fix: respect reasoning_config effort for custom providers

agent/transports/chat_completions.py

@@ -421,7 +421,12 @@ def build_kwargs(
                 if gh_reasoning is not None:
                     extra_body["reasoning"] = gh_reasoning
             else:
-                extra_body["reasoning"] = {"enabled": True, "effort": "medium"}
+                _effort = "medium"
+                if reasoning_config and isinstance(reasoning_config, dict):
+                    _e = (reasoning_config.get("effort") or "").strip().lower()
+                    if _e in {"low", "medium", "high", "xhigh"}:
+                        _effort = _e
+                extra_body["reasoning"] = {"enabled": True, "effort": _effort}
 
         if provider_name == "gemini":
             raw_thinking_config = _build_gemini_thinking_config(model, reasoning_config)

run_agent.py

@@ -4862,6 +4862,24 @@ def _supports_reasoning_extra_body(self) -> bool:
             opts = self._lmstudio_reasoning_options_cached()
             # "off-only" (or absent) means no real reasoning capability.
             return any(opt and opt != "off" for opt in opts)
+        # Custom/user-defined providers: check providers.<name>.models.<model>.reasoning
+        # in config.yaml. This lets users opt in to reasoning for models served
+        # through custom endpoints (CLIProxyAPI, vLLM, Ollama, etc.) without
+        # adding a dedicated provider profile.
+        try:
+            from hermes_cli.config import load_config as _load_rc_cfg
+            _rc_cfg = _load_rc_cfg()
+            _rc_providers = _rc_cfg.get("providers")
+            if isinstance(_rc_providers, dict):
+                _rc_prov = _rc_providers.get(self.provider or "")
+                if isinstance(_rc_prov, dict):
+                    _rc_models = _rc_prov.get("models")
+                    if isinstance(_rc_models, dict):
+                        _rc_mcfg = _rc_models.get(self.model or "")
+                        if isinstance(_rc_mcfg, dict) and _rc_mcfg.get("reasoning"):
+                            return True
+        except Exception:
+            pass
         if "openrouter" not in self._base_url_lower:
             return False
         if "api.mistral.ai" in self._base_url_lower:

tests/agent/transports/test_chat_completions.py

@@ -252,6 +252,27 @@ def test_reasoning_default(self, transport):
         )
         assert kw["extra_body"]["reasoning"] == {"enabled": True, "effort": "medium"}
 
+    def test_reasoning_effort_from_config(self, transport):
+        """reasoning_config.effort must propagate, not be hardcoded to medium."""
+        msgs = [{"role": "user", "content": "Hi"}]
+        for effort in ("low", "medium", "high", "xhigh"):
+            kw = transport.build_kwargs(
+                model="gpt-4o", messages=msgs,
+                supports_reasoning=True,
+                reasoning_config={"effort": effort},
+            )
+            assert kw["extra_body"]["reasoning"] == {"enabled": True, "effort": effort}
+
+    def test_reasoning_effort_invalid_falls_back_to_medium(self, transport):
+        """Unknown effort values fall back to medium, not crash."""
+        msgs = [{"role": "user", "content": "Hi"}]
+        kw = transport.build_kwargs(
+            model="gpt-4o", messages=msgs,
+            supports_reasoning=True,
+            reasoning_config={"effort": "garbage"},
+        )
+        assert kw["extra_body"]["reasoning"] == {"enabled": True, "effort": "medium"}
+
     def test_nous_omits_disabled_reasoning(self, transport):
         from providers import get_provider_profile
         profile = get_provider_profile("nous")

tests/run_agent/test_run_agent.py

@@ -6694,6 +6694,46 @@ def test_xiaomi_models_are_treated_as_reasoning_capable(self):
             agent.model = model
             assert agent._supports_reasoning_extra_body() is True, model
 
+    def test_custom_provider_model_reasoning_opt_in(self):
+        """providers.<name>.models.<model>.reasoning: true enables reasoning for custom providers."""
+        from unittest.mock import patch as mock_patch
+        agent = self._make_agent()
+        agent.provider = "cliproxyapi"
+        agent.base_url = "http://localhost:8317/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.model = "mimo-v2.5-pro"
+        cfg = {
+            "providers": {
+                "cliproxyapi": {
+                    "models": {
+                        "mimo-v2.5-pro": {"reasoning": True},
+                    }
+                }
+            }
+        }
+        with mock_patch("hermes_cli.config.load_config", return_value=cfg):
+            assert agent._supports_reasoning_extra_body() is True
+
+    def test_custom_provider_without_reasoning_flag_returns_false(self):
+        """Custom provider without reasoning: true must still return False (backwards compat)."""
+        from unittest.mock import patch as mock_patch
+        agent = self._make_agent()
+        agent.provider = "cliproxyapi"
+        agent.base_url = "http://localhost:8317/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.model = "mimo-v2.5-pro"
+        cfg = {
+            "providers": {
+                "cliproxyapi": {
+                    "models": {
+                        "mimo-v2.5-pro": {"context_length": 131072},
+                    }
+                }
+            }
+        }
+        with mock_patch("hermes_cli.config.load_config", return_value=cfg):
+            assert agent._supports_reasoning_extra_body() is False
+
 
 class TestMemoryContextSanitization:
     """sanitize_context() helper correctness — used at provider boundaries."""