Users stop receiving agent responses in longer ChatKit threads involving client tool calls

[pstepaniuk]

Summary

When a new thread is started and one of its earliest agent responses invokes a client tool, the user can continue chatting for a few turns, but eventually stops receiving agent replies.

The workflow typically fails at the first Agent node with No tool call found for function call output with call_id (...), even though the referenced tool calls are visible in the conversation history.

Impacted Users

Multiple users have reported the same behavior:

• Bug Report: onClientTool handler executes but agent fails to receive output #94 (comment)
• Bug Report: onClientTool handler executes but agent fails to receive output #94 (comment)

Expected Behavior

Users should continue receiving agent responses regardless of conversation length or the number of client tool invocations.

Minimal Reproducible Example

1. Use Agent Builder to create and publish a simple workflow with a single Agent node (e.g. gpt-5) that has a client tool configured.
   📸 Example

   
2. Use a basic ChatKit configuration to run the chat.

   const { control } = useChatKit({
     api: {
       getClientSecret,
     },
     onClientTool: async (tool) => ({
       toolCalled: tool.name,
     }),
   });

3. Start a brand-new thread via the ChatKit UI.
4. In the first user message, prompt the agent to trigger the client tool.
   📸 Example

   
5. Keep chatting in the same thread by sending several more user messages (typically around 5 is enough); no additional tool calls are required.
6. Observe that subsequent user messages no longer receive agent responses and the Agent node errors are logged.

[chriskyle-oai]

Thanks for the detailed report @pstepaniuk. We're investigating this issue right now and hope to have an update for you soon. Appreciate your patience while we look into this, I'll follow up in a bit with an updated status!

[chriskyle-oai]

We've identified the issue and are working on a fix right now, will follow up once it is deployed. Thanks again for your patience!

[chriskyle-oai]

The fix should now be fully deployed!

[grahamplace]

@chriskyle-oai thanks for working on this 🙏

I'm still having the issue where client tool calls seem to kill the agent silently.

One recent ChatKit thread with this issue: cthr_691bb400ce948196b75e1a9b2b26553105b2a445c5e770b3

The logs show

• User message (which I expect to trigger a client tool call)
• Client tool call, Status: completed
• Task group (cti_691bb40367a88196b0178c814984677b05b2a445c5e770b3)

Then nothing else in logs. No error shown in the frontend in ChatKit either, it continues to show the "..." state like it's working

I am able to add another user message, which changes the frontend ChatKit state to showing Thought for a while as the last agent message (where it was "..."), and I do see the user message in the logs, but the agent never responds to user messages from that point forward

Also worth pointing out, the original issue report specifies "in longer ChatKit threads" but in my case the issue happens with just one user message + one client tool call

[grahamplace]

This is pure speculation — is it possible that the cause is the agent process hitting a model rate limit behind the scenes and failing silently?

My org is only Usage tier 1, and I did hit the gpt-5.1 TPM limit the other day when developing on this workflow

[pstepaniuk]

@chriskyle-oai Thank you! No further issues on my end so far ✅

[trevorcreech]

confirmed that @grahamplace's issue is from hitting rate limits—moving to DM and we'll work to surface those better going forward

[grahamplace]

@chriskyle-oai I'm seeing the same behavior OP and I reported on this issue (I have no way to tell if it's the same cause -- in my case rate limit errors under the hood)

Posted details in the forum here: https://community.openai.com/t/agent-builder-agent-process-silently-dies/1368412