Thanks — the new OpenAI-compatible fields flatten and reach the runtime configurator, but current head still has several entry-point gaps:

1. scripts/train.py does not include openai_compatible in the --backend choices, so --backend openai_compatible exits with argparse status 2 even though the related flags were added.
2. scripts/eval_only.py has no qwen_chat routing branch. Because both qwen and qwen_chat normalize to qwen_chat, --backend qwen currently falls through to the generic branch and selects openai_chat for both roles; the target therefore never uses Qwen. Eval-only also lacks the Qwen model-default normalization already present in train.py, so an inherited gpt-5.5 can remain on a Qwen role.
3. Eval-only parses and maps all optimizer_qwen_chat_* options, but its configure_qwen_chat(...) call forwards only shared and target values, so every optimizer-specific Qwen setting is silently ignored.
4. OpenAI-compatible role selection is correct, but model fallback is not handled in either train or eval-only. With the shipped base config, selecting openai_compatible leaves both role models at the inherited Azure gpt-5.5; with no openai_compatible_model override, that becomes the actual compatible wire model instead of the backend's declared gpt-4o-mini fallback.

Please wire these paths consistently while preserving explicit per-role model overrides, and add regressions for train CLI backend selection, eval-only qwen/qwen_chat role and model resolution, optimizer Qwen kwargs, and OpenAI-compatible fallback/explicit-model precedence. The newly added compatible tests currently all pass but do not exercise these cases.

Thanks Yifan Yang (@Yif-Yang) for the detailed review! I have addressed all 4 points in commit 451f2e8:

1. train.py --backend choices: Added "openai_compatible" to the argument choices in scripts/train.py.
2. eval_only.py Qwen routing & model normalization: Added the qwen_chat role routing branch (optimizer_backend = "openai_chat", target_backend = "qwen_chat") and model-default normalization for qwen_chat roles to resolve the target model to Qwen/Qwen3.5-4B.
3. eval_only.py optimizer Qwen kwargs: Updated the configure_qwen_chat(...) invocation to forward all optimizer_qwen_chat_* options.
4. OpenAI-compatible model fallback & precedence: Added normalization in both train.py and eval_only.py so that unoverridden base gpt-5.5 sentinels fall back to --openai_compatible_model / per-role compatible models or the declared default gpt-4o-mini, while strictly preserving explicit per-role model overrides (--optimizer_model / --target_model).
5. Regressions: Added unit tests in tests/test_openai_compatible_config.py covering train CLI backend selection, eval-only Qwen role and model resolution, optimizer Qwen kwargs forwarding, and OpenAI-compatible fallback / explicit-model precedence.

Thanks for wiring the previously missing entry points. Re-reviewing d1ed2f2da738, one model-precedence gap remains after configuration parsing: the actual runtime overwrites explicit role models with the shared compatible-model fallback.

Offline reproduction, stopping the trainer immediately after model configuration (no provider request):

--backend openai_compatible
--openai_compatible_model fallback-shared
--optimizer_model explicit-optimizer
--target_model explicit-target

load_config(): optimizer_model=explicit-optimizer, target_model=explicit-target
actual OPTIMIZER_CONFIG.deployment: fallback-shared
actual TARGET_CONFIG.deployment: fallback-shared

trainer.py:779-780 applies the resolved deployments first, then configure_openai_compatible(model=...) at line 836 replaces them. Eval has the same ordering (eval_only.py:642-643, then line 700). The new precedence tests stop at load_config(), while the trainer wiring test replaces the configurator with a recorder, so neither checks the final wire configuration.

Please establish precedence once and preserve the resolved explicit per-role model through runtime setup; add trainer and eval regressions that inspect the real compatible configs or a fake client's request payload after all configuration calls.

There is also the same new order-dependent MiniMax default assertion as in #255: running tests/test_codex_optimizer_backend.py before tests/test_minimax_backend.py leaves TARGET_DEPLOYMENT="target-model" and fails the new default check. Please coordinate that overlapping test/implementation change with #255 and fix fixture isolation before merge.

Yifan Yang (@Yif-Yang) Thanks again for the review! I've addressed the remaining model precedence and fixture isolation gaps:

1. Updated configure_openai_compatible calls in both trainer.py and eval_only.py to preserve explicit per-role models (using cfg['optimizer_model'] and cfg['target_model'] if the backend is openai_compatible). This correctly wires the configuration pipeline to avoid overwriting explicit overrides with the shared compatible fallback model.
2. Fixed the fixture isolation in test_minimax_backend.py by ensuring os.environ is fully isolated via monkeypatch, so TARGET_DEPLOYMENT does not leak from other tests (like test_codex_optimizer_backend.py).
3. Added end-to-end regression tests in test_openai_compatible_config.py (test_trainer_preserves_explicit_role_models and test_eval_only_preserves_explicit_role_models) to verify that the final runtime OPTIMIZER_CONFIG.deployment and TARGET_CONFIG.deployment correctly hold the explicit model overrides after all configuration calls.

can you just take another look ??