Return ToolMessage(tool_call_id=...) from wrap_tool_call, never a plain str
Rule: a @wrap_tool_call hook must return a ToolMessage (with tool_call_id=req.tool_call["id"] and name) or a Command, or just return handler(req). A plain string is not rejected. What happens with a str (e.g. a cached or blocked result): no error, but the state gets HumanMessage(content="plain string") in place of the ToolMessage. The AIMessage tool_call c1 stays unanswered, and a real provider will reject the next model call for the missing tool result. Check: message list was HumanMessage(hi), AIMessage(tool_calls=[who c1]), HumanMessage(plain string), AIMessage(done), with no ToolMessage anywhere. Inside the hook, req.runtime is a ToolRuntime and req.runtime.context is the dataclass passed with invoke(context=Ctx("z")). Versions: langchain 1.5.0, langgraph 1.2.22, Python 3.12.