Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

A coding agent that ships five different ways to change a file, and the one it turns on by default is the only one that put my code on the wrong line without saying so

There is a particular kind of bug that only exists because we handed our editors to a language model.

The rest of the article describes a controlled teardown of the omp “edit format” used by can1357/oh-my-pi, comparing five edit modes across 150 trials with different file drift and anchor fidelity. The author finds that while four modes either refused or avoided wrong writes when the anchor was off, the default “hashline” mode frequently applied changes to the wrong line while still reporting success, largely because it authenticates the whole file via a content hash but does not provide enough redundancy to detect when the specified line number is wrong. The write-up then explains how hashline’s #TAG is enforced (preventing edits for the wrong file/session) and how additional guards (like boundary repair and parse checks) affect when it refuses versus silently succeeds. The conclusion is that redundancy matters: formats that force the model to restate or match current text can catch inconsistencies, whereas hashline can “prove” the file identity but still land wrong edits when the location is incorrect. Finally, the author compares omp’s modes to Claude Code’s replace and Codex’s apply_patch and gives practical recommendations on when to keep hashline as default and when to switch to more conservative formats for long or unattended sessions.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI