Windows: Voice Chat Fork loses parent project context, model selection, and AGENTS startup behavior

What version of the Codex App are you using?

• Codex desktop app on Windows: 26.814.5167.0
• Model selected in the parent task: GPT-5.6 Terra, Medium reasoning

What platform are you using?

Windows desktop app.

What issue are you seeing?

Starting Voice Chat from a new task created in a multi-folder Codex project creates a separate Voice Chat Fork that does not preserve the parent task's effective project context.

In this reproduction, the parent task was started in an All Repos project with 2 configured source folders:

• C:\Github
• a separate shared coordinator folder

The parent task showed GPT-5.6 Terra with Medium reasoning selected. After clicking the Voice Chat button:

• The child is titled Voice Chat Fork.
• The child composer shows GPT-5.6 Terra Light, rather than the parent's Medium selection.
• The child does not follow the parent's/global AGENTS.md startup instruction when greeted with the coordinator activation phrase.
• Instead of performing the required context load and task rename, Voice responds generically.
• Inspection of the resulting task shows it running from a generated realtime-voice-chat-* workspace rather than retaining the parent project workspace.

The failure is visible on the first voice reply, before any project work begins. The child still displays Full access, so this is not explained by a missing sandbox permission.

This was observed immediately after updating the desktop app on August 18, 2026. The same workflow previously preserved the expected coordinator behavior.

Steps to reproduce

1. Create or open a multi-folder Codex project with a root AGENTS.md instruction that has a distinctive first-message behavior.
2. Start a new task in that project and select a non-default model/reasoning combination.
3. Click the Voice Chat button in the task composer.
4. In the Voice Chat Fork, say the distinctive activation phrase from AGENTS.md.
5. Compare the parent and child:
   • selected model/reasoning;
   • effective project/source-folder context;
   • first-response behavior.

Expected behavior

Voice Chat should remain attached to the task that launched it, or its fork should inherit all relevant task configuration:

• multi-folder project and source folders;
• effective AGENTS.md instructions;
• selected model and reasoning effort;
• sandbox/access configuration;
• task-level startup behavior.

If Voice Chat intentionally creates an isolated task, the UI should clearly state that it will not inherit project context and should offer an option to retain it.

Actual behavior

The Voice Chat Fork appears to receive a generated workspace and a different model/reasoning setting. It ignores the parent task's visible project-specific behavior on its first response.

Why this matters

For workflows that use AGENTS.md as durable startup context, Voice Chat becomes unreliable: the user has to re-explain identity, project scope, and operating rules after starting voice. It also creates a risk that voice work runs under a different model/reasoning setting than the task the user explicitly configured.

Additional information

I can provide screenshots showing the parent All Repos task with Terra Medium selected and the resulting Voice Chat Fork with Terra Light selected. I can also provide task identifiers privately if useful; they are omitted here because they are local task metadata.

This is distinct from tool-handler failures during Realtime Voice. The problem occurs at Voice Chat Fork creation and is observable immediately in the first reply.

Start at the Voice Chat button and Voice Chat Fork creation flow, reproducing the issue in a Windows multi-folder project with a distinctive AGENTS.md instruction and non-default model reasoning. Compare the parent and child task's project folders, startup behavior, model setting, and access configuration; done means the fork preserves those settings or clearly communicates and offers the intentional isolation described in the issue.