In-Context Learning Problem in Multi-Node Agents: How History Poisons Tool Calls | Dark Hacker News