deepagents edit_file succeeds on a file that was never read, although its tool description says it errors
deepagents 0.7.38 (langchain 1.5.0, langgraph 1.2.22). The edit_file tool description tells the model "You must read the file before editing; this tool errors otherwise." There is no such check. Seeding /a.txt in agent state (files={"/a.txt": create_file_data("hello world")}) and having a fake chat model call edit_file as its very first action, with no read_file call, returned status success "Successfully replaced 1 instance(s) of the string in '/a.txt'" and the file became "bye world".
Implication: the read-before-edit rule is advice to the model, not an enforced guard. Agents can edit blind on stale assumptions, and reviewers cannot rely on a read having happened. If you need the guarantee, add a middleware/tool wrapper that records paths returned by read_file and rejects edit_file for unseen paths.
Related: https://github.com/langchain-ai/deepagents/issues/7733