---

name: k8s-kaniko-job description: > Sandbox build executor for Sealos deploys: build a container image from source with an in-cluster Kaniko Job and push it to GHCR, for environments without a Docker daemon (Brain managed Devbox sandboxes). Use only when a managed deploy (SEALAI_DEPLOY_MODE=managed) needs an image built from project source; local interactive deploys build with docker buildx instead (see the use-sealos skill). allowed-tools: Bash(kubectl:), Bash(python3:), Bash(tar:), Bash(curl:), Bash(bash:*)

K8s Kaniko Job

Builds linux/amd64 images inside the Kubernetes namespace when no Docker daemon exists. One script owns the whole flow — do not hand-roll Jobs or Secrets around it:

python3 scripts/kaniko-build.py --image ghcr.io/<owner>/<repo>:<tag> \
  [--context <dir>] [--dockerfile <path-relative-to-context>] \
  [--build-arg KEY=value ...]

Resolve scripts/ against this skill's directory. Run the script from the project workspace so it finds .sealos/build-runtime.json.

How it works

tar the context (excludes .git/.sealos/.versitygw-*)
  → DevBox-local VersityGW S3 store (POSIX dir, served on port 1319)
  → Kaniko Job in the current namespace pulls s3://... via the
    Job-reachable endpoint from .sealos/build-runtime.json
  → pushes the tagged image to ghcr.io with a build-only registry Secret
  → digest captured from the pod termination message

Everything is resolved automatically, in order: CLI flags → .sealos/build-runtime.json (written by the control plane: Job-reachable s3Endpoint, S3 secretKeyRef, build deadline) → DevBox runtime env (KANIKO_CONTEXT_POSIX_DIR, S3_ENDPOINT, SEALOS_DEVBOX_JWT_SECRET, ...) → defaults. The Job runs with backoffLimit: 0, an active deadline capped at 1800s, ttlSecondsAfterFinished: 3600, and the current ServiceAccount.

Preconditions

• kubectl pointed at the sandbox namespace (injected kubeconfig; the script never selects a region, workspace, or other namespace).
• GITHUB_TOKEN with the write:packages scope. The target image owner must equal the token's login, lowercased: ghcr.io/<login>/<repo>:<tag>. Tag with the commit SHA or a timestamp, never latest.
• Namespace permissions to create Jobs/Secrets and read Pods/logs.
• A Dockerfile inside the context (write one first if missing — rules in ../use-sealos/references/build.md §1; the buildx/registry sections of that file do not apply here).

Reading the result

On success stdout is JSON with digest and image_ref (ghcr.io/<owner>/<repo>@sha256:...). Prefer image_ref in deployment manifests — the digest pin survives tag mutation. If digest is null (rare: termination message lost), fall back to the tag reference in image.

pull reports downstream pull behavior:

• anonymous — the package is public; no pull secret needed.

• private — GHCR packages are private by default. Create a pull secret in the namespace and reference it from the workload (imagePullSecrets: [{name: <app>-pull}], fixed literal name):

  kubectl create secret docker-registry <app>-pull \
    --docker-server=ghcr.io --docker-username=<login> \
    --docker-password="$GITHUB_TOKEN" \
    --dry-run=client -o yaml | kubectl apply -f -
  

• indeterminate — verify by watching the first pod pull, and add the pull secret if it hits ErrImagePull.

Failure triage

The script reports the Job and namespace on failure without echoing raw Pod logs or Kubernetes error text, which may contain credentials. Inspect the named Job and Pods directly, then summarize only redacted findings.

Symptom	Cause → fix
Job pod ImagePullBackOff on the executor image	cluster cannot pull gcr.io/kaniko-project/executor → report; there is no local fallback
Kaniko log: error uploading context / S3 connection refused	the Job cannot reach the VersityGW endpoint → check .sealos/build-runtime.json.s3Endpoint; never point the Job at 127.0.0.1
Kaniko log: 401/403 on push	token scope or owner mismatch → the script validates both before creating the Job; inspect the token and image owner privately
Dockerfile build error	fix the Dockerfile in the workspace and rerun; each run creates a fresh Job
Job deadline exceeded	build too slow → trim the context (.dockerignore), use smaller base images

Build-only Secrets (use-sealos-ghcr-auth-*) are namespace-local with no TTL; the Job itself is garbage-collected after an hour. Leave cleanup to sandbox teardown unless the user asks.

Never print GITHUB_TOKEN, S3 credentials, or Secret payloads. Never pass secrets through --build-arg — build args are visible in the Job spec. --render-only prints a synthetic, redacted Job manifest for structural inspection; it does not reveal values from the runtime contract or build args.