Skip to main content

Troubleshoot testing and publishing

Identify the exact agent, version, and surface before changing configuration.

The private browser test does not start​

Allow microphone access, confirm the active workspace, check usable credits and payment status, and retry after the editor shows a successful save. Chat Agents cannot start voice Playground sessions.

The browser test uses old behavior​

Confirm the save state completed before the test session was created. Close the current test, verify the draft, and start a new test. A session snapshot does not update while it is running.

This is expected after later edits. A shared link freezes the Playground configuration at creation. Revoke it and create a new link from the intended saved draft.

It may be missing, revoked, expired, at its attempt limit, blocked by insufficient workspace credits, or blocked by payment lock. The sender can inspect active links and create a replacement if appropriate.

Connected tools are absent in a shared test​

The link was likely created with Allow connected tools off, or the integration is disconnected. Knowledge search remains available independently. Create a new link if the risk review permits tools; an existing snapshot cannot be edited.

Publish is cancelled before validation​

The editor could not flush the latest draft. Restore connectivity, make a small change, wait for a successful save, reload to confirm persistence, and publish again.

Publish is blocked by configuration errors​

Read the first diagnostic, fix the referenced architecture, model, voice, language, or required setting, and rerun the test. Resolve prompt and function warnings separately.

Publish says no draft exists​

Open the editor to initialize the draft, confirm its contents, wait for autosave, and retry.

Publish is unavailable because nothing changed​

The saved draft matches the published configuration and function/tool snapshots. Confirm you edited the correct agent and workspace.

Live calls still use the old behavior​

Saving and testing do not publish. Confirm the intended entry is Live, then start a new call. An existing call stays on the version loaded at its start.

Restore did not change production​

Restore creates a draft. Test that draft, publish it as a new version, and start a new production interaction.

The test record lacks a summary or outcome​

Browser voice tests do not currently run the live post-call AI analyzer. For a live call, processing can be delayed; wait, refresh the call record, and then investigate post-call processing if it remains incomplete.

Credits were charged for a test​

Private and shared browser voice tests are billed. Shared tests draw from the owner's workspace. Compare the stored duration and tested configuration rate, then inspect usage aggregation and separately priced tool activity.

Escalation evidence​

Provide workspace, agent ID, test or call ID, version, surface, time and timezone, expected result, transcript excerpt, visible diagnostic, tool name, and external result. Redact phone numbers and never send credentials or secure share tokens.