fix configured provider slash model routing

api/config.py

@@ -2515,8 +2515,19 @@ def model_with_provider_context(model_id: str, model_provider: str | None = None
     if provider == "openrouter":
         return f"@{provider}:{model}"
 
-    # For non-OpenRouter slash IDs, keep the ID intact so existing custom/proxy
-    # base_url routing and portal-provider handling remain in charge.
+    # Explicit providers configured in config.yaml (for example local llama.cpp,
+    # Ollama, LM Studio, vLLM, or other OpenAI-compatible endpoints) must keep
+    # their provider hint even when the model ID is HuggingFace-style and
+    # contains '/'. Otherwise a selected local model such as
+    # 'unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL' inherits the default provider
+    # (e.g. openai-codex) and is sent to the wrong backend.
+    providers_cfg = cfg.get("providers") if isinstance(cfg, dict) else {}
+    if isinstance(providers_cfg, dict) and provider in providers_cfg:
+        return f"@{provider}:{model}"
+
+    # For non-OpenRouter slash IDs without an explicit configured provider,
+    # keep the ID intact so existing custom/proxy base_url routing and
+    # portal-provider handling remain in charge.
     if "/" in model:
         return model
 

api/routes.py

@@ -3713,6 +3713,15 @@ def _resolve_compatible_session_model_state(
     """
     model = str(model_id or "").strip()
     requested_provider = _clean_session_model_provider(model_provider)
+    if model and requested_provider and model.startswith(f"@{requested_provider}:"):
+        try:
+            from api.config import cfg as _active_cfg
+
+            providers_cfg = _active_cfg.get("providers") if isinstance(_active_cfg, dict) else {}
+        except Exception:
+            providers_cfg = {}
+        if isinstance(providers_cfg, dict) and requested_provider in providers_cfg:
+            return model, requested_provider, False
     if model and requested_provider:
         # Only safe when the model itself does not carry an ``@provider:model``
         # qualifier — qualified strings require the catalog to decide whether

tests/test_issue1855_resolve_model_provider_fast_path.py

@@ -30,7 +30,6 @@ Coverage:
    through).
 """
 import pathlib
-import re
 from unittest.mock import patch
 
 
@@ -175,6 +174,40 @@ class TestSlowPathStillFires:
             "cross-provider artifacts (#1253)."
         )
 
+    def test_configured_provider_qualified_model_skips_catalog(self):
+        """Configured providers already carry trusted routing context."""
+        import api.config as config
+        from api.routes import _resolve_compatible_session_model_state
+
+        old_cfg = dict(config.cfg)
+        config.cfg["model"] = {
+            "provider": "openai-codex",
+            "default": "gpt-5.5",
+        }
+        config.cfg["providers"] = {
+            "local-llama": {
+                "base_url": "http://127.0.0.1:8088/v1",
+                "api_key": "test-key",
+            },
+        }
+        try:
+            with patch("api.routes.get_available_models") as mock_catalog:
+                mock_catalog.side_effect = AssertionError("catalog should not be called")
+                result = _resolve_compatible_session_model_state(
+                    "@local-llama:unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL",
+                    "local-llama",
+                )
+        finally:
+            config.cfg.clear()
+            config.cfg.update(old_cfg)
+
+        assert result == (
+            "@local-llama:unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL",
+            "local-llama",
+            False,
+        )
+        assert mock_catalog.call_count == 0
+
     def test_no_requested_provider_goes_to_slow_path(self):
         """When provider isn't supplied, slow path must repair from catalog."""
         from api.routes import _resolve_compatible_session_model_state

tests/test_provider_mismatch.py

@@ -522,6 +522,37 @@ def test_non_openrouter_slash_model_provider_context_stays_unqualified():
     assert runtime_model == "anthropic/claude-sonnet-4.6"
 
 
+def test_configured_provider_slash_model_keeps_provider_context():
+    """Configured OpenAI-compatible providers need explicit context for slash IDs."""
+    import api.config as config
+
+    old_cfg = dict(config.cfg)
+    config.cfg["model"] = {
+        "provider": "openai-codex",
+        "default": "gpt-5.5",
+    }
+    config.cfg["providers"] = {
+        "local-llama": {
+            "base_url": "http://127.0.0.1:8088/v1",
+            "api_key": "test-key",
+        },
+    }
+    try:
+        runtime_model = config.model_with_provider_context(
+            "unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL",
+            "local-llama",
+        )
+        model, provider, base_url = config.resolve_model_provider(runtime_model)
+    finally:
+        config.cfg.clear()
+        config.cfg.update(old_cfg)
+
+    assert runtime_model == "@local-llama:unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL"
+    assert model == "unsloth/gemma-4-12b-it-GGUF:UD-Q4_K_XL"
+    assert provider == "local-llama"
+    assert base_url == "http://127.0.0.1:8088/v1"
+
+
 def test_cursor_acp_slash_model_always_gets_provider_hint():
     """ACP subprocess models with '/' must not fall through to config default."""
     import api.config as config