What problem does it solve? Moving parallel work off a local machine onto a Kubernetes cluster is error-prone: jobs get lost, retries re-run finished work, and results scatter across pods that disappear. This Skill turns a list of independent tasks into a wave of Kubernetes Jobs on the envision-compute GKE cluster, records every result in Google Cloud Storage, and makes re-runs idempotent so completed tasks never execute twice. ## Core Features & Use Cases - Two task kinds: generic container commands run as one Indexed Job, or Claude Code plan executions run as per-task Jobs in the claude-executor image with budget caps and branch pushes. - Idempotent replay: re-dispatching a wave keyed by wave id adopts prior exit-0 results, archives stale failures, and only schedules pending tasks. - Roadmap execution: run_roadmap.py converts a GSD planning directory into dependency-batched phases with checkpointed state in GCS and resume support. - Use Case: Lint two directories of a repository in parallel on the cluster instead of a laptop: normalize the two tasks into a wave manifest, dry-run to inspect the Job YAML, dispatch, then collect per-task exit codes and logs from GCS. ## Quick Start Ask the assistant to dispatch a wave of independent shell commands or Claude Code plans to the GKE cluster by first building a manifest with normalize_wave.py, then running dispatch.py and collect.py from the skill's scripts directory.
