Asking an LLM to edit a file it has never seen results in the model fabricating the contents based on the filename. Always. The read-before-edit guard tracks a per-session set of read files and rejects Edit calls on anything outside that set. The agent reads the error in its tool result, course-corrects by reading the file, then retries. It is the difference between a coding agent that helps and one that silently corrupts your repo.
Engineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.