App integration can read context but cannot apply edits on macOS

ChatGPT Bugs
1

Environment

  • ChatGPT Classic 1.2026.184 (1784145287)
  • macOS Tahoe 26.5.2
  • Tested apps:
    • TextEdit
    • Visual Studio Code
    • Xcode

Problem

ChatGPT can access and read content from shared macOS apps, but edits generated by ChatGPT are not applied to the target application.

The issue is reproducible across multiple apps, so it does not appear to be specific to TextEdit, VS Code, or Xcode.

Steps to reproduce

  1. Open TextEdit and create a new empty document.
  2. Share the TextEdit document with ChatGPT.
  3. Ask ChatGPT to write TEST into the document.
  4. ChatGPT generates an edit/patch and reports that the change was sent.
  5. Check the TextEdit document.

Actual result

The document remains unchanged. TEST does not appear in TextEdit.

The same behavior occurs in Visual Studio Code and Xcode: ChatGPT can access the shared app context, but generated edits are not applied.

Expected result

When ChatGPT generates an edit for a shared application, the change should be applied to the corresponding document/file in that application.

Permissions / troubleshooting performed

The following troubleshooting steps were already performed:

  • Restarted ChatGPT completely.
  • Restarted the target applications.
  • Removed and re-added app/document sharing.
  • Tested with newly created documents/files.
  • Tested multiple applications: TextEdit, VS Code, and Xcode.
  • Disabled and re-enabled ChatGPT under:
    System Settings → Privacy & Security → Accessibility.

After re-enabling Accessibility permission, ChatGPT displayed the dialog:

Enable Access Permissions

Clicking Open System Settings opened the macOS Accessibility settings. ChatGPT was enabled there successfully, but the permission dialog inside ChatGPT did not disappear.

Only after completely quitting and restarting ChatGPT did the permission dialog disappear and app context access start working again.

However, write/edit operations still did not work after the restart.

Additional observation

Under:

System Settings → Privacy & Security → Automation

there is no entry for ChatGPT.

Under Accessibility, ChatGPT can only be enabled or disabled.

Current behavior

  • Reading/shared app context: works
  • Detecting the shared document/app: works
  • Generating an edit/patch: appears to work internally
  • Applying the edit to the target app: fails
  • Affected apps: TextEdit, VS Code, Xcode

This suggests the issue may be in the macOS app-integration write/edit path rather than in the individual target applications.

2

Can someone from OpenAI have a look on this please?

3

I’m seeing the same issue on macOS.

ChatGPT Classic can detect the shared app and read its contents correctly. It also calls the editing tool and generates the proposed edit, but after a few seconds, the tool aborts and nothing is applied.

I’ve reproduced this issue with both:

  • ChatGPT Classic 1.2026.183
  • ChatGPT Classic 1.2026.184

Troubleshooting steps attempted:

  • Reinstalled the app integration
  • Reset permissions
  • Restarted all related apps and services
  • Tested in a Temporary Chat with Thinking enabled

Accessibility permission is enabled, and content reading continues to work normally.

Affected apps:

  • Visual Studio Code
  • Apple Notes

Relevant Console logs:

Could not match any valid candidates of any source.
containerToPush is nil, will not push anything to candidate receiver.

Since two different Classic versions fail in exactly the same way, this seems more likely to be a server-side editing-tool regression than a local permissions or installation issue.

Has anyone received confirmation from OpenAI that this is being investigated?

4

Unfortunately I didn’t get any confirmation. This bug is very annoying. Can someone from OpenAI please investigate this? Do you need any further informations?

5

In my humble opinion this is not a bug. It is an auto sabotage of the tool to forbid the use.

Currently, using this technique is requiring far more less tokens that agent modes and the industry does not want to communicate about this product in a way to make a maximum amount of agent mode user.

The finally did figure out that with this mode we did became what we rely want. Augmented humans, those whose are using ai as a prolongation of there own brains.

In my case I am a software engineer who is building a complex platform on my own.

During the last 2 years I was using this tool to write all the code, snippets after snippets, by using a complete transversal approach.

I have 32 repositories and when I work on a feature this can be at the same time on many of them. I edit files simultaneously in multiple repository and I am the keeper of the global consistency because I know the product from A to Z.

With the agent mode my product my productivity is dropping to a zero.

Before I was not writing any prompts, I was just using discussions as a global context and the tool to edit and making visible in the context, only a limited set of file, to think with the discussion and then applying some very precise patches.

Now I could no more use the outputs given by the discussion. The agent mode is far away of this intelligent and efficient I was making AI.

But the industry is blind and is full of mediocratic approach and people are not seeing the cost and inefficient workflow of prompt editing toward direct code editing.

OpenAI have making impossible the editions of files in VSCODE from the plugin (guessed named OBOE from my old chatGPT discussion) in the remote environnement. This was working 2 years ago but was removed and the edition was possible only on local vscode session.

I did set an NFS bridge over wireguard to access the remote environments in a local manner on my macOS computer and this was enough to get back this feature working.

But now the openAI team seems to have breaking this also.

In my opinion they will never bring back our tool. They are seeking for the monkey coders agentic’s modes.

I did not figure out the solution to this big problem on my side, but this is real and a serious lost for my workflow and they want to get the money from the dummy agents that are producing tremendously big contexts and generate the expensive tokens payment from the industry doom.

The discussion was not consuming the API and they kind of want us to pay more and keep the discussion artificially cheap and more dumb than before when it comes to a technical software subject.

They want to keep the discussion for the normal people use to ask if they are wearing colors properly or if there haircut is fitting.