Users stop receiving agent responses after refreshing the page during a client tool call

[pstepaniuk]

Summary

Following the previously reported and now resolved issue (#96) that caused users to stop receiving agent responses (thanks again for fixing it!), we have identified another scenario that can lead to the same problem.

This consistently occurs when a user sends a chat message that makes the agent workflow call a client tool, and the user refreshes the page before the client tool call completes.

As a result, when the user returns to the same thread, they stop receiving agent responses, and the following error is logged for the workflow's first Agent node:

Error code: 400 - {'error': {'message': 'No tool output found for function call <call ID>', 'type': 'invalid_request_error', 'param': 'input', 'code': None}}

Expected Behavior

Users should continue to receive agent responses after refreshing the page at any time.

Minimal Reproducible Example

1. Use Agent Builder to create and publish a simple workflow with a single Agent node (e.g. gpt-5.1) that has a client tool configured.

   📸 Example

   

2. Use a basic ChatKit configuration that:

   • Includes a client tool handler that takes at least a few seconds to complete (for easier reproduction)
   • Sets the initial thread to the previously used thread ID when the page is refreshed (for extra convenience - optional)

   const { control: chatKitControl } = useChatKit({
     api: {
       getClientSecret,
     },
     onClientTool: async (tool) => {
       // Simulate tool processing delay
       await new Promise((resolve) => setTimeout(resolve, 10_000));
   
       return {
         toolCalled: tool.name,
       };
     },
     initialThread,
     onThreadChange: handleThreadChange,
   });

3. Start a brand-new thread via the ChatKit UI.

4. In the first user message, prompt the agent to trigger the client tool, and refresh the page as soon as the tool starts executing.

   📸 Example

   

5. Observe that subsequent user messages no longer receive agent responses and the Agent node errors are logged.

[grahamplace]

@pstepaniuk I wonder if enabling "Continue on error" would at least keep the agent process going in this case. Unfortunately, testing that theory is currently blocked by this bug that prevents continue on error configuration from persisting: https://community.openai.com/t/agent-builder-ignoring-destroying-error-handling-logic/1364264/2

[pstepaniuk]

@grahamplace Thanks for sharing this! Enabling "Continue on error" isn't really feasible for my use case, so I hope the root cause gets resolved soon 🤞

I also saw your other forum thread reporting a similar issue (though not necessarily the same cause), so I'll keep following that one as well.