editfidelity
A byte-level regression suite for the file-edit tools coding agents use. Eight tools get the same one-line edit on twelve hazard files; 17 of the 96 results are defects.
- Year
- 2026
- Kind
- Evaluation · Python
- Status
- Shipped

Coding agents change files through small edit tools. editfidelity gives eight of them, from OpenHands, aider, SWE-agent, cline and goose, the same one-line edit on twelve hazard files: CRLF and mixed line endings, a BOM, UTF-16, Latin-1, tabs, no final newline, an exec bit, a symlink, a NUL byte and an empty file. The result is compared byte for byte with the expected output. No model, no API key, no network.
Of the 96 results, 17 are defects: 15 silent corruptions, where the tool reported success and changed something it was not asked to, and 2 crashes. Both OpenHands tools and SWE-agent's rewrite a CRLF file to LF and say it worked. SWE-agent's also expands every tab to spaces. One OpenHands tool replaces a symlink with a regular file and leaves the target unedited. goose's edit tool is exact or a clean reject on every hazard.
Every safe result is pinned, so a new release that breaks one fails the check. The limits: one edit shape, a one-line change in the middle of a small file; Linux only; and only the entry point it drives, not every route an agent takes to the disk. Each result names the release it was measured on.
