Fix issue #654 Respect --model parameter when using custom base URL

[codearranger]

TLDR

Fixed the --model parameter being ignored when using a custom base URL. The CLI now correctly respects the model specified via -m or --model flags when using alternative API endpoints.

Dive Deeper

The issue was in the createContentGeneratorConfig function in packages/core/src/core/contentGenerator.ts. When using OpenAI auth type (triggered when a custom base URL is set), the code was overriding the model selection with the OPENAI_MODEL environment variable or falling back to DEFAULT_QWEN_MODEL, completely bypassing the model specified through the CLI's --model parameter.

The fix ensures proper priority ordering:

1. Model from config (includes CLI --model parameter) - highest priority
2. Environment variable (OPENAI_MODEL or QWEN_MODEL) - fallback
3. Default model constant - last resort

This same logic was applied to both OpenAI and QWEN OAuth authentication paths to ensure consistency.

Reviewer Test Plan

To validate this fix:

1. Test with custom base URL and model parameter:

   export OPENAI_API_KEY="your-key"
   export OPENAI_BASE_URL="https://your-custom-endpoint.com"
   qwen --model gpt-4 "test prompt"
   # Should use gpt-4, not qwen3-coder-plus or coder-model

2. Test that environment variable still works as fallback:

   export OPENAI_MODEL="gpt-3.5-turbo"
   qwen "test prompt"  # No --model flag
   # Should use gpt-3.5-turbo from env var

3. Test precedence (CLI flag should override env var):

   export OPENAI_MODEL="gpt-3.5-turbo"
   qwen --model gpt-4 "test prompt"
   # Should use gpt-4 from CLI, not gpt-3.5-turbo from env

Testing Matrix

	🍏	🪟	🐧
npm run	✅	❓	❓
npx	❓	❓	❓
Docker	❓	❓	❓
Podman	❓	-	-
Seatbelt	❓	-	-

Linked issues / bugs

Fixes #654

[Mingholy]

Looks good. But just to confirm: users can change the model manually in interactive mode, model from config(using config.getModel()) is not always the same as cli args. What's your expectation in this case?

[bbartels]

Looks good. But just to confirm: users can change the model manually in interactive mode, model from config(using config.getModel()) is not always the same as cli args. What's your expectation in this case?

I suppose that is fine, it should prioritise the --model parameter, but if the user changes it interactively then it should be preferring that.

[bbartels]

Would be good to get this change in @Mingholy as the current behaviour is quite confusing