Point hivectl at an issue and it runs Claude Code in a throwaway Kubernetes pod that opens a pull request. It's per-project — run it inside an app repo and every pod, secret, and log is scoped to that project, so one cluster hosts fleets for many repos at once. The tracker is GitHub Issues in the same repo as the code; each issue's Definition of Done and Plan are the authoritative success criteria.
Roles, the in-pod review gate, DoD & Plan, the k8s sandbox, image flow, the label lifecycle, spend caps.
Every hivectl command, grouped: fleet, driver, project, local cluster, infra setup.
The four worker modes plus qa — what each does, and the DoD / Plan / template conventions they share.
Image pulls, LoadBalancer on Mac, cost caps, missing DoD/Plan, recreating the cluster.
npm i -g @anupam/hive # installs the `hivectl` command
# one-time, per machine
hivectl bootstrap # doctor → local-cluster-up → setup
# one-time, per project
cd ~/my-app
hivectl init # scaffold .hive/config.yaml
hivectl labels sync # push status:*/type:* labels into the repo
hivectl agent-setup # build the pod image + create the creds secret
# fire the fleet
hivectl fire 42 --type=feature-implementation
hivectl tail 42 # live-tail the pod
hivectl merge 42 # merge issue #42's reviewed PR
hivectl driver # hands-off orchestration loop