We see occasional cases where our Agent Builder workflow stops responding to user messages, with no error shown in the thread’s logs.
It might be related to client tool calls, but I’m not positive about that.
- User sends messages
- Agent responds
- Client tool calls are used successfully, user continues to go back and forth with agent.
- Then at some point, another client tool is called successfully, but that is the last response the agent ever gives. In the logs we see:
- CLIENT TOOL CALL
- Status: completed
- Output: { “success”: true, … }
- TASK GROUP
- Created is after the client tool call
- Type: Thought
- USER: User message
- User says something normal, should get a response from agent
- Instead, agent never responds
- USER: User message
- USER: User message
- …. etc, user never gets another response from Agent
- CLIENT TOOL CALL
- The Status of the thread is still shown as “Active”
Similar issues:
- Users stop receiving agent responses in longer ChatKit threads involving client tool calls · Issue #96 · openai/chatkit-js · GitHub (I’m on this thread, last time it turned out silent rate limit errors were happening under the hood. We couldn’t see those errors in the logs. Once our usage tier was increased to avoid rate limits, the issue was resolved. Not sure if the same thing is happening since they don’t show in Dashboard)
- Users stop receiving agent responses after refreshing the page during a client tool call · Issue #107 · openai/chatkit-js · GitHub (in my case, it’s not because of mid-client tool refresh, and we don’t see any error logged as described on this issue)
Example threads:
- https://platform.openai.com/logs/cthr_6930d7d7bea881948b3d4a1eda43db6000b2783dce918c1e
- https://platform.openai.com/logs/cthr_692f996bd4208195b2345f71b3d66ce70ffae3760f062495
Unfortunately, I don’t have a way to deterministically repro.