You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Client tools never execute against spec-compliant AG-UI servers (plain success RUN_FINISHED) #1464
When useChat is connected to a third-party AG-UI server (Pydantic AI's AGUIAdapter in my case), the model calls a registered .client() tool, the tool call is rendered, and then the run just ends. The .client() handler is never invoked and no continuation run is started.
The server emits the canonical AG-UI sequence for a frontend tool:
Client-tool execution is only triggered from StreamProcessor.handleInterrupts (activities/chat/stream/processor.ts), i.e. when RUN_FINISHED carries outcome.type: "interrupt" with a metadata.kind: "client_tool" interrupt. That's the shape TanStack's own chat() emits, but it isn't what AG-UI specifies.
Per ag-ui-protocol/ag-ui#2731, a frontend tool call ends the run with a plain success outcome. It may include pendingToolCallIds; otherwise the client derives the unanswered calls from the stream. The client then executes the tools and continues with role: "tool" messages in the next run's history. There is no success-outcome path that calls onToolCall, so nothing fires.
Expected
On a successRUN_FINISHED, the client should dispatch onToolCall for tool calls that were started but not answered by TOOL_CALL_RESULT, using outcome.pendingToolCallIds when present. The rest of the pipeline already exists (pendingToolExecutions, addToolResultForClientTool, continuation). The interrupt path would stay unchanged.
Repro
Pydantic AI agent exposed via AGUIAdapter, no server-side ping tool.
Ask the agent to call ping and report the result → handler never runs, conversation stops.
Your Minimal, Reproducible Example - (Sandbox Highly Recommended)
_
Screenshots or Videos (Optional)
No response
Do you intend to try to help solve this bug with your own PR?
None
Terms & Code of Conduct
I agree to follow this project's Code of Conduct
I understand that if my bug cannot be reliable reproduced in a debuggable environment, it will probably not be fixed and this issue may even be closed.
Activity
fix(ai): run the client tools that a success RUN_FINISHED leaves pend…