Skip to content

Rename verification

Verification of the install and upgrade paths after the install-safety follow-up merged in PR #54. The checkout under test is dad280a, release v0.21.1; the last release before the rename is v0.20.2.

Published module

From an empty module cache and an empty build cache, go install github.com/urmzd/humagen/cmd/humagen@v0.21.1 passed on macOS arm64 with Go 1.27.1. go version -m identifies the installed module as github.com/urmzd/humagen v0.21.1. Running humagen version succeeded and printed dev (none, unknown): a plain Go install does not supply the release linker flags. Installation did not require a checkout or local replacements.

Container

The unmodified root Dockerfile built successfully for Linux arm64 with its default Go 1.26.10 and Node 24 base images. The resulting image is humagen:codex-verification, digest sha256:e0c912a52157601f87a28e342eef973c0a66af8cd126379cc7909594d8e2be00. Its configured user is 10001:10001.

Both docker run --rm humagen:codex-verification version and the same image with --entrypoint /usr/local/bin/warp exited successfully. The second invocation printed the expected compatibility note:

humagen: the command is humagen now; warp still runs it

Both printed dev (none, unknown) for the version because this build used the Dockerfile's default build arguments. Initial downloads were slow and pnpm retried several tarballs, but the build completed without source changes.

Slack

pnpm install --frozen-lockfile and pnpm validate:slack passed with Node 24.15.0 and pnpm 11.6.0, with network access to Slack's public blocks.validate endpoint. Every sandbox fixture was accepted, including the three parts of longTurn (Slack returned warnings for those parts).

The app's live payload validation passed all 74 tests, with zero failures and zero skips. This checks Block Kit acceptance; it does not install an app into a Slack workspace or exercise an actual Socket Mode connection.

Kubernetes upgrade

The documented Helm upgrade passed on kind 0.33.0 / Kubernetes 1.37 with Helm 4.3.0, using the actual v0.20.2 chart and Go binary for the baseline and the new root Dockerfile image for the upgrade. The default Postgres store stayed configured through the old WARP_POSTGRES_URL Secret key.

Revision 2 retained the Deployment UID, PVC UID and marker contents, Secret UID and bytes, and container name warp. The new pod became ready and /healthz returned ok. This verifies an in-place Deployment update rather than a reinstall. The script records the before/after JSON and logs under its VERIFY_ARTIFACTS directory.

The separate Flux upgrade passed with Flux 2.9.6, helm-controller 1.6.5 and source-controller 1.9.6. Its GitRepository first fetched v0.20.2, then commit 39f7f48; the HelmRelease switched from ./deploy/helm/warp to ./deploy/helm/humagen with chartNameChangeStrategy: InPlaceUpdate and nameOverride: warp. The Deployment, PVC and Secret UIDs remained the same, the Secret bytes and volume marker survived, the container stayed named warp, and the new pod became ready with /healthz returning ok.

The original checks used the default Postgres slot. The memory-store regression below was subsequently fixed and both store variants were checked again. An earlier attempt was interrupted when the host's Colima VM stopped; the completed checks above were rerun after the existing VM and isolated cluster recovered.

Memory-store reuse-values regression

Issue #62 records a live Helm 4.3.0 failure for an existing memory-store release. The old release is healthy with an override containing stores.kind: memory and stores.postgres: null. Its rendered spec omits the null key. An upgrade with --reuse-values preserves postgres: null in the new spec, and the pod exits with:

humagen: /etc/warp/humagen.yaml: document 1 (System warp): spec: unknown field "postgres"

The failing atomic upgrade rolled back. A regular helm template render drops the null and therefore did not expose this live upgrade behavior.

The chart now copies the System values and omits only null store entries. Configured inactive blocks, including empty maps and false, still reach the unchanged strict spec decoder and are refused. Postgres without a database reference is still refused. The regression injects retained values after Helm's normal coalescing and covers both store kinds and custom null blocks; zero execution limits, false settings and empty lists stay intact. CI runs these checks through make chart-check with Helm 4.3.0.

On October 25, 2026, both the default Postgres and memory variants passed a fresh kind upgrade from the actual v0.20.2 chart and binary to the fixed chart with the current v0.25.1 Go source. Each used the documented --reuse-values, nameOverride and image flags. No extra store override was passed at upgrade. The Deployment, PVC and Secret UIDs, Secret bytes, volume marker and container name stayed unchanged, and both pods became ready with /healthz returning ok. The memory release's stored Helm values still contain postgres: null, while its generated spec contains only stores: {kind: memory}. The runtime was built from this branch and layered onto the verified Alpine image; console assets were unchanged. Helm 3 and old-to-old upgrades have not been checked.

Reproduction

Use an empty directory for the module and build caches when checking the published module:

verify_dir=$(mktemp -d)
GOMODCACHE="$verify_dir/mod" GOCACHE="$verify_dir/cache" \
  GOBIN="$verify_dir/bin" \
  go install github.com/urmzd/humagen/cmd/humagen@v0.21.1
"$verify_dir/bin/humagen" version

Build and check both command names from the repository root:

docker build -t humagen:codex-verification .
docker run --rm humagen:codex-verification version
docker run --rm --entrypoint /usr/local/bin/warp \
  humagen:codex-verification version

The isolated cluster upgrade checks are in scripts/verify-upgrade.sh. They need kind, kubectl, Helm, Flux 2.9 or later, Python 3 and yq 4. The script uses its own kubeconfig, test namespaces and dummy Secret, and compares the Deployment, claim and Secret UIDs, Secret bytes, a volume marker, container name and readiness across both upgrades. It deletes only a cluster it created itself, unless KEEP_CLUSTER=1 is set.

The old runtime tested here uses the real v0.20.2 Go binary on the new image's shared Alpine runtime and console assets. No published container image exists for either release. To reproduce that old runtime on arm64:

git worktree add --detach .worktrees/install-pre-rename v0.20.2
old_image_dir=$(mktemp -d)
(
  cd .worktrees/install-pre-rename
  CGO_ENABLED=0 GOOS=linux GOARCH=arm64 \
    go build -trimpath -o "$old_image_dir/warp" ./cmd/warp
)
cat > "$old_image_dir/Dockerfile" <<'DOCKERFILE'
FROM humagen:codex-verification
USER root
RUN rm -f /usr/local/bin/warp
COPY warp /usr/local/bin/warp
USER 10001:10001
ENTRYPOINT ["/usr/local/bin/warp"]
DOCKERFILE
docker build -t humagen:codex-pre-rename "$old_image_dir"
docker pull postgres:16-alpine
OLD_CHECKOUT="$PWD/.worktrees/install-pre-rename" \
  OLD_IMAGE=humagen:codex-pre-rename NEW_IMAGE=humagen:codex-verification \
  scripts/verify-upgrade.sh

The fixture keeps the chart's default Postgres store, with an isolated Postgres 16 Deployment and Service in each test namespace and an old-name database reference in the preserved Secret. Its test-only OIDC client uses Google's public discovery issuer. No sign-in or real credentials are involved. Both rendered default specs are checked by their actual runtime images before installation. The script also installs a separate memory release and reuses its null backend values through an upgrade. Set VERIFY_FLUX=0 to check both local chart upgrades before its branch is available to the Flux source.