Skip to main content
Start with the browser console on the affected product page:
version identifies the loaded SDK build. getState() shows the last accepted, serializable initialization snapshot. It does not expose the identity token or prove that the server accepted its signature.

​
When nothing appears

After publishing, reload an already-open product tab. Configuration can remain cached briefly even after the publish cache is cleared, so allow up to two minutes before treating the old version as a deployment failure.

​
Identity requests fail

For a 401 or 403 response, confirm that:
  • userId is stable and exactly matches the JWT sub;
  • the JWT is signed with the current workspace secret using HS256;
  • exp is in the future and no more than seven days away;
  • the token was minted by the customer server, not the browser; and
  • the intended environment is in Observe or Enforce as expected.
getState().initProps.verified means a non-empty token was supplied to the SDK. Use a successful server request and the workspace identity coverage view to verify acceptance.

​
An action is unavailable

Check the complete chain in order:
  1. The action is reviewed, enabled, and published.
  2. A published Skill capability contains the action.
  3. The Skill is attached to the active Wizard or Resource Center Agent.
  4. The experience is published.
  5. For a client-side callback, the current page includes it in init({ actions }) or the latest complete setActions() replacement.
  6. For a Session API action, the customer-owned executor recognizes the published tool key.
Open the action’s Activity tab to distinguish unavailable, declined, invalid-input, and failed outcomes. See Recover or roll back configuration before changing Production.

​
Resource Center or Guidance is incomplete

  • For Resource Center Search, confirm the knowledge articles are indexed and returned by the runtime. Articles are delivered automatically after ingestion; there is no article Publish button. If results remain stale, follow Knowledge delivery.
  • For Ask, confirm Ask in your own words is enabled and its reserved Agent has the intended published capabilities and knowledge scope.
  • For a reviewed guide, confirm both the guide and workspace Guidance configuration are published.
  • For a goal-based Guidance run, confirm Guidance is enabled and the target page is supported.
  • For a stale Resource Center card, confirm its referenced guide, article, callback, or connector is still available, then republish the Resource Center.

​
Guidance stops before the first useful step

Retry the same safe Test goal once. If it fails again with the same message before selecting a target or presenting an instruction:
  1. End the Test session so its temporary token is revoked.
  2. Confirm the saved Guidance draft, attached knowledge, and target URL are still the intended inputs.
  3. Open Guidance → Analytics and inspect the failed run timeline instead of relying only on its generated theme or aggregate outcome.
  4. Record the exact time, environment, workspace, goal, page URL, and visible error before escalating.
A failure before the first useful step does not by itself identify a bad target, prompt, model, or service. Preserve the run evidence so the responsible layer can be diagnosed without guessing.

​
The SDK was destroyed

destroy() is terminal for that installed SDK instance. Repeating init() after destroy() is ignored. Reload the full page to create a new instance. Adding the CDN script again in the same document is treated as a duplicate install and does not revive the destroyed instance.

​
Collect evidence before escalating

Record the product URL, time, workspace, Test or Production context, SDK version, affected entity and version, relevant Network status, and the exact [dubot] console message. Avoid including the identity token, verification secret, customer content, or unnecessary personal data.