _||  _ \ \ \ / /
   | |  _   / _ \  |  \| |  | |  | |_) | \ V / 
   | |_| | / ___ \ | |\  |  | |  |  _ <   | |  
    \____|/_/   \_\|_| \_|  |_|  |_| \_\  |_|  

GANTRY

The harness holds it, checks it, and never flies with it.

Automate the full SDLC of an OI Nebula project from inside the coding agent you already use.

GANTRY installs a set of /gantry slash commands into Claude Code, Cursor, Codex, Windsurf, GitHub Copilot and Gemini CLI. You point it at a project, and it runs the delivery cycle: read the current state, pick the work that matters, make the change, prove it works, and hand the project back.

The build-run-test work happens on your machine with your real toolchain — your local Docker, your test runner, a real browser — so what gets marked done is what actually ran. The platform's own containers are never used as a runtime.

1. Create a project

GANTRY drives a project hosted on OI Nebula. Create one there first, then note three things from the project console:

2. Install

pip install oi-gantry
gantry sync            # writes the commands into every coding agent on this machine

gantry sync --target claude (or cursor, codex, windsurf, copilot, gemini, antigravity) installs for one agent only. --scope project --dir <path> installs into a single project folder instead of your user config.

Full automation (on by default)

A GANTRY firing is unattended: every phase from pull onwards runs docker compose, git, the project's package manager and its test runner. So sync also grants the agent unattended read/write/execute/network access — for Claude Code it merges permissions.defaultMode: "bypassPermissions" into that scope's settings.json. Your existing settings are merged, never overwritten, and gantry sync --no-automation installs the commands without the grant (expect a prompt per action, and an unattended /loop firing to stall on the first one). --uninstall subtracts exactly the keys the installer added.

The other five agents have no config file this installer can write, so it prints the one-line change to make in each — Codex approval_policy = "never", Gemini --yolo, and the auto-run/Turbo/auto- approve toggles in Cursor, Windsurf and Copilot.

Install into a scope you are willing to run unattended: --scope project --dir <path> confines the grant to one project folder, while the default --scope user applies it everywhere.

3. Run it

Open your coding agent and type:

/gantry 10m /loop
{
  URL:        https://your-project-host,
  API_KEY:    your-key,
  Project_ID: proj-xxxx
}

That runs the delivery cycle on a repeating 10-minute firing: each pass takes the project, works as much of the backlog as it can, records what it did, and hands the project back before the next one starts. Change 10m to whatever interval suits the work.

The three values are the ones from step 1. They are read once at the start of the session and verified with a single read-only call before anything is touched.

For a single pass instead of a loop, just type /gantry.

Commands

/gantry on its own runs the whole cycle. You rarely need the individual commands — they exist so you can step in at any point, and so the agent can tell you exactly where it is.

Understand

Change

Decide and hand back

Ways to call it

Each loop firing is a complete session — it takes the project, works, and releases it. Nothing is held between firings, so a firing that dies never leaves the project locked for the next one.

Sub-commands are /gantry:<name> in Claude Code and Gemini CLI, and /gantry-<name> in Cursor, Codex, Copilot, Windsurf and Antigravity — the difference is how each agent namespaces its commands, not a difference in behaviour.

Questions are answered, not executed. When a request is ambiguous, GANTRY takes the cheaper, reversible path — reply, or ask — rather than making a change you did not ask for.

Supported agents

Claude Code · Cursor · Codex · Windsurf · GitHub Copilot · Gemini CLI · Antigravity

One package installs all seven. gantry sync after every upgrade.

Antigravity reads workflows from a different folder depending on scope — .agents/workflows/ in a project, global_workflows/ under ~/.antigravity for the user scope — and the installer handles that for you. Both are taken from the shipped build's own path constants, not inferred.

Staying current

gantry selfcheck            # is this copy current?
gantry selfcheck --update   # upgrade and rewrite every agent's commands

Sessions check this themselves before they start work, so an out-of-date copy is caught before it can run a cycle under superseded rules. If the index cannot be reached, the session carries on — a version check is never allowed to block delivery.

Requirements

Python 3.9+. No third-party dependencies. Docker must be installed and runnable on the machine you run GANTRY from — it is the only runtime for build, test, migrations and verification, and a session cannot fall back to the platform's containers. A browser is needed for the e2e steps. Both use whatever your project already uses.

© Orion Innovation. See CHANGELOG.md for what changes between releases.