Skip to content

[Bug] Computer Use can't take images to see context, can't press some keys #457

Description

@Auditormadness9

Codex Web GPT version

5.0.6

Codex version

26.908.40834

Platform

Windows x64

OS version

Windows 11 25H2

ChatGPT account tier

Plus

Integration mode

Full harness (MCP)

Selected model

ChatGPT Web - High

What happened?

I am using Codex to fix a Garry's Mod map for me that I recently ported from another game with Astra. Everything is okay but it had bugs like npcs spawning without weapons or objectives not pointing with correct waypoints, so I tried to use Zero Risk to fix the issue, it initiated Computer Use, launched the game, and tried to make a few changes then test it in game by screenshotting things for itself. Unfortunately, I checked the chat and it said that it wanted to feed the image to the model but it was not accepted/environment doesn't support images. Theoretically, I can understand why this can happen given that we work with Temporary Chats where image uploads are not allowed, however, a lot of times when I manually re-send prompt (because Zero Risk) it opens up regular chat instead of temporary, so I thought maybe this issue won't happen now, yet it still happens, even though regular chat supports image uploads.

Is it possible to perhaps convert images to base64 or something? I've seen this happen when copying my Google Vertex prompts where I had image uploads as a python script, it encoded them into base64 or some long string.

Unfortunately if this is not supported I cannot really use it since it is very important in my use case, the harness ends up cancelling the job mid way and saying that it didn't have necessary tools or openai safety checks blocked it so it couldn't conduct the necessary "smoke tests".

Minimal reproduction

  1. Start a task with ChatGPT Web - Zero Risk Light selected (or Native2, whatever you use, I use ZR so I can only report on this), preferrably one that requires visual checks (eg. game doesn't render thing properly)
  2. Paste the prompt in browser and click sent
  3. Wait until it reaches moment where it initiates Computer Use to check things
  4. Observe the reasoning chains in chat, in many of them you can spot image uploads but clicking on them shows in text that environment doesn't support images
  5. Wait a bit longer and Codex finishes the job saying it couldn't conduct the tests properly since it didn't have necessary tools.

Complete final error

Native app bindings are unavailable for windows

unsupported key: asciitilde

<image content omitted because you do not support image input>

Privacy-safe diagnostics

codex-web-gpt-diagnostics-2026-09-12.json

Image Image

Other Codex wrappers or route owners

none

Preflight checklist

  • I tested the latest release and, for catalog or turn failures, fully restarted Codex including its background process.
  • I stated whether the failure also occurs in a fresh Codex task, or explained why that does not apply.
  • I attached only a privacy-safe export, or explained why the exporter could not produce one; I removed any remaining private content.
  • I searched existing issues for this exact final error.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P2Medium: important but scoped or non-blockingbugSomething isn't workingneeds reproNeeds a matching local, platform, or account reproduction.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions