| name | kgym-deploy |
| description | Deploy, upgrade, shut down, or troubleshoot kGymSuite for live-kBench in local Docker Compose or GCP/remote multi-server environments. Use when Codex is asked to deploy kGym, start local kGym services, configure deployment/local or deployment/gcp, run kgym.py new-deploy/upgrade/down/config-artifact-reg, prepare kGym storage, or verify scheduler/dashboard/worker services. |
kGym Deploy
Workflow
First determine the target environment from the user request:
- Local development: use Docker Compose with
kGym/deployment/local/compose.yml.
- GCP or remote multi-server: use
kGym/kgym.py and a copied deployment directory under kGym/deployment/<name>.
Read references/deployment.md for concrete commands, config fields, and verification checks.
Local Deployment
- Work from
kGym/.
- Ensure userspace images exist in the local storage bucket before expecting VM jobs to run.
- Build with
DEPLOYMENT=local docker compose -f deployment/local/compose.yml --project-directory . build.
- Start core services before workers:
kmq, kscheduler, kdashboard, then kbuilder, kvmmanager, and optionally kprebuilder.
- Verify with
docker compose ps, scheduler docs at http://localhost:8000/docs, and dashboard at http://localhost:3000.
GCP or Remote Deployment
- Create a deployment directory from
deployment/gcp, then update config.json, kgym-runner.env, and image tags in compose.yml.
- Confirm GCP prerequisites before running commands: Compute VMs, Artifact Registry, GCS bucket, IAM permissions, and authenticated
gcloud.
- Build and push images before
kgym.py upgrade or first remote bring-up.
- Run
python kgym.py <deployment> new-deploy, then python kgym.py <deployment> config-artifact-reg <artifact-registry-host>.
- Upload
buildroot.raw and bullseye.raw to the configured GCS bucket under userspace-images/.
Safety
- Do not run
down, remove volumes, or overwrite deployment configs unless the user asked for that action.
- Do not invent project IDs, bucket names, deployment names, hostnames, or registry hosts.
- For troubleshooting, inspect service logs and config before recommending rebuilds.
