fix(llm): OpenRouter reasoning-off — use effort=minimal (gpt-5 compat)

ciris_engine/logic/services/runtime/llm_service/service.py

@@ -1509,7 +1509,7 @@
         raise RuntimeError("All endpoints exhausted without success")
 
     @staticmethod
     def _build_reasoning_off_extras(base_url: str, model_name: str) -> Dict[str, Any]:
         """Return the ``extra_body`` keys that disable reasoning for this
         specific (provider, model) combination.
 
@@ -1547,6 +1547,23 @@
         model = (model_name or "").lower()
 
         if "openrouter.ai" in base:
+            # Reasoning-OFF contract: pass `reasoning.enabled=false` so any
+            # reasoning-capable model that honours it (claude-sonnet-4.6,
+            # openai/gpt-5-chat, deepseek, etc.) emits zero reasoning tokens
+            # as before. `effort=minimal` is *not* a disable on those models —
+            # it's the lowest-effort setting and still emits reasoning tokens.
+            #
+            # OpenRouter explicitly rejects `reasoning.enabled=false` for the
+            # gpt-5 reasoning family with 400 "Reasoning is mandatory for this
+            # endpoint and cannot be disabled." (verified 2026-05-16 against
+            # `openai/gpt-5`). For that narrow case fall back to
+            # `effort=minimal`, which the gpt-5 family honours by emitting
+            # 0 reasoning tokens (cost ~$0.0002 on a trivial prompt). The
+            # chat-tuned variant `openai/gpt-5-chat` is reasoning-capable, not
+            # reasoning-mandatory, and stays on the universal `enabled=false`
+            # path.
+            if model.startswith("openai/gpt-5") and "chat" not in model:
+                return {"reasoning": {"effort": "minimal"}}
             return {"reasoning": {"enabled": False}}
 
         if "together" in base:

tests/ciris_engine/logic/services/runtime/llm_service/test_llm_service_coverage.py

@@ -325,10 +325,33 @@ def test_deepinfra_carries_vllm_and_reasoning_enabled_keys(self):
         assert extra_body["chat_template_kwargs"] == {"enable_thinking": False}
         assert extra_body["reasoning"] == {"enabled": False}
 
-    def test_openrouter_carries_reasoning_enabled_false(self):
+    def test_openrouter_default_carries_reasoning_enabled_false(self):
+        # Default reasoning-OFF contract on OpenRouter: send `enabled=false`
+        # so reasoning-capable models (claude, gpt-5-chat, deepseek, …) emit
+        # zero reasoning tokens. `effort=minimal` is *not* a disable on those
+        # models, so we only use it for the narrow reasoning-mandatory case
+        # below.
         extra_body = self._build_for("https://openrouter.ai/api/v1", "kimi/k2-0130-preview")["extra_body"]
         assert extra_body["reasoning"] == {"enabled": False}
 
+    def test_openrouter_gpt5_uses_effort_minimal(self):
+        # OpenRouter rejects reasoning.enabled=false for the gpt-5 reasoning
+        # family with 400 "Reasoning is mandatory for this endpoint and cannot
+        # be disabled." (verified 2026-05-16 against openai/gpt-5). For that
+        # narrow case we fall back to effort=minimal, which gpt-5 honours by
+        # emitting 0 reasoning tokens.
+        for model in ("openai/gpt-5", "openai/gpt-5-mini", "openai/gpt-5-nano"):
+            extra_body = self._build_for("https://openrouter.ai/api/v1", model)["extra_body"]
+            assert extra_body["reasoning"] == {"effort": "minimal"}, model
+
+    def test_openrouter_gpt5_chat_stays_on_enabled_false(self):
+        # gpt-5-chat is the chat-tuned (reasoning-capable, not mandatory)
+        # variant and honours reasoning.enabled=false. It must stay on the
+        # universal disable path so it doesn't accidentally re-enable
+        # reasoning at lowest-effort under our reasoning-OFF contract.
+        extra_body = self._build_for("https://openrouter.ai/api/v1", "openai/gpt-5-chat")["extra_body"]
+        assert extra_body["reasoning"] == {"enabled": False}
+
     def test_local_endpoint_carries_vllm_key(self):
         extra_body = self._build_for("http://localhost:8080/v1", "llama")["extra_body"]
         assert extra_body["chat_template_kwargs"] == {"enable_thinking": False}
@@ -453,9 +476,11 @@ def test_openrouter_includes_provider_config(self):
             retry_state=RetryState(),
         )
 
-        # Verify reasoning is disabled
+        # gpt-4o is reasoning-capable, not reasoning-mandatory — it honours
+        # `enabled=false`. Reasoning-mandatory gpt-5 models would take the
+        # `effort=minimal` fallback instead (covered separately above).
         assert "extra_body" in extra_kwargs
         extra_body = extra_kwargs["extra_body"]
         # Provider config may or may not be present depending on env
         assert "reasoning" in extra_body
-        assert extra_body["reasoning"]["enabled"] is False
+        assert extra_body["reasoning"] == {"enabled": False}