What duplicate file checks miss inside Codex session files
We looked at 1,467 Codex session files on a Mac we use every day. A whole-file duplicate check found nothing, because no two files were identical. Inside the files, 76.9% of the decoded image bytes were copies of an image that had already appeared.
TL;DR
- The problem. A whole-file duplicate check of 1,467 Codex session files found nothing, because no two files were identical.
- What we found. Inside the files, 76.9% of the decoded image bytes were copies of an image that had already appeared.
- What it means. To find that space you have to look inside each conversation file. Comparing whole files misses it.
What we measured
In September 2026 we read every Codex session file on one actively used Mac, 1,467 files and 16.509 GB in total. We only read them. Nothing was rewritten, moved or deleted during the study.
Codex keeps screenshots and other images inside the session file as encoded text. We pulled out every image payload we could recognize, decoded it, and took a SHA-256 hash of the image bytes. Two occurrences with the same hash are the same image, whichever file they sit in.
What we found
The files held 62,014 image occurrences but only 12,104 distinct hashes. Decoded, the images came to 5.987 GB. Keeping the first occurrence of each image and counting everything after it as a copy, the copies added up to 4.605 GB, which is 76.9% of the decoded image data.
A duplicate file finder compares whole files, so on this Mac it would have reported no duplicates at all. The repetition lives inside the conversations. We did not try to work out why each copy was written, which would be a separate study.
Limits of this study
This is one Mac and the way it happens to be used, so the share on another machine could be much higher or much lower. Payloads we could not recognize as images were left out of the image count. It is a study of file contents and says nothing about how accurate Mint’s cleanup is.
The 76.9% is measured against decoded image bytes only. Session files also hold text, and disk usage is counted differently again, so please don’t read the figure as an estimate of space you could get back.
Real-data results: compression and a separate Undo test
In a September 26 Mint-team experiment, a different, file-level compression method in a development build saved 1.29 GB on 3.13 GB of real Codex files: about 41%, or 428.42 MB per conversation on average. All four files were byte-identical and Codex read the same 1,598 turns. Separately, an October 15 test with public Mint 1.0.99 reduced storage by 35.1% in eight eligible conversation copies; continuation and Undo content checks passed. Different methods and files, with separate dated evidence. The 76.9% image-repetition figure above is not a measured savings result.
A case you can check on your own Mac
This published study establishes repeated content, not a successful optimization benchmark. To test the product claim, record both the space change and whether the work still functions. Use a backed-up, non-sensitive test conversation and keep an independent copy; a test report should never include private chat contents.
- Record the Mac, macOS, Mint edition/build and Codex version, test date, selected action and sample count. Note running tasks, APFS snapshots and backups that can affect free space.
- Before any change, record the selected files’ logical and allocated sizes, free space on the same volume, and the scan estimate separately. Note how to reopen the conversation and which pictures or dependent conversations must remain usable.
- Run only the supported action offered for that sample. Record completed, skipped and refused items and the actual result. Keep backup and recovery data in the accounting; do not delete them just to improve the number.
- Reopen the conversation and check the required text and images. Where applicable, continue dependent conversations and test eligible Undo. Record failures and unsupported checks, not just passes.
- Re-measure free space immediately and after the next normal work session, ideally again after 24 hours. Report regrowth and concurrent downloads or writes. Publish anonymized measurements with the method and limitations; a second reviewer can repeat the same protocol.
What this means for cleaning up Codex
Finding the same image twice does not make it safe to rewrite a conversation. Another conversation may be built on top of this one, you may still want the screenshots, and Codex only supports certain changes to its own files. Settle those questions before you look at any hash.
The current version of Mint has three actions for Codex conversations. Optimize shrinks the space that supported repeated screenshots take and keeps the pictures Codex reads when you continue the conversation. It can also store eligible PNG images in less space with the same pixels, and you can undo it as long as the original data and history it needs are still there. Keep text only keeps the conversation text and permanently removes embedded images, audio and video, leaving placeholders instead of small previews. Delete permanently removes the whole selected conversation through Codex. Support depends on the tool, the format and the version, so read the action and its confirmation in your installed edition before you go ahead.
Frequently asked questions
Does 76.9% mean I can free 76.9% of my Codex storage?
No. The figure is a share of the decoded image bytes in this one sample. Session files also contain text, and the study made no estimate of recoverable space.
Why didn’t a duplicate file finder catch this?
Duplicate finders compare whole files, and no two of the 1,467 session files were identical. The repeated images sit inside different files.
Does this study measure how accurate Mint’s cleanup is?
No. It is a content study on a single Mac. Measuring accuracy would take labeled samples from many machines and app versions.
See what Codex keeps on your Mac
Download Mint free and open the Disk tab to see what your AI tools have stored before you decide what to change.