About 15 minutes from an empty repository to a verified, scoped change. Every step has a headless variant for CI and IDE agents. Use a throwaway branch if the repository is not empty.
The steps
| # | Goal | Interactive | Headless |
|---|---|---|---|
| 1 | Install the CLI | npm install -g @jeger-ai/opengantry | npx @jeger-ai/opengantry --version |
| 2 | Scaffold | gantry init --tutorial | gantry init --yes --no-ci |
| 3 | Planner identity | gantry planner set "$(git config user.email)" | same (needs git config user.email) |
| 4 | Health check | gantry doctor | gantry doctor --json, exit 0 |
| 5 | Legislate | gantry legislate "First change" --msn MSN-0001 --skill-key logic --gate-command "npm test" | same, plus --yes |
| 6 | Planner commit | Review the YAML, then git commit -m "[MSN-0001] legislate mission" | same |
| 7 | Executor env | eval "$(gantry runtime env --mission .gitagent/missions/MSN-0001.*.yaml)" | gantry runtime env --mission … --json |
| 8 | Make the change | Edit inside the mission’s tmvc_roots | gantry runtime exec --mission … -- <agent command>, so the agent runs inside the process boundary |
| 9 | Record evidence | Append the gate output as a unique line to EXECUTOR_LOG.md | same |
| 10 | Verify | gantry verify --mission … | gantry verify --mission … --json |
| 11 | Audit | git log --grep='MSN-0001' --oneline | same |
Use the mission id and file name that step 5 wrote.
The same graph, day to day
# 1. The Planner reviews the mission before any executor runs
gantry legislate "Fix login spinner" --msn MSN-0001 --skill-key ui --gate-command "npm test"
git add .gitagent/missions/MSN-0001.<slug>.yaml
git commit -m "[MSN-0001] legislate mission"
# 2. Pin the mission and load the executor environment
gantry mission pin .gitagent/missions/MSN-0001.<slug>.yaml
eval "$(gantry runtime env)"
# 3. Verify and audit
gantry verify --fix
gantry verify --json | jq -r '.error_code // "passed"'
git log --grep='MSN-0001' --oneline
gantry doctorgantry legislate --from-intent "<story>" proposes the scope from the repository and asks you to approve it (Missions).
Pinning is optional but saves typing: gantry mission pin <file> makes verify and runtime env default to that mission, gantry mission pin with no argument shows the pin, and gantry mission unpin clears it.
What success looks like
gantry verifyexits0, with the gate and every trace row PASS.git log --grep='MSN-'shows the legislation commit from an allowlisted Planner email.gantry mission pinwith no argument prints the pinned mission.
When a step fails
Capture what happened. It is the most useful adoption feedback there is:
step: <number>
expected: <what should happen>
actual: <what happened>
error_code: <GXT_* from --json, if any>Look the code up in Troubleshooting, then file an issue or add it to your team runbook.
Compare it yourself
The repository ships a contrast benchmark: an improvised agent orchestrator script next to the same task under OpenGantry. Clone the repository, run npm ci && npm run build, then npm run examples:benchmark (-- --json for machine output). The specimens live in examples/contrast-agent-script/ and examples/gantry-minimal/. The raw script compresses common anti-patterns into one file; most teams do not ship a monolithic orchestrator, but without a Git-native protocol, structure and auditability are improvised.
Next, wire your agent.