ollama - 💡(How to fix) Fix Codex App integration: switching GUI permissions from Custom(config.toml) to Full Access can break patch tool calls

What is the issue?

I am using ollama launch codex-app on macOS to run Codex Desktop App with a local Ollama model.

After launching Codex App through Ollama and selecting a local model, I changed the Codex Desktop GUI permission mode from Custom (config.toml) to Full Access using the permission button in the app.

After that, when asking Codex to modify files, the model output indicated that the patch tool call failed.

The workaround was to keep the GUI permission mode set to Custom (config.toml), and configure file write permissions in ~/.codex/config.toml instead:

sandbox_mode = "workspace-write"

I intentionally did not use danger-full-access, because I only wanted the model to write within the workspace and did not want to grant unrestricted filesystem access.

Steps to reproduce

1. On macOS, start Codex Desktop App through Ollama:

   ollama launch codex-app --model <local-model-name>

2. Open Codex Desktop App and select the local Ollama model.

3. In the Codex Desktop GUI, use the permission button to switch from Custom (config.toml) to Full Access.

4. Ask Codex to modify a file in the current workspace.

5. Observe that the model output reports a patch tool call failure.

Expected behavior

When using ollama launch codex-app, Codex Desktop permissions should remain consistent and tool calls such as patch/file edits should continue to work.

If Full Access is not supported or may conflict with the configuration generated/used by Ollama, the integration should either:

• preserve and recommend Custom (config.toml) for Ollama-launched Codex App sessions;
• document that users should configure sandbox_mode = "workspace-write" in ~/.codex/config.toml;
• warn users not to switch to Full Access from the GUI in this setup;
• or ensure the generated Codex App configuration remains compatible with GUI permission changes.

Workaround

Keep Codex Desktop GUI set to Custom (config.toml) and configure permissions in ~/.codex/config.toml:

sandbox_mode = "workspace-write"

This allows patch/file modification tools to work without using danger-full-access.

Related issues checked

I searched for related Codex App integration issues, but did not find this specific permission-mode problem. These look related to Codex integration in general, but appear to be different issues:

• https://github.com/ollama/ollama/issues/15912
• https://github.com/ollama/ollama/issues/16177
• https://github.com/ollama/ollama/issues/16188
• https://github.com/ollama/ollama/issues/14752

Environment

• OS: macOS
• Ollama version: 0.24.0 client
• Integration: ollama launch codex-app
• App: Codex Desktop App
• Local model: local Ollama model, e.g. qwen3.6 / gemma4