[Bug] Codex Desktop rejects external-model custom sub-agent with ChatGPT account #897

Pre-submission

• I searched existing issues and did not find a matching report.
• I reviewed CONTRIBUTING.md.
• No API key, request header, credential, or customer data is included.

Category

Model configuration / Agent framework compatibility

Agent framework

Codex

Severity and frequency

• Severity: Blocking for custom worker delegation; the main Codex session remains usable.
• Frequency: Reproduced on each attempted delegation.

Description

A named custom sub-agent configured to use an external OpenAI-compatible model provider cannot start from Codex Desktop when the manager session is authenticated with a ChatGPT account.

This test used a third-party QwenCloud model rather than a Z.ai model. I am reporting it here because this repository accepts Codex-framework feedback and the same Codex account/provider boundary may affect Z.ai/GLM custom model profiles. If this is outside Z.ai's scope, please redirect the report to the correct integration owner.

The external provider appears not to have been reached; the failure looks like local/session model validation.

Steps to reproduce

1. On Windows, configure a user-scoped named Codex worker using an external OpenAI-compatible provider. Store the provider credential only in a user environment variable.
2. Restart Codex Desktop.
3. Open a local repository chat authenticated with a ChatGPT account.
4. Ask the manager chat to spawn the named worker qwen_strong, configured for qwen3.8-max.
5. Observe the 400 error below.
6. Add new named workers such as qwen_fast and qwen_auto.
7. From the already-open chat, attempt to spawn qwen_fast.
8. Observe that the worker type is not recognized.

Expected behavior

Codex should either:

• start the named worker through its configured external provider, or
• clearly document that ChatGPT-authenticated Codex Desktop sessions cannot delegate to external-provider custom agents.

When a new worker profile is added, the UI should also explain whether a full app restart or a new chat is required before the worker name is available.

Actual behavior

First attempt:

Agent errored: {"type":"error","status":400,"error":{"type":"invalid_request_error","message":"The 'qwen3.8-max' model is not supported when using Codex with a ChatGPT account."}}

After adding a smaller worker profile in the same open client session:

unknown agent_type 'qwen_fast'

No repository files were changed by either attempt.

Environment

• OS: Microsoft Windows NT 10.0.26200.0
• Codex CLI: 0.156.1
• Client: Codex Desktop, local repository chat
• Authentication: ChatGPT account
• Provider channel: external OpenAI-compatible provider
• Date reproduced: 2026-10-01 (Asia/Singapore)
• ZCode version: Not applicable; the issue occurred in Codex Desktop

Questions for maintainers

1. Does the Codex framework currently support named external-provider sub-agents when the manager session uses ChatGPT authentication?
2. Are Z.ai/GLM custom model profiles subject to the same restriction?
3. Is a full app restart sufficient after adding worker profiles, or must a new local chat also be created?