Environment
- Surface: ChatGPT Web / Code Mode
- Codex version: Not exposed to the user in this ChatGPT web environment
- OS / browser: ChatGPT running in a desktop web browser; connected tool host is Linux
- Model / plan: GPT-5.6 Sol / ChatGPT Plus
Bug
A Code Mode tool path that previously returned runtime-generated screenshots to the current ChatGPT conversation stopped delivering images.
Last known working: September 17, 2026
First observed failing: September 18, 2026
The underlying screenshot/capture operation still succeeds and produces JPEG data, but the image no longer becomes available to the model in the conversation. The resulting tool output now reports:
Image output is not supported.
Expected behavior:
A runtime-generated image returned by the tool should be delivered as an Image content item and become available to the multimodal model, as it did previously.
Actual behavior:
Capture succeeds locally, but the image does not reach the model’s visual context and the Image output path returns Image output is not supported.
Text output through the tool environment continues to work normally. User-uploaded image attachments in the same ChatGPT web experience also remain visible to the model, so this does not appear to be a general vision-input failure.
Reproduction
- From ChatGPT Web / Code Mode, invoke a connected Linux tool that captures a screenshot and returns it as an Image tool-result/content item.
- Confirm that the capture operation successfully produces JPEG data.
- Attempt to return that image to the current ChatGPT conversation.
- Observe that no image becomes available to the model and the result reports:
Image output is not supported.
This workflow was working before September 18 without changes to the intended workflow.
Diagnostics
Exact error:
Image output is not supported.
Timing:
Last confirmed working September 17, 2026. First observed failing on the morning of September 18, 2026 (US Eastern Time). The failure has remained reproducible since then.
Additional observations:
- The Linux-side capture path can still produce JPEG data.
- Text tool output continues to work.
- Normal user-uploaded images are still visible to the model.
- The failure therefore appears specific to runtime-generated Image tool output being relayed back into the ChatGPT conversation.
Could you confirm whether runtime-generated Image tool results are expected to be supported on this ChatGPT / Code Mode surface? If so, this appears to be a regression. If not, is there a currently supported mechanism for returning a runtime-generated screenshot to the same multimodal conversation?
9-21-2026
Update: I was able to isolate this further by testing the same HomeNode MCP screenshot tool with Codex as a separate consumer.
The tool returns a standard MCP result containing content[type=image] with base64 JPEG data. Codex successfully consumes that image and can accurately describe screen contents that are not present in the accompanying metadata.
The same image-result path in ChatGPT Web / Code Mode still does not make the image available to the model.
This appears to narrow the issue to consumer/runtime handling of MCP image content rather than screenshot capture or JPEG generation.
I’ve added the full isolation result to the corresponding Codex GitHub issue: #46927.