CI / continuous validation¶

This repo has six GitHub Actions workflows that validate changes, publish the documentation site, track upstream images, and release signed chart artifacts.

Workflow	Trigger	Role
validate-chart.yaml	Pull requests that change the chart, tests, validation scripts, or this workflow	Lint, generated-docs drift checks, and isolated kind install/test scenarios before merge.
deploy-docs.yaml	Pull requests and `main` pushes touching anything the site renders (`docs/`, the READMEs, `charts/`, `examples/`, `mkdocs.yml`, ...), manual runs, and release refreshes	Build the site with strict link checks and deploy it to GitHub Pages outside pull requests.
cron-fetch-image.yaml	Every 6 hours or manually	Detect a newer upstream image, open an appVersion update PR, and create upstream-review issues.
propose-release.yaml	Manual	Consume pending Changesets into one reviewable release PR without publishing.
release-chart.yaml	Push to `main` that changes `charts/hermes-agent/Chart.yaml`	Tag `vX.Y.Z`, publish OCI and Helm Repository artifacts, cosign-sign OCI, and refresh Pages.
verify-release.yaml	After a successful `release-chart` run	Re-verifies the published, signed artifact end to end.

How the workflows connect¶

flowchart LR
    C["cron-fetch-image<br/>every 6 hours"] --> U["appVersion update PR"]
    C --> I["upstream-review issues"]
    P["Chart-related PR"] --> V["validate-chart"]
    P --> B["deploy-docs<br/>build only"]
    S["Site-only PR"] --> SB["deploy-docs<br/>build only"]
    U --> V
    U --> B
    V --> M["Merge chart change"]
    B --> M
    SB --> SM["Merge site change"]
    M --> R["propose-release<br/>manual"]
    R --> RP["Release PR"]
    RP --> RV["Review and validation"]
    RV --> RM["Merge release PR"]
    RM --> RC["release-chart"]
    RC --> VR["verify-release"]
    RC --> D["deploy-docs<br/>publish"]
    M --> D
    SM --> D

The scheduled path stops at a normal pull request and review issues. It never publishes a chart directly. Publishing starts only after a maintainer prepares, reviews, and merges the Changesets release PR.

validate-chart¶

Two jobs run on every functional change:

`lint`¶

helm lint, helm template, render-only assertions for the dashboard block (.github/scripts/assert-dashboard-render.sh), and a helm-docs drift check - if charts/hermes-agent/README.md is out of date relative to README.md.gotmpl, the job fails. Always run make docs after editing values.yaml and commit the result.

`test`¶

Seven scenarios run as a matrix, each on its own ephemeral kind cluster (a separate runner) - fully isolated, with native per-job status, timeout, and failure diagnostics instead of one bundled log. The PR checks list shows them separately: test (message), test (existing-claim), test (team), test (security-hardened), test (bootstrap-overwrite), test (config-migration), and test (dashboard-ingress). Scenario logic lives in .github/scripts (lib.sh + one script per scenario) rather than inline in the workflow.

A changes job skips test entirely for version-bump-only commits (where the chart behavior is unchanged).

message scenario: scenario-message.sh ¶

Install with chart-managed storage.
Run the chart's hermes doctor test hook (the same Job as helm test, but invoked directly so it can't stall on the hook watch).
Only on trusted runs (an NVIDIA_API_KEY secret is present): inject a skill onto the PVC, then do one hermes chat round-trip through NVIDIA NIM.

The CI_MODELS pool is failover only - the round-trip passes on the first model that answers, not every model. The chat invocation mirrors the chart's own test hook: hermes chat -m <model> --provider nvidia -q <prompt> --max-turns N.

The live Discord notification step (workflow-level, after the scenario script) only runs for this scenario, since it's the only one with Discord enabled.

existingClaim scenario: scenario-existing-claim.sh ¶

Exercises persistence.existingClaim (the ability to mount a pre-existing PVC instead of one the chart creates - PR #37):

Create a ci-shared-pvc PVC outside the chart.
Install with --set persistence.existingClaim=ci-shared-pvc.
Confirm the hermes doctor test hook passes.
Exec into the pod and write then read ${HERMES_HOME}/ci-claim-probe.txt.
Re-run hermes doctor.

This is a smoke test: it proves the chart binds to a pre-created PVC and starts cleanly. Persistence across restarts/upgrades remains a follow-up.

The separate team scenario provisions hermes-team-knowledge as RWX, mounts it read-write on the leader and read-only on a member, writes a probe from the leader, cross-reads it from the member, and confirms that a member write is rejected. It therefore covers the multi-instance shared-knowledge boundary without treating the volume as a task handoff path.

The security-hardened scenario installs values-hardened.yaml in a namespace enforcing Pod Security Standards restricted. It verifies the rendered security contexts and runs hermes doctor in the admitted read-only-root-filesystem pod.

The bootstrap-overwrite scenario validates the durable config contract. It installs with bootstrap.overwrite=false, writes a marker into the running pod's config.yaml, and then upgrades with a harmless Pod annotation to force the seed init container to run again. The marker must still exist in the new pod, proving that runtime edits survive an upgrade.

The config-migration scenario uses Hermes v2026.9.21 and a persistent PVC. It checks first boot, replacement with bootstrap.overwrite=true, preservation and migration with false, a safe refusal for an explicitly unsupported version, and the documented operator migration-floor step. It verifies the resulting schema version and the pre-migration backup.

The dashboard-ingress scenario installs a pinned ingress-nginx controller in its own kind cluster and reaches the dashboard through a TLS Ingress. The pod must turn Ready only once the dashboard listens (the readiness probe), so the Ingress answers right after helm --wait; the auth gate must reject unauthenticated and wrong-password requests; and sign-in must set Secure, HttpOnly __Host- session cookies because the controller pod is listed in dashboard.trustedProxies. A control upgrade without a trusted proxy checks that the cookies are not Secure, which is the documented reason to set it. The host is a reserved .test name resolved with curl --resolve and the certificate is self-signed, so no DNS or public certificate is involved; never use a *.localhost host here, because upstream treats loopback hosts as a development setup and does not set Secure for them. Real DNS and Let's Encrypt issuance are verified by hand on a real cluster.

Fork PRs¶

Fork PRs don't receive repository secrets, so the chat round-trip (and any live Discord check) is skipped and the run falls back to doctor-only, which is safe and still meaningful.

deploy-docs¶

Its path filter covers every repository source rendered by the site, which is broader than docs/: the root README, SECURITY, and CONTRIBUTING files in both supported languages; charts/; examples/; mkdocs.yml; main.py; hooks.py; and requirements.txt are all included. A chart-only pull request therefore runs this workflow too, because the chart README is part of the site.

Pull requests run mkdocs build --strict and upload the site artifact but do not deploy it. A matching change on main, a manual run, or a refresh requested by release-chart deploys the built site through the GitHub Pages Actions API.

The workflow merges index.yaml and packaged chart archives from the gh-pages branch into the site output. The branch is a Helm Repository data store; deploy-docs is the only workflow that publishes the Pages site.

cron-fetch-image¶

The schedule 0 */6 * * * runs at 00:00, 06:00, 12:00, and 18:00 UTC. A manual dispatch runs the same process.

Fetch date-based nousresearch/hermes-agent tags from Docker Hub.
Compare the newest tag with the chart's current appVersion.
If a newer image exists, open an appVersion bump PR with a minor Changeset.
Independently review intervening upstream release notes with NVIDIA NIM and create one labeled issue for each chart-relevant follow-up.

The version bump and upstream-review jobs are independent. Maintainers can merge the image update after its normal checks without waiting for every follow-up issue to be implemented.

propose-release¶

A maintainer starts this workflow when pending Changesets are ready. It combines them into one release PR and synchronizes the release manifest, Chart.yaml, Artifact Hub annotations, generated chart docs, and versioned examples. It does not publish a chart.

release-chart¶

Merging the release PR changes the chart version on main and starts this workflow. If the version tag does not already exist, it:

creates the vX.Y.Z tag and GitHub Release;
packages and publishes the chart to OCI;
signs the OCI artifact with keyless cosign;
updates the Helm Repository data on gh-pages; and
triggers deploy-docs so the documentation site and Helm Repository index are served together.

verify-release¶

After a release is published, this workflow proves the whole supply chain against the artifact users actually pull (namespace verify-hermes-chart):

cosign verify the OCI artifact against this repo's Actions OIDC identity.
helm install from the OCI registry (not local source).
Run the same hermes doctor + chat round-trip as validate-chart.

A failure here means the published artifact or its signature is broken.

Running the equivalents locally¶

make lint        # helm lint
make template    # render manifests
make docs        # regenerate the chart README (helm-docs) - commit the result
make test        # install + helm test (needs a cluster/kind)

See Local development guide for setting up a local kind cluster and a Discord + NVIDIA dev loop.