Skip to content

Local Harness + Celln hands-on session

Historical record

This page records a Celln development milestone (September 2026). Interfaces and limits it describes may since have changed or been replaced by the Celln fleet. For current behaviour see Celln Backend and Celln Fleet Installation.

This is a bounded, eight-hour developer session on the explicitly selected kind-celln-deployed cluster. It is not a production installation, a persistent conversation, or a reboot-survival qualification. Framework is not used.

The session keeps the real signed catalogue, approved runtime and borrowed tools, host issuer, router, KVM dispatcher, namespace-scoped controller, authenticated API with embedded UI, and authenticated NATS event bus running together. The DeepSeek credential stays in a private host file, never in the browser or a Pod. The UI gets a distinct random API credential. Capability discovery gets its own read-only router credential, not an execution credential.

Start and verify

Use the prerequisites and explicit paths in the live catalogue guide. Build the API image from the current integration worktree and load it into all Kind nodes using the Podman image-archive workflow in the installation guide. Set the ordinary live proof variables, including the explicit kubeconfig, pinned kernel/initrd checksums, signed public fixture, native Harness package, binaries and image tags. Add:

export CELLN_LIVE_INTERACTIVE=1
export CELLN_LIVE_NATS_BINARY=/absolute/path/to/nats-server
export CELLN_LIVE_ISSUER_PROCESS=1
export CELLN_LIVE_AUTOMATIC_ISSUANCE=1
export CELLN_LIVE_OPERATOR_ADMISSION=1
export CELLN_LIVE_BROWSER_SUBMISSION=1
export CELLN_PAUSE_TEST_CONTROLLER=1

Run test/integration/test-celln-catalogue-harness.sh within the isolated Podman private network namespace, as in the live proof guide. Do not start another session while this one owns the paused proof controller. Session startup incurs two real, bounded DeepSeek runs and refuses readiness unless both complete via the actual browser, including explicit Harness selection and ordered lending.

The printed evidence directory contains interactive-ready.json: the namespace, session PID, expiry, two successful run names and the private API-token file location. No token contents are printed. The issuer's read-only Kubernetes token and unique TLS certificates last nine hours; the session stops after eight, leaving time for finalizer cleanup. This is not automatic credential renewal.

Forward this session's embedded UI/API from your normal host shell:

kubectl --context kind-celln-deployed --namespace SESSION_NAMESPACE \
  port-forward --address 127.0.0.1 service/browser 9090:8080

Open http://127.0.0.1:9090/. Log in with the session API token, then select the exact namespace from interactive-ready.json. Do not run make web-dev-serve: it targets the separate historical sympozium-system API and can start a second, mismatched frontend. The session API already includes the matching built UI.

Try it

  1. Open Runs → New Run and select Agent agent.
  2. Choose runtime in the one-run Harness selector (or inherit that Agent default).
  3. Select Celln as the backend and deepseek-chat as the model.
  4. Lend uppercase@v1, then length@v1, in that order. This exact combination is prepared and registered; arbitrary combinations are not automatically admitted.
  5. Submit: Call uppercase with text hello, then call length with the returned text. Use both tools exactly once in that order. Report the uppercase text and its length.
  6. Open the resulting run and select Result. The answer is displayed first; Raw Harness output preserves the original event trace. Reported tool events are presentation data, not independent proof of authority or isolation.

You can submit more independent tasks against this selection. Each run permits at most three model requests, 512 output tokens per request, 1,536 total output tokens, and only the approved tools and model endpoint. This is a small tool-lending demo, not a general shell or a conversational Pi/Hermes session. The event stream connection is real, but native Celln answers are delivered on completion, not token-streamed. Topology shows the Agent's relationship to the native runtime.

Stop

Send SIGUSR1 to the session PID from its readiness file, after checking that it is still the owned celln-catalogue-setup.test process. The session archives all its AgentRuns to session-runs-at-stop.json, deletes them with UID preconditions while the controller can still finalize, tears down its own namespace and host processes, and restores the original proof controller. Stop the corresponding UI port-forward too. Other namespaces and Framework are untouched.

The eight-hour session is not resumable after reboot. A restart creates a fresh namespace, new trust identities and new evidence; it performs admission again. interactive-ready.json is historical after stop: also check that no interactive-stopped.json exists and that the PID, Pods and UI are live.