fix(codex): return Codex CLI-compatible /v1/models response

Codex CLI deserializes /v1/models as {"models": [...]} (ModelsResponse)
while the proxy returns OpenAI-compatible {"object": "list", "data": [...]}.
Detect Codex CLI via client_version query param and passthrough the
upstream models.json catalog.

Fixes: #59

Author: b3nw

src/proxy_app/main.py

@@ -1548,7 +1548,24 @@ async def list_models(
         enriched: If True (default), returns detailed model info with pricing and capabilities.
                   If False, returns minimal OpenAI-compatible response.
         refresh: If True, triggers all providers to refresh their upstream model listings and updates the metadata cache.
+
+    Codex CLI sends ``client_version`` as a query parameter and deserializes
+    the response as ``{"models": [...]}``.  When we detect that parameter we
+    return the upstream models.json catalog verbatim so the CLI's model
+    picker/refresh works without warnings.
     """
+    # Codex CLI compat: return upstream {"models": [...]} catalog when
+    # the request carries a client_version query param (Codex-specific).
+    client_version = request.query_params.get("client_version")
+    if client_version is not None:
+        try:
+            from rotator_library.providers.codex_provider import get_raw_models_catalog
+            raw_catalog = get_raw_models_catalog()
+            if raw_catalog is not None:
+                return {"models": raw_catalog}
+        except ImportError:
+            pass
+
     if refresh:
         # Trigger model registry refresh if available
         if hasattr(request.app.state, "model_info_service"):

src/rotator_library/providers/codex_provider.py

@@ -171,6 +171,8 @@ def _derive_ws_endpoint() -> str:
 # Module-level cache for dynamic model data
 _models_cache: Optional[Dict[str, Any]] = None
 _models_cache_time: float = 0.0
+# Raw upstream models.json for Codex CLI passthrough ({"models": [...]})
+_raw_models_catalog: Optional[List[Dict[str, Any]]] = None
 
 
 def _fetch_models_from_github() -> Optional[Dict[str, Any]]:
@@ -198,6 +200,9 @@ def _fetch_models_from_github() -> Optional[Dict[str, Any]]:
             lib_logger.warning("[Codex] models.json from GitHub had empty models list")
             return None
 
+        global _raw_models_catalog
+        _raw_models_catalog = models_list
+
         base_models = []
         reasoning_efforts = {}
         context_limits: Dict[str, int] = {}
@@ -343,6 +348,19 @@ def get_model_plan_access() -> Dict[str, set]:
     return _get_model_data().get("plan_access", {})
 
 
+def get_raw_models_catalog() -> Optional[List[Dict[str, Any]]]:
+    """
+    Return the raw upstream models.json catalog for Codex CLI passthrough.
+
+    Codex CLI expects {"models": [...]} with full ModelInfo objects (slug,
+    display_name, context_window, etc.) rather than the OpenAI-compatible
+    {"object": "list", "data": [...]}.  Ensures the cache is populated
+    before returning.
+    """
+    _get_model_data()
+    return _raw_models_catalog
+
+
 def _model_requires_paid_tier(model: str) -> bool:
     """Check if a model is restricted to paid plans only (not available on free-class plans)."""
     plan_map = get_model_plan_access()