コンテンツにスキップ

DevOps Roadmap

This document tracks CI/CD pipeline improvements for hermes-agent-helm. Items are derived from a source-control and DevOps review of the current .github/workflows/ configuration.

Grouped by area and ordered by priority within each group.

Source Control

🔴 High priority

  • [ ] Enable branch protection on main
  • Require at least one PR before merging (no direct push to main)
  • Add validate-chart as a required status check

  • Risk today: a broken commit can hit main and immediately trigger release-chart.yaml without any CI gate

  • [ ] Include workflow file changes in validate-chart trigger

  • Current path filter watches charts/hermes-agent/** and .github/workflows/validate-chart.yaml only
  • Changes to cron-fetch-image.yaml, release-chart.yaml, propose-release.yaml, etc. land on main unvalidated
  • Add .github/workflows/** to the paths: trigger (or at minimum lint YAML with actionlint)

🟡 Medium priority

  • [ ] Resolve the unused dev branch strategy
  • validate-chart triggers on [dev, main] but all work goes directly to main; the dev branch has never been used in practice
  • Either adopt a dev → main merge flow, or remove dev from trigger lists and clean up CONTRIBUTING.md so the documented strategy matches reality

Pipeline Design

🟡 Medium priority

  • [ ] Split release-chart into independent OCI and gh-pages jobs
  • Currently a single job runs OCI push → gh-pages deploy in sequence
  • If OCI push succeeds but gh-pages fails, re-running skips everything (tag-existence guard) - partial deployment with no automatic recovery

  • Splitting into two jobs (with a shared needs: on a tagging step) allows each to be re-run independently

  • [ ] Close the "release commit is not tested" gap

  • Changesets release PRs touch only generated version metadata, Chart.yaml, docs, and CHANGELOG.md, so functional=false → kind cluster test is skipped
  • The last merged commit before shipping is never integration-tested
  • Options: (a) add a lightweight smoke test that always runs on a release PR, or (b) document the gap and accept it as a known trade-off given that the preceding feature commit was tested

🟢 Low priority

  • [ ] Abstract the runner label into a repository variable
  • All workflows hard-code ubuntu-26.04-arm; if that image is deprecated or has an outage, every workflow fails simultaneously
  • Use runs-on: ${{ vars.RUNNER_LABEL || 'ubuntu-latest' }} so the fallback can be changed without a code edit
  • See fallback design note below

Security

🟢 Low priority

  • [ ] Add Helm template security scanning
  • No kubesec, trivy, checkov, or kube-score scan on rendered templates today
  • Suggested insertion point: new step in validate-chart / lint job, running helm template | trivy config - or kubesec scan -

  • Cosign signing proves provenance but does not validate what was signed

  • [ ] Generate and attach SBOM to each release

  • release-chart.yaml signs the OCI artifact but produces no Software Bill of Materials
  • syft or cosign attest --predicate (CycloneDX/SPDX) can be added as a post-push step

Operations

🟢 Low priority

  • [ ] Prune old chart packages from gh-pages
  • release-chart.yaml uses keep_files: true; every release accumulates a new .tgz on the gh-pages branch indefinitely

  • Add a cleanup step that retains the last N versions (e.g., 10) and removes older .tgz files before regenerating index.yaml

  • [ ] Harden the AI advisor fallback in upstream-review

  • If the NVIDIA NIM endpoint is down or quota is exhausted, the upstream-review job silently fails or produces no GitHub issue
  • Add explicit continue-on-error: true with a step that posts a fallback comment/issue when the AI call fails, so the failure is visible

Notes

Runner fallback note

GitHub Actions does not natively support "try runner A, fall back to runner B." The runs-on: [label-a, label-b] array syntax means all labels must match, not first available. The repository-variable approach is the lowest-overhead workaround: a single settings change propagates to all workflows instantly.

Functional diff filter design

The changes job in validate-chart deliberately skips the kind cluster test for release-only commits (version bump + docs). This is an intentional cost/speed trade-off; the item above ("release commit is not tested") captures the known gap for future review, not necessarily for immediate fixing.

Last reviewed: 2026-06-29