Test and troubleshoot page tools

Try every tool the way a visitor would before you launch. While you test, turn on Show actions (the deployment's Chat widget tab → Behavior) so each call shows as a card.

Checklist#

1. Each tool: ask for it in plain words. The card reads Completed and the answer uses the result.
2. Tools that ask first: Approve runs the tool; Reject doesn't, and the card reads Denied.
3. Forms: the agent fills them in; with toolautosubmit the form is submitted; password fields stay empty.
4. Reload mid-call: start a slow tool and reload the page. The card reads Not completed, and the tool doesn't run twice.
5. Stop: start a slow tool and select Stop. The card reads Not completed, and the agent doesn't reply.
6. A new message mid-call: the tool stops and the agent answers the new message.
7. Changing pages: move to a page without a tool and ask for it again. The agent says it can't do that there.
8. Two tabs: open your site in two tabs and send from one. Only that tab runs the tool.
9. Failures: make a tool throw. The card reads Error, and the agent explains.
10. Your team's view: find the conversation in Inbox, and the tool in the Tools column of Analytics → AI Model.

Test in Chrome, Safari and Firefox, and on a phone.

Follow every call with the toolResult event#

const chat = await HelpAI.ready();
chat.on('toolResult', ({ toolCallId, name, status, durationMs }) => {
  analytics.track('help_ai_tool', { name, status, durationMs });
});

The event never carries the input or the result. Answering one of the agent's questions in the chat fires it too, with only toolCallId.

Common problems#

When you report a problem#

Include:

• The conversation id: (await HelpAI.ready()).state.conversationId in the browser console.
• The request body of /pai/stream-message (DevTools → Network), which lists the tools the message offered in clientTools.
• Any warnings in the browser console, and a screenshot of the card.