Skip to content

speech-runtime runs in k3d and on the PC from its attested ref, on CPU, with its caller table as a ConfigMap #119

Closed

Description

@FlexOr2

speech-runtime runs in the namespace speech-runtime from its attested private GHCR ref, on CPU, with its caller table as a ConfigMap. It is applied by the Kustomization apps and proven in the job cluster on pull requests and on main, then on the PC. It has no route and no caller yet: hopin and presentator wire to it later, and it goes live with P7 (#116), which brings Keycloak.

Caller

Facts

speech-runtime's head, 07.11.2026, from speech-runtime's main 8adecbd and the image's own files:

  • The ref: ghcr.io/overnightworks/speech-runtime:2026.11.07.1@sha256:18d513fa334de8cb18065a8c6e50162e8d89323eeb20c47f398cfb0b23362f27. It is signed by quality's image.yml, has SLSA v1 provenance, and is private.
  • The user is 1001. The root is read-only; the only writable path is /tmp (tmpfs of 64 MiB in compose, where uploads over 1 MiB are spooled). The voices (/opt/voices) and the hearing model (/opt/hearing) are baked in and only read, so the service needs no egress apart from Keycloak's key set.
  • Port 8080, with SPEECH_HOST=0.0.0.0.
  • Readiness: GET /health answers 200 while the models still load, so readiness is compose's exec probe: both speaking.ready and hearing.ready true, start period 30 s.
  • Settings to set: SPEECH_CONFIG_FILE (the caller table's path), SPEECH_HOST=0.0.0.0, SPEECH_DEVICE=cpu, SPEECH_VOICE_CACHE=/opt/voices, SPEECH_HUGGINGFACE_CACHE=/opt/hearing, SPEECH_HEARING_MODEL=small. SPEECH_GPU_VOICES stays unset; Qwen and the card are P8's. The defaults' owner is speech-runtime's docs/wire.md.
  • No Secret of its own. The caller table is plain configuration, and the key set is public.
  • The caller table for this image:
    • [identity]: issuer, jwks_url (https), audience = "speech-service", workload_role = "speech-workload", group_claim = "groups", service_group = "/services/speech-service";
    • [apps.hopin] and [apps.presentator], each with workload_client, voices = ["piper"] and hear = true (the operator's ruling of 07.11.2026, speech-runtime#184).
    • The next image drops hear and voices, and a removed key stops the start, so the ConfigMap changes with the image tag in one pull request.

Decided by the head

  • The pull Secret. The namespace receives its own ghcr-pull from platform-secrets, like hopin's, so it joins the namespaces layer with restricted and the prune annotation. The operator puts it with one more secrets:put-registry for the namespace speech-runtime (O2), the path hopin's took. P4: Flux decrypts SOPS-age secrets from platform-secrets #13's P4d include, the credential kept once for every namespace, waits for its own slice, because a SOPS MAC over metadata.namespace may rule out a namespace transform before decryption. In the job, the artifact gets the same job-made ghcr-pull in this namespace too.
  • The issuer until P7 is https://auth.hallucinai.de/realms/hallucinai, as in hopin's manifest. P7 moves it.
  • A NetworkPolicy admits 8080 only from the namespaces hopin and presentator.
  • The image is large (about 9.8 GB). The job measures whether k3d on a standard runner pulls and starts it, and reports disk and time. A runner that cannot hold it is reported to the head before any workaround. When the measured pull and start exceed apps' timeout of 5 minutes, that timeout and the job's wait on the Kustomizations rise to the measurement with a margin, in clusters/pc/apps.yaml and the job.

Done when

  • D1 In the job, the Deployment speech-runtime pulls its attested ref through ghcr-pull and becomes Available with one Ready pod (both models ready) in a namespace that enforces restricted. images is green on the ref.
  • D2 It runs on CPU with a read-only root, /tmp the only writable path, user 1001, and no service account token; the caller table comes from a ConfigMap.
  • D3 A Pod in another namespace than hopin and presentator gets no answer on 8080.
  • D4 On the PC, apps is Ready at main's revision with the pod Ready.
  • D5 AGENTS.md names apps/speech-runtime/, and quality hygiene . prints hygiene: clean.

Operator actions

  • O1 On GHCR, in the package speech-runtime under "Manage Actions access", add overnightworks/platform with Read. Without it, the job cannot pull.
  • O2 One secrets:put-registry for the namespace speech-runtime in platform-secrets, as for hopin.
version = 1
now = "PR #120 is built; its pull and images check wait for O1. O2 is done: platform-secrets#13 holds the encrypted pull Secret and merges after this item lands."
next = "Review PR #120 while O1 is done, then rerun its CI and measure the pull against apps' timeout."
done_when = "speech-runtime runs in k3d and on the PC from its attested private ref through ghcr-pull, Available with both models ready, on CPU in a restricted namespace with a read-only root and its caller table from a ConfigMap, reachable on 8080 only from hopin and presentator; AGENTS.md names apps/speech-runtime/ and quality hygiene . prints hygiene: clean."

scope = [".github/workflows/ci.yml", "AGENTS.md", "apps/speech-runtime/speech-runtime.yaml", "clusters/pc/apps.yaml", "infrastructure/namespaces/namespaces.yaml"]

size = "M"

whole = "One Deployment, its Namespace and its job proof have no caller apart, and AGENTS.md describes them in the same change."

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions