Tools
LangChain 1.4.2 stops human-in-the-loop edits from dropping the model's own tool calls
langchain 1.4.2, released 2026-09-18, preserves model-generated tool calls when a human edits a tool call in a HITL interrupt, and adds a notice to ToolMessage. The prior release, 1.4.1 on 09-16, preserved open MCP object arguments through the same path. Both are cases where the approval step was quietly rewriting what the model actually asked for, which is the worst place for silent mutation since the human believes they approved the original.
Source
↳ Follow the thread