[Desktop][Custom provider] Thread title generation silently falls back from gpt-5.6-luna to deepseek-v4-pro

What version of the Codex App are you using (From “About Codex” dialog)?

26.908.4834.0

What subscription do you have?

Third-party API Key (DeepSeek via CC Switch)

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

When the selected main model is deepseek-v4-flash from a custom provider, Codex Desktop's automatic thread-title generation requests the hard-coded gpt-5.6-luna model with feature=thread_title and reasoningEffort=low.

Because the custom provider does not expose gpt-5.6-luna, the title-generation request silently falls back to deepseek-v4-pro instead of using the selected deepseek-v4-flash model. The main conversation still uses deepseek-v4-flash.

This causes a background feature to ignore the user-selected model and may result in unexpected model usage and cost. The fallback is not visible in the UI and no warning is shown.

What steps can reproduce the bug?

Feedback ID: 01a097a7-513e-7d63-b608-2ee6c9d7677c

What is the expected behavior?

Automatic thread-title generation should not silently select a different model from the one selected by the user.

When gpt-5.6-luna is unavailable from a custom provider, Codex should use the currently selected model, skip title generation, show a clear warning, or provide a dedicated title-generation model setting.

In this case, selecting deepseek-v4-flash should not result in a background request being sent to deepseek-v4-pro.

Additional information

Feedback ID: 01a097a7-513e-7d63-b608-2ee6c9d7677c

1. Configure a custom OpenAI-compatible provider that exposes deepseek-v4-flash and deepseek-v4-pro.
2. Select deepseek-v4-flash as the main model in Codex Desktop.
3. Create a new thread and send the first message.
4. Allow Codex to generate the thread title automatically.
5. Inspect the Codex Desktop log and the provider/proxy request log.

Observed evidence:

• Codex Desktop log:
  feature=thread_title
  model=gpt-5.6-luna
  reasoningEffort=low
  inputTokens=13502
  outputTokens=115
  cachedTokens=13440

• Actual request received by the custom provider:
  model=deepseek-v4-pro
  request_model=deepseek-v4-pro
  status=200

• Timestamp: 2026-09-13 06:05 CST

• The main conversation uses deepseek-v4-flash.

Start by locating the desktop thread-title generation entry point and the custom-provider model fallback path; the issue provides a reproducible configuration and log evidence but names no files or tests. Reproduce with deepseek-v4-flash selected, inspect the title-generation and provider request logs, and verify that title generation no longer silently uses deepseek-v4-pro.