Repository navigation
Release the run worker and factory images on a version tag - #424
Conversation
A v* tag should publish the run worker and the five run-unit images like the four services: versioned, multi-arch, signed, with an SBOM. Records the catalogue handoff for the agent images and the gate fixes. Part of #391
release.yml now publishes all ten images docker.yml publishes as :edge, multi-arch, signed and with an SBOM. The agent images read their model catalogue from their own build scripts on amd64 first. Every apk-based image gets APK_UPGRADE_BUST. The gate's .env gains the run worker's required OIDC secret, and the gate renders the factory profile so a missing variable fails it. ApkUpgradeIsNotCachedTest reads both workflows, and a new check fails when release.yml and docker.yml publish different images. Part of #391
SPIRE_VERSION may now be a release version for every image, so the "sha-<short> only" caveats go. Records the release half in CICD-AND-PACKAGING section 11, the 2026-11-14/15 live pull of four factory images in HISTORY, and what stays unproven in UNVERIFIED. Part of #391
artyomsv
left a comment
There was a problem hiding this comment.
First review at ad424d9. The release matrix mirrors docker.yml (same fields, same Dockerfiles), the agent images read their model catalogue from their own build scripts before setup-buildx, every action is pinned by SHA, a manual run stays a dry run with no switch to publish, cosign signs each image by digest, and the new arch check fails when the two workflows publish different images or build one from a different Dockerfile. The test classes are within the limits (146, 190, 61 lines; misdeclaredImages is 27 lines). PR CI is green. In the dry run (43657172281) the gate and 7 of 10 builds had passed when I looked, including both agent images with their catalogue reads; the three JVM services were still running under QEMU. I found no defect. Two LOW notes about what the documents promise before a real tag has run.
Findings: 2 LOW. Reviewed head ad424d9832 as @artyomsv.
artyomsv
left a comment
There was a problem hiding this comment.
✅ APPROVED
Posted as a comment: GitHub does not let an account approve or request changes on its own pull request. The label carries the verdict.
Agent findings: 0 resolved, 2 open. CI: 13/14 checks passing. Head ad424d9832, reviewed by @artyomsv.
Open findings:
- [docs/L-1] LOW — #424 (comment)
- [docs/L-2] LOW — #424 (comment)
Dry run 43657172281 built every image for amd64 and arm64 and passed the gate. A real tag that fails one image is recovered by re-running the failed jobs; a version is usable once its GitHub release exists. Part of #391
artyomsv
left a comment
There was a problem hiding this comment.
✅ APPROVED
Posted as a comment: GitHub does not let an account approve or request changes on its own pull request. The label carries the verdict.
Agent findings: 2 resolved, 0 open. CI: 13/14 checks passing. Head badc54cb79, reviewed by @artyomsv.
code-spire-beaver
left a comment
There was a problem hiding this comment.
Reviewed badc54c against 66dd409 across security, correctness, repository rules and tests, including the final documentation-only update. The release matrix matches the ten edge images, the distribution/catalogue preparation follows the existing build recipes, and the new guards detect the pre-change omissions. Release dry run 43657172281 passed at ad424d9, whose workflow and build inputs are unchanged at the reviewed head. No actionable findings in the changed code.
Findings: none. Reviewed head badc54cb79 as @code-spire-beaver.
code-spire-beaver
left a comment
There was a problem hiding this comment.
✅ APPROVED
Reviewed head badc54cb795e84fe8502a70527ab570f34b08846 against base 66dd409f0c35e303d73eb655a1ab5c46fb98fd77 (global/release-factory-images), including the final documentation-only delta from ad424d9. No actionable findings in the changed code.
- Security: checked the tag-only push/sign boundary, synthetic gate credentials, catalogue output and build arguments, pinned actions and package contexts. No tenant/authentication path changes. CI Semgrep's differential scan ran 85 rules on 9 files with zero findings; the full scan ran 442 rules on 1,730 files with zero findings. Gitleaks and CodeQL passed.
- Correctness: checked all ten entries against docker.yml, the SERVICE/context/Dockerfile mappings, installDist preparation, catalogue extraction before setup-buildx and transfer to both architectures, compose interpolation and the publish/release dependency chain. Read the changed Java signatures and all callers.
- Rules: checked root CLAUDE.md, documentation consistency, method/class/parameter limits and the four commit messages. Applied the user's commit conventions over the generic conventional-commit template. No nested CLAUDE.md, project rule files, matching prior review state or prescribed memory files were present.
- Tests: reasoned through removal of the release entries and APK_UPGRADE_BUST argument; the new parity and per-workflow checks fail on those pre-change omissions. Existing parser checks reject an absent include block, and the per-image guard rejects an empty parsed matrix. No production application code changed.
CI run 43660975170 is green at badc54c: fast tests, service tests and assemble, and dashboard type checking/tests/build. Dashboard: 127 files and 1,342 tests passed. The Gradle logs do not publish test counts, so no Java test count is claimed. The changed spire-arch tests also executed rather than coming from cache in the initial run 43657168378 at ad424d9. Manifest invariants, kind installation, the manifest gate, CodeQL, Semgrep and gitleaks passed on the latest head. No local builds, tests or linters were run.
Applicable global rule files: clean-code-java.md, clean-code-general.md, observability-and-logging.md, secrets-and-env-handling.md, documentation-maintenance.md, tech-debt-tracking.md, blog-material-capture.md, reference-pattern-adoption.md, git-workflow.md, local-environment.md, local-port-ranges.md, no-synthetic-data.md, no-local-builds.md and pr-review-labels.md.
Release dry run 43657172281 passed at ad424d9: the gate, all ten amd64/arm64 image builds, both catalogue preparations, chart packaging, draft release creation and draft deletion. Its workflow, tests and build sources are unchanged by badc54c; the final delta records this result and describes recovery from a partially published tag.
Residual limits: a manual dry run does not push or sign images. It does not demonstrate a live arm64 deployment or a packaged factory review, and the gate starts only the reviewer stack while rendering the factory profile. Those operational gaps remain documented in UNVERIFIED.md.
Agent findings: 2 resolved, 0 open. CI: 13/14 checks passing. Head badc54cb79, reviewed by @code-spire-beaver.
What
release.ymlnow publishes the same 10 images asdocker.yml: the 4 services, the run worker, the publisher, the model proxy, the egress proxy, and the Codex and Claude Code agents. Each image gets::<version>and:latest;all behind the existing e2e round-trip gate. The agent images read their model catalogue on amd64 with their own build script, then pass it to the multi-arch build.
Gaps fixed on the way:
.envlackedSPIRE_OIDC_RUN_WORKER_SECRET, whichcompose.ghcr.ymlrequires. The gate now also renders the compose file with--profile factory.release.ymlpassed noAPK_UPGRADE_BUST.ApkUpgradeIsNotCachedTestnow checks both workflows, and a new check fails if the two workflows publish different images.The docs say
SPIRE_VERSIONcan be a release for every image. They also record the GHCR live proof: on the dev stack,spire-test#106 was reviewed with the Codex agent, publisher, model proxy and egress proxy all pulled from GHCR atsha-66dd409.Design:
docs/superpowers/specs/2026-11-15-release-factory-images-design.md.Verification
CI on this PR. Then a
workflow_dispatchdry run ofrelease.ymlon this branch, which builds all 10 images for both architectures and runs the gate. It pushes nothing.Part of #391
🤖 Generated with Claude Code