[macOS] Codex loads images from unrelated conversations by scanning ~/.codex/generated_images

Environment

• Product: Codex desktop experience in ChatGPT for macOS
• App version: 26.825.32147 (released Aug 28, 2026)
• Model: 5.6 Sol Ultra
• Subscription: ChatGPT Pro
• macOS: Tahoe 26.6.2
• Hardware: 16-inch MacBook Pro, Apple M5 Max, 128 GB memory
• Feedback ID: 01a05075-fc76-74c1-b20a-07d31b85ff2a

Description

While working in one conversation, Codex unexpectedly searched the shared local generated-image directory and opened images belonging to other, unrelated conversations. The active conversation did not ask Codex to inspect those conversations or their images.

The transcript shows Codex running a command beginning with:

find /Users/<username>/.codex/generated_images -type f -name '*.png' -mtime -3 -print0 | xargs -0 ls -lt ...

It then displayed “Viewed 5 images”, including several prior-session assets unrelated to the active task.

This appears to break conversation/project isolation. It also adds irrelevant images to the model context, wastes image-token/compute usage, reduces work efficiency, and likely contributes to faster consumption of usage limits. It can additionally cause the model to use the wrong visual references and raises a cross-conversation privacy concern.

Observed sequence / steps to reproduce

1. Open a Codex conversation in ChatGPT for macOS.
2. Work on an image-based UI task whose references are limited to the current conversation.
3. In the affected run, Codex searches for recent PNG files under the global ~/.codex/generated_images directory.
4. Codex opens images from unrelated conversations and adds them to the active context, shown as “Viewed 5 images.”

Actual behavior

Codex scans a shared local generated-image directory and inspects unrelated images without an explicit request or confirmation.

Expected behavior

Codex should only inspect images explicitly attached, generated, or referenced in the current conversation or project. Images from other conversations should remain isolated and should never be loaded into the active context unless the user explicitly selects them.

Impact

• Unnecessary image-token and compute consumption, causing usage limits to be depleted faster.
• Slower and less focused task execution.
• Risk of incorrect output based on unrelated visual assets.
• Cross-conversation privacy and context-isolation concern.

Additional information

Observed on Aug 29, 2026. The feedback report was uploaded under the ID above. Screenshots captured the shell search, the “Viewed 5 images” event, and thumbnails that were unrelated to the active conversation.

No source file or test is named. Start by reproducing the macOS scenario and tracing how Codex discovers files under ~/.codex/generated_images and associates them with the active conversation. Done means unrelated images are not inspected or added to context, while explicitly attached or referenced images still work.