Dockerized Service Release Deployment Workflow

Dockerized Service Release and Deployment Workflow

Purpose

Create or audit a durable release contract for a Dockerized backend or cloud service without treating a developer worktree or a production host as a build machine.

The default contract is:

1. Feature worktrees are development-only and may run native source tests, but never Linux artifact or image builds.
2. A clean GitHub Actions checkout of a version tag validates the release commit and builds the published OCI image.
3. The tag must resolve to a commit reachable from origin/main.
4. The registry digest, tag, commit, image reference, build run, and provenance/SBOM choices are written into a release manifest.
5. The workflow publishes a GitHub Release with that manifest attached: normal SemVer tags become normal releases, and recognized prerelease tags become GitHub prereleases.
6. Publishing the GitHub Release triggers deployment: normal releases target protected production; enabled prereleases target test.
7. The deployment job invokes a provider-specific adapter with image@sha256:... only after the selected environment permits it.
8. The adapter verifies service health and records the prior manifest/digest so rollback deploys a previous exact digest rather than rebuilding source.

This is a durable building-block change: it removes the ambiguity between a developer build, a CI artifact, and a deployed artifact. A project can later change registries or providers without weakening the tag-to-digest deployment contract.

When To Use

• Use this skill when adding or reviewing Dockerized-service release automation, an OCI registry publication path, GitHub Release deployment triggers, production approval gates, image-digest deployment, release manifests, health checks, or rollback guidance.
• Use it for backend services and cloud workloads regardless of application language when Docker or OCI images are the release artifact.
• Use server-side-swift:docker-workflow for Swift Dockerfile, Compose, Linux runtime, and image-entrypoint work.
• Use the official provider plugin or provider documentation for the final deployment adapter. This skill does not create cloud resources, configure a cloud account, or guess a provider command.

Source Check

Inspect the repository's Dockerfile, CI workflows, release process, registry settings, deployment files, health endpoint, provider configuration, and rollback procedures first. Then verify the current platform behavior against these official sources:

• GitHub Actions environments
• GitHub Actions OIDC
• GitHub Releases
• GitHub Actions release-event triggers
• GitHub Actions token-trigger behavior
• GitHub Container registry
• Docker GitHub Actions image publishing
• Docker BuildKit attestations

Translate those sources into the actual project choices: registry, tag policy, environment policy, provider identity, health URL, and rollback command.

Build Ownership Preflight

Before starting an image build, record the deployment target's architecture, operating system, available memory, disk capacity, container runtime, and any provider constraints. Select the OCI platform from that evidence. For example, an Apple-silicon developer machine targeting an x86-64 Linux host builds linux/amd64; do not let the developer machine's architecture choose the release artifact by accident.

Build every cloud Linux archive and OCI image in a clean GitHub Actions checkout. Developer machines and production hosts must not build, cross-compile, or smoke-test cloud images. Record OCI artifacts by registry digest and record archive artifacts by SHA-256 checksum in the release manifest; deployment must consume that exact recorded identity. A production host may only pull or load a finished immutable artifact and run it. Do not clone application source, resolve dependencies, compile Swift, or run docker build on a production host.

One build owns its Docker client session until it exits. Preserve the original progress-producing shell or durable log stream; do not replace it with blind polling. When an agent must return for a build, GitHub Action, deployment approval, or health gate, it records the immutable identity (tag, digest, run, environment, and target), reuses a live matching host-native continuation while the gate remains pending and healthy, and performs one fresh inspection on wakeup. Continue only if every recorded identity equals the fresh tag, digest, run, environment, and target; on any mismatch, stop and create a new continuation packet rather than build, deploy, diagnose, roll back, or otherwise advance. Do not delete/recreate the continuation after an unchanged snapshot; create/update only after it fires or becomes stale, with at least five minutes between rechecks. Codex uses heartbeat; Hermes uses a continuable cronjob with deliver="origin" and attach_to_session=true. Pause/delete it when the gate resolves, fails, is cancelled, or its identity changes. Do not leave a shell waiting or create a one-to-four-minute polling loop. While that build is active, do not run Docker status, image-inspection, build-history, Buildx, Compose, or second-build commands unless the runtime explicitly documents concurrent client access as safe. If an orchestration wrapper returns before its child exits, inspect the real process rather than trusting the wrapper result, report that the build is still active, and wait before starting another package-manager or container command.

Hermes Compatibility

This is portable instruction-only guidance and is exported through Socket's Hermes skill tap. It does not install a GitHub App, configure registry credentials, provision a cloud account, or bundle a deployment adapter for Hermes or Codex.

Required Decisions Before Editing

Confirm these decisions. Stop rather than silently choosing a production target:

• registry and image name
• supported target platform or platforms
• tag pattern and the protected integration branch, normally main
• native repository validation command and GitHub Actions image smoke-test command
• GitHub App with only the repository contents: write permission; store its client ID as RELEASE_PUBLISH_APP_CLIENT_ID and private key as RELEASE_PUBLISH_APP_PRIVATE_KEY
• registry-specific OCI_REGISTRY_USERNAME repository variable and OCI_REGISTRY_PASSWORD secret, unless the adopting repository deliberately narrows the template to GHCR
• GitHub production environment reviewers, permitted stable tags, and bypass policy
• whether a test environment exists; when it does, enable ENABLE_PRERELEASE_DEPLOYMENTS=true as a repository variable and configure its test-tag policy
• cloud account, project, region, workload, and OIDC trust policy for the deployment adapter
• health endpoint, expected response, timeout, and failure diagnostics
• source of the prior successful release manifest for rollback

Template Set

Copy these template assets into the target repository and replace every {{...}} placeholder before enabling them:

• release-container.yml.tmpl: tag-anchored validation, immutable image publication, manifest creation, and automated normal-release or prerelease publication.
• deploy-production.yml.tmpl: a GitHub Release published trigger, protected production or enabled test environment selection, manifest download, exact-digest adapter call, health verification, and failure guidance.
• release-manifest.json.tmpl: the release-record schema.
• deploy-production-image.sh.tmpl: a deliberately blocked provider-adapter seam.
• verify-production-health.sh.tmpl: a deliberately blocked health-check seam.

The shell templates fail closed until a project replaces their placeholders. Do not treat the template itself as a deploy command.

Release Workflow

1. Create feature work in a branch-backed worktree. Use native source checks for feedback only; do not build or publish a Linux release image from that worktree.
2. Open and merge the reviewed change into main through the repository's normal CI and review gates.
3. Create an annotated SemVer tag from the reviewed main commit and push it.
4. The tag workflow checks out the tag in a clean hosted runner, fetches origin/main, and rejects a tag commit that is not reachable from it.
5. Run the repository validation from that clean checkout.
6. Build and push the image once. Capture the registry output digest, then run the container smoke test against that exact published digest. Publish provenance and SBOM attestations where the chosen registry supports them.
7. Accept only vMAJOR.MINOR.PATCH stable tags or the recognized prerelease forms -alpha, -beta, -rc, and -test, each with an optional dot/hyphen suffix. Reject all other tag forms before publishing.
8. Create release-manifest.json from the tag, resolved commit, release kind, immutable image reference, workflow run URL, and attestations. Mint a short-lived GitHub App installation token in the job, then publish a normal GitHub Release for a stable tag or a GitHub prerelease for a recognized prerelease tag.

The automated publisher needs a GitHub App installation token because a release created with the workflow's GITHUB_TOKEN does not trigger the release-published deployment workflow. Configure the App with only the required repository access and mint its token during the job rather than storing a reusable PAT.

The deployment workflow must listen only for release: [published]. It must not deploy directly on a tag push, a push to main, or a pull request.

Production Deployment and Rollback

1. The release-published workflow downloads release-manifest.json from the exact release event.
2. It validates that the manifest tag matches the release event and that image is a fully qualified name@sha256:digest reference.
3. A stable release enters the GitHub production environment. Configure required reviewers, stable-tag restrictions, and no-administrator-bypass where the repository plan supports them. A recognized prerelease enters test only when the repository has explicitly set ENABLE_PRERELEASE_DEPLOYMENTS=true; otherwise it publishes without deployment.
4. Authenticate to the provider using short-lived OIDC credentials when supported. Do not put long-lived cloud credentials in repository secrets merely to run this workflow.
5. Pass only the exact digest reference to the provider deployment adapter. The adapter must not run docker build, clone source, or reinterpret a mutable tag.
6. Run the repository-defined health check against the deployed release. On failure, preserve deployment logs and stop; do not silently roll forward or rebuild.
7. Roll back by selecting the prior successful GitHub Release manifest and rerunning the same exact-digest deployment path after the required production approval. Do not rebuild an old commit on a host.
8. For an asynchronous provider rollout, record the release tag, immutable digest, environment, deployment/run URL, and health target; reuse a live matching host-native continuation while the gate is pending and healthy, inspect those identities on wakeup, and only then decide whether to advance, diagnose, or roll back. Continue only if the fresh release, digest, environment, deployment/run URL, and health target exactly match the record; on a mismatch, stop and create a new continuation packet. Do not delete/recreate it after an unchanged snapshot; create/update only after it fires or becomes stale, with at least five minutes between rechecks.

Security and Supply-Chain Rules

• Scope workflow permissions minimally. The release workflow needs contents: read and packages: write; its short-lived GitHub App token has only the separate contents: write authority needed to publish the release. The deploy workflow should have read-only contents plus id-token: write only when the provider adapter uses OIDC.
• Keep registry write access in the release workflow. Production deployment should normally need only pull/read access plus provider authorization.
• Do not place secrets in Docker build arguments, image layers, caches, release manifests, logs, or committed environment files.
• Treat restored CI caches as untrusted input; do not cache secrets or make a release depend on an unrebuildable cache.
• Use mutable version tags only as discovery aliases. The release manifest and deployment adapter must use the immutable digest.
• Require a real health signal. A TCP connection alone is not a readiness guarantee unless the service explicitly documents it as one.

Validation

Before enabling the workflow:

1. Validate YAML and shell syntax after template substitution.
2. Run the repository's native test suite locally when appropriate, then run the image build and smoke test only in GitHub Actions.
3. Use a recognized prerelease tag in a repository with test enabled to prove that the Actions runner publishes a GitHub prerelease, produces the manifest and registry digest, and deploys its exact digest to test.
4. Inspect the manifest against the registry and release tag after automated publication.
5. Confirm a stable tag creates a normal GitHub Release, pauses at production approval, and passes image@sha256:..., not a tag, to the deployment adapter.
6. Exercise the health-check failure path and a rollback drill using a known prior release digest before treating the workflow as release-ready.

Guardrails

• Do not deploy or modify a cloud account while creating this reusable guidance or template set.
• Do not use a worktree, developer laptop, or production host to build the image that a release deploys.
• Do not start an image build before recording the target platform and host-resource constraints; stop and correct an architecture mismatch, insufficient capacity, or interrupted owning session before a deployment mutation.
• Do not probe Docker concurrently with an active build, start a second build after an early-returning wrapper, or substitute a new shell session for the original progress-producing build owner.
• Do not publish a GitHub Release from a tag that is not anchored to the protected integration branch.
• Do not publish a release with the workflow GITHUB_TOKEN when it must trigger another workflow; use the dedicated release-publisher token.
• Do not deploy an unrecognized, draft, deleted, edited, or merely-created release. Stable releases deploy only to production; prereleases deploy only to enabled test.
• Do not deploy from latest, a branch name, a short SHA tag, or any other mutable image identifier.
• Do not substitute an SSH host copy, docker compose build, or remote source checkout for the provider deployment adapter.
• Do not claim rollback readiness until a prior release manifest and exact-digest adapter path have both been verified.

Output Shape

Return:

1. Release contract: branch, tag pattern, registry, image platforms, validation, manifest location, and GitHub Release ownership.
2. Deployment contract: provider adapter, OIDC/credential boundary, production environment policy, health signal, and rollback source.
3. Templates: copied assets and every required substitution.
4. Validation: local, CI, registry, release, environment-approval, health, and rollback evidence.
5. Risk: unresolved provider commands, approval policy, plan limitations, health semantics, or rollback gaps.