Codex CLI: Full-Resolution Image Inspection for Precision Visual Tasks
CodexCodex CLI 0.115.0 enables supported models to request full-resolution image inspection via view_image and codex.emitImage(..., detail: "original"). Previously, image inputs were processed at a standard downsampled detail level; the new detail: "original" parameter bypasses downsampling to give the model access to the image at native resolution. The feature is opt-in and most impactful for precision tasks: UI debugging, dense diagram analysis, code screenshot review, and pixel-level visual comparison.
Sources & Mentions
3 external resources covering this update
Full-Resolution Image Inspection
Codex CLI 0.115.0 introduces a new image detail mode that allows supported models to inspect images at their original full resolution. The feature surfaces via two APIs:
view_image— the built-in image viewing tool used within agent contextscodex.emitImage(..., detail: "original")— the programmatic API for emitting images from within the JavaScript REPL or custom tool code
Prior to this release, image inspection in Codex operated at a standard detail level, applying downsampling to manage token costs. For most use cases — attaching a screenshot for context or referencing a diagram — standard detail is sufficient. But for precision visual tasks, downsampled images can cause missed details or misinterpretations.
The detail: "original" Parameter
Passing detail: "original" tells Codex to deliver the image to the model at full native resolution, bypassing the downsampling pipeline entirely. This gives the model access to every pixel of the original file. The tradeoff is higher token consumption — developers should use this selectively and budget accordingly.
The feature is strictly opt-in. Existing workflows that do not pass detail: "original" are unaffected and continue to use the default detail level.
Use Cases
Practical applications where full-resolution image inspection adds meaningful value:
- UI debugging — inspecting screenshots of web or native app interfaces to identify layout defects or rendering issues visible only at full resolution
- Diagram analysis — reading engineering, architecture, or data-flow diagrams with dense annotations that downsampling makes illegible
- Code screenshot review — when source code is shared as an image rather than text, full resolution ensures all characters are legible to the model
- Visual diffs — comparing before/after screenshots of a UI or document at pixel level for QA purposes
Model Compatibility
Not all models handle full-resolution image tokens equally. Developers should verify that the model in use supports high-detail image input before enabling this mode in production workflows.